Datasets:
Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +260 -0
- parse/dev/2nJdh_C-UWe/2nJdh_C-UWe.md +413 -0
- parse/dev/2nJdh_C-UWe/2nJdh_C-UWe_content_list.json +1125 -0
- parse/dev/2nJdh_C-UWe/2nJdh_C-UWe_middle.json +0 -0
- parse/dev/2nJdh_C-UWe/2nJdh_C-UWe_model.json +0 -0
- parse/dev/3tYvDb4dwab/3tYvDb4dwab.md +292 -0
- parse/dev/3tYvDb4dwab/3tYvDb4dwab_content_list.json +1596 -0
- parse/dev/3tYvDb4dwab/3tYvDb4dwab_middle.json +0 -0
- parse/dev/3tYvDb4dwab/3tYvDb4dwab_model.json +0 -0
- parse/dev/6HN7LHyzGgC/6HN7LHyzGgC.md +398 -0
- parse/dev/6HN7LHyzGgC/6HN7LHyzGgC_content_list.json +2043 -0
- parse/dev/6HN7LHyzGgC/6HN7LHyzGgC_middle.json +0 -0
- parse/dev/6HN7LHyzGgC/6HN7LHyzGgC_model.json +0 -0
- parse/dev/6dlC7E1H_9/6dlC7E1H_9_middle.json +0 -0
- parse/dev/6dlC7E1H_9/6dlC7E1H_9_model.json +0 -0
- parse/dev/89ia77nZ8u/89ia77nZ8u.md +0 -0
- parse/dev/89ia77nZ8u/89ia77nZ8u_content_list.json +0 -0
- parse/dev/89ia77nZ8u/89ia77nZ8u_middle.json +0 -0
- parse/dev/89ia77nZ8u/89ia77nZ8u_model.json +0 -0
- parse/dev/9XFSbDPmdW/9XFSbDPmdW.md +0 -0
- parse/dev/9XFSbDPmdW/9XFSbDPmdW_content_list.json +0 -0
- parse/dev/9XFSbDPmdW/9XFSbDPmdW_middle.json +0 -0
- parse/dev/A6X9y8n4sT/A6X9y8n4sT.md +308 -0
- parse/dev/A6X9y8n4sT/A6X9y8n4sT_content_list.json +1297 -0
- parse/dev/A6X9y8n4sT/A6X9y8n4sT_middle.json +0 -0
- parse/dev/A6X9y8n4sT/A6X9y8n4sT_model.json +0 -0
- parse/dev/B4tkwuzeiY/B4tkwuzeiY_content_list.json +0 -0
- parse/dev/B4tkwuzeiY/B4tkwuzeiY_middle.json +0 -0
- parse/dev/B4tkwuzeiY/B4tkwuzeiY_model.json +0 -0
- parse/dev/DAxQXzdq8SF/DAxQXzdq8SF_model.json +0 -0
- parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc.md +304 -0
- parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc_content_list.json +1379 -0
- parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc_middle.json +0 -0
- parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc_model.json +0 -0
- parse/dev/IJNDyqdRF0m/IJNDyqdRF0m.md +429 -0
- parse/dev/IJNDyqdRF0m/IJNDyqdRF0m_content_list.json +0 -0
- parse/dev/IJNDyqdRF0m/IJNDyqdRF0m_middle.json +0 -0
- parse/dev/IJNDyqdRF0m/IJNDyqdRF0m_model.json +0 -0
- parse/dev/LSKlp_aceOC/LSKlp_aceOC.md +271 -0
- parse/dev/LSKlp_aceOC/LSKlp_aceOC_middle.json +0 -0
- parse/dev/Loek7hfb46P/Loek7hfb46P.md +0 -0
- parse/dev/Loek7hfb46P/Loek7hfb46P_content_list.json +0 -0
- parse/dev/Loek7hfb46P/Loek7hfb46P_middle.json +0 -0
- parse/dev/Loek7hfb46P/Loek7hfb46P_model.json +0 -0
- parse/dev/MbCAOMGsZXC/MbCAOMGsZXC.md +239 -0
- parse/dev/MbCAOMGsZXC/MbCAOMGsZXC_content_list.json +1053 -0
- parse/dev/MbCAOMGsZXC/MbCAOMGsZXC_middle.json +0 -0
- parse/dev/MbCAOMGsZXC/MbCAOMGsZXC_model.json +0 -0
- parse/dev/O476oWmiNNp/O476oWmiNNp.md +0 -0
- parse/dev/O476oWmiNNp/O476oWmiNNp_content_list.json +0 -0
.gitattributes
CHANGED
|
@@ -4259,3 +4259,263 @@ parse/train/nLYMajjctMh/nLYMajjctMh_layout.pdf filter=lfs diff=lfs merge=lfs -te
|
|
| 4259 |
parse/train/nLYMajjctMh/nLYMajjctMh_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4260 |
parse/train/nLYMajjctMh/nLYMajjctMh_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4261 |
parse/train/sqZ-b0a6Wm/sqZ-b0a6Wm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4259 |
parse/train/nLYMajjctMh/nLYMajjctMh_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4260 |
parse/train/nLYMajjctMh/nLYMajjctMh_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4261 |
parse/train/sqZ-b0a6Wm/sqZ-b0a6Wm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4262 |
+
parse/train/sqZ-b0a6Wm/sqZ-b0a6Wm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4263 |
+
parse/train/sqZ-b0a6Wm/sqZ-b0a6Wm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4264 |
+
parse/train/SJJySbbAZ/SJJySbbAZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4265 |
+
parse/train/SJJySbbAZ/SJJySbbAZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4266 |
+
parse/train/SJJySbbAZ/SJJySbbAZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4267 |
+
parse/train/J_VRu0ZFVX-/J_VRu0ZFVX-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4268 |
+
parse/train/J_VRu0ZFVX-/J_VRu0ZFVX-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4269 |
+
parse/train/J_VRu0ZFVX-/J_VRu0ZFVX-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4270 |
+
parse/train/uys9OcmXNtU/uys9OcmXNtU_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4271 |
+
parse/train/uys9OcmXNtU/uys9OcmXNtU_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4272 |
+
parse/train/uys9OcmXNtU/uys9OcmXNtU_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4273 |
+
parse/train/rJxAo2VYwr/rJxAo2VYwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4274 |
+
parse/train/rJxAo2VYwr/rJxAo2VYwr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4275 |
+
parse/train/rJxAo2VYwr/rJxAo2VYwr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4276 |
+
parse/train/TRDAFiwDq8A/TRDAFiwDq8A_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4277 |
+
parse/train/TRDAFiwDq8A/TRDAFiwDq8A_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4278 |
+
parse/train/TRDAFiwDq8A/TRDAFiwDq8A_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4279 |
+
parse/train/KJNcAkY8tY4/KJNcAkY8tY4_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4280 |
+
parse/train/KJNcAkY8tY4/KJNcAkY8tY4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4281 |
+
parse/train/KJNcAkY8tY4/KJNcAkY8tY4_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4282 |
+
parse/train/YNA3UEPCUp/YNA3UEPCUp_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4283 |
+
parse/train/YNA3UEPCUp/YNA3UEPCUp_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4284 |
+
parse/train/YNA3UEPCUp/YNA3UEPCUp_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4285 |
+
parse/train/BJesDsA9t7/BJesDsA9t7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4286 |
+
parse/train/BJesDsA9t7/BJesDsA9t7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4287 |
+
parse/train/BJesDsA9t7/BJesDsA9t7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4288 |
+
parse/train/cLYyCXHU7g1n/cLYyCXHU7g1n_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4289 |
+
parse/train/cLYyCXHU7g1n/cLYyCXHU7g1n_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4290 |
+
parse/train/cLYyCXHU7g1n/cLYyCXHU7g1n_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4291 |
+
parse/train/HJxV-ANKDH/HJxV-ANKDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4292 |
+
parse/train/HJxV-ANKDH/HJxV-ANKDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4293 |
+
parse/train/HJxV-ANKDH/HJxV-ANKDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4294 |
+
parse/train/d0MtHWY0NZ/d0MtHWY0NZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4295 |
+
parse/train/d0MtHWY0NZ/d0MtHWY0NZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4296 |
+
parse/train/d0MtHWY0NZ/d0MtHWY0NZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4297 |
+
parse/train/SkfrvsA9FX/SkfrvsA9FX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4298 |
+
parse/train/SkfrvsA9FX/SkfrvsA9FX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4299 |
+
parse/train/SkfrvsA9FX/SkfrvsA9FX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4300 |
+
parse/train/rkl85oRqYX/rkl85oRqYX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4301 |
+
parse/train/rkl85oRqYX/rkl85oRqYX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4302 |
+
parse/train/rkl85oRqYX/rkl85oRqYX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4303 |
+
parse/train/xgGS6PmzNq6/xgGS6PmzNq6_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4304 |
+
parse/train/xgGS6PmzNq6/xgGS6PmzNq6_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4305 |
+
parse/train/xgGS6PmzNq6/xgGS6PmzNq6_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4306 |
+
parse/train/ByeMPlHKPH/ByeMPlHKPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4307 |
+
parse/train/ByeMPlHKPH/ByeMPlHKPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4308 |
+
parse/train/ByeMPlHKPH/ByeMPlHKPH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4309 |
+
parse/train/rklOg6EFwS/rklOg6EFwS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4310 |
+
parse/train/rklOg6EFwS/rklOg6EFwS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4311 |
+
parse/train/rklOg6EFwS/rklOg6EFwS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4312 |
+
parse/train/tKlYQJLYN8v/tKlYQJLYN8v_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4313 |
+
parse/train/tKlYQJLYN8v/tKlYQJLYN8v_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4314 |
+
parse/train/tKlYQJLYN8v/tKlYQJLYN8v_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4315 |
+
parse/train/QgX15Mdi1E_/QgX15Mdi1E__layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4316 |
+
parse/train/QgX15Mdi1E_/QgX15Mdi1E__span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4317 |
+
parse/train/QgX15Mdi1E_/QgX15Mdi1E__origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4318 |
+
parse/train/rJedV3R5tm/rJedV3R5tm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4319 |
+
parse/train/rJedV3R5tm/rJedV3R5tm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4320 |
+
parse/train/rJedV3R5tm/rJedV3R5tm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4321 |
+
parse/train/BkgBvsC9FQ/BkgBvsC9FQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4322 |
+
parse/train/BkgBvsC9FQ/BkgBvsC9FQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4323 |
+
parse/train/BkgBvsC9FQ/BkgBvsC9FQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4324 |
+
parse/train/RLRXCV6DbEJ/RLRXCV6DbEJ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4325 |
+
parse/train/RLRXCV6DbEJ/RLRXCV6DbEJ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4326 |
+
parse/train/RLRXCV6DbEJ/RLRXCV6DbEJ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4327 |
+
parse/train/d7KBjmI3GmQ/d7KBjmI3GmQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4328 |
+
parse/train/d7KBjmI3GmQ/d7KBjmI3GmQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4329 |
+
parse/train/d7KBjmI3GmQ/d7KBjmI3GmQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4330 |
+
parse/train/gl3D-xY7wLq/gl3D-xY7wLq_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4331 |
+
parse/train/gl3D-xY7wLq/gl3D-xY7wLq_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4332 |
+
parse/train/gl3D-xY7wLq/gl3D-xY7wLq_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4333 |
+
parse/train/QHUUrieaqai/QHUUrieaqai_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4334 |
+
parse/train/QHUUrieaqai/QHUUrieaqai_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4335 |
+
parse/train/QHUUrieaqai/QHUUrieaqai_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4336 |
+
parse/train/S1MB-3RcF7/S1MB-3RcF7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4337 |
+
parse/train/S1MB-3RcF7/S1MB-3RcF7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4338 |
+
parse/train/S1MB-3RcF7/S1MB-3RcF7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4339 |
+
parse/train/roNqYL0_XP/roNqYL0_XP_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4340 |
+
parse/train/roNqYL0_XP/roNqYL0_XP_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4341 |
+
parse/train/roNqYL0_XP/roNqYL0_XP_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4342 |
+
parse/train/zkHlu_3sJYU/zkHlu_3sJYU_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4343 |
+
parse/train/zkHlu_3sJYU/zkHlu_3sJYU_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4344 |
+
parse/train/zkHlu_3sJYU/zkHlu_3sJYU_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4345 |
+
parse/train/SkT5Yg-RZ/SkT5Yg-RZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4346 |
+
parse/train/SkT5Yg-RZ/SkT5Yg-RZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4347 |
+
parse/train/SkT5Yg-RZ/SkT5Yg-RZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4348 |
+
parse/train/HkE0Nvqlg/HkE0Nvqlg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4349 |
+
parse/train/HkE0Nvqlg/HkE0Nvqlg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4350 |
+
parse/train/HkE0Nvqlg/HkE0Nvqlg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4351 |
+
parse/train/SyNPk2R9K7/SyNPk2R9K7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4352 |
+
parse/train/SyNPk2R9K7/SyNPk2R9K7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4353 |
+
parse/train/SyNPk2R9K7/SyNPk2R9K7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4354 |
+
parse/train/BJxkOlSYDH/BJxkOlSYDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4355 |
+
parse/train/BJxkOlSYDH/BJxkOlSYDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4356 |
+
parse/train/BJxkOlSYDH/BJxkOlSYDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4357 |
+
parse/train/XxP75wV6JGH/XxP75wV6JGH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4358 |
+
parse/train/XxP75wV6JGH/XxP75wV6JGH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4359 |
+
parse/train/XxP75wV6JGH/XxP75wV6JGH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4360 |
+
parse/train/SyfXKoRqFQ/SyfXKoRqFQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4361 |
+
parse/train/SyfXKoRqFQ/SyfXKoRqFQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4362 |
+
parse/train/SyfXKoRqFQ/SyfXKoRqFQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4363 |
+
parse/train/B1DmUzWAW/B1DmUzWAW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4364 |
+
parse/train/B1DmUzWAW/B1DmUzWAW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4365 |
+
parse/train/B1DmUzWAW/B1DmUzWAW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4366 |
+
parse/train/rJ4km2R5t7/rJ4km2R5t7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4367 |
+
parse/train/rJ4km2R5t7/rJ4km2R5t7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4368 |
+
parse/train/rJ4km2R5t7/rJ4km2R5t7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4369 |
+
parse/train/LmUJqB1Cz8/LmUJqB1Cz8_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4370 |
+
parse/train/LmUJqB1Cz8/LmUJqB1Cz8_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4371 |
+
parse/train/LmUJqB1Cz8/LmUJqB1Cz8_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4372 |
+
parse/train/S1l2IyrYPr/S1l2IyrYPr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4373 |
+
parse/train/S1l2IyrYPr/S1l2IyrYPr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4374 |
+
parse/train/S1l2IyrYPr/S1l2IyrYPr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4375 |
+
parse/train/S1GDXzb0b/S1GDXzb0b_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4376 |
+
parse/train/S1GDXzb0b/S1GDXzb0b_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4377 |
+
parse/train/S1GDXzb0b/S1GDXzb0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4378 |
+
parse/train/jM76BCb6F9m/jM76BCb6F9m_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4379 |
+
parse/train/jM76BCb6F9m/jM76BCb6F9m_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4380 |
+
parse/train/jM76BCb6F9m/jM76BCb6F9m_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4381 |
+
parse/train/rkgqN1SYvr/rkgqN1SYvr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4382 |
+
parse/train/rkgqN1SYvr/rkgqN1SYvr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4383 |
+
parse/train/rkgqN1SYvr/rkgqN1SYvr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4384 |
+
parse/train/S1fHmlbCW/S1fHmlbCW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4385 |
+
parse/train/S1fHmlbCW/S1fHmlbCW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4386 |
+
parse/train/S1fHmlbCW/S1fHmlbCW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4387 |
+
parse/train/HygQBn0cYm/HygQBn0cYm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4388 |
+
parse/train/HygQBn0cYm/HygQBn0cYm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4389 |
+
parse/train/HygQBn0cYm/HygQBn0cYm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4390 |
+
parse/train/x_n34KpwAvI/x_n34KpwAvI_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4391 |
+
parse/train/x_n34KpwAvI/x_n34KpwAvI_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4392 |
+
parse/train/x_n34KpwAvI/x_n34KpwAvI_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4393 |
+
parse/train/BkLhzHtlg/BkLhzHtlg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4394 |
+
parse/train/BkLhzHtlg/BkLhzHtlg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4395 |
+
parse/train/BkLhzHtlg/BkLhzHtlg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4396 |
+
parse/train/IMPnRXEWpvr/IMPnRXEWpvr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4397 |
+
parse/train/IMPnRXEWpvr/IMPnRXEWpvr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4398 |
+
parse/train/IMPnRXEWpvr/IMPnRXEWpvr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4399 |
+
parse/train/r1saNM-RW/r1saNM-RW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4400 |
+
parse/train/r1saNM-RW/r1saNM-RW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4401 |
+
parse/train/r1saNM-RW/r1saNM-RW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4402 |
+
parse/train/SJeLopEYDH/SJeLopEYDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4403 |
+
parse/train/SJeLopEYDH/SJeLopEYDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4404 |
+
parse/train/SJeLopEYDH/SJeLopEYDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4405 |
+
parse/train/B1ZvaaeAZ/B1ZvaaeAZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4406 |
+
parse/train/B1ZvaaeAZ/B1ZvaaeAZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4407 |
+
parse/train/B1ZvaaeAZ/B1ZvaaeAZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4408 |
+
parse/train/MlFcgL2AP4d/MlFcgL2AP4d_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4409 |
+
parse/train/MlFcgL2AP4d/MlFcgL2AP4d_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4410 |
+
parse/train/MlFcgL2AP4d/MlFcgL2AP4d_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4411 |
+
parse/train/ryl8-3AcFX/ryl8-3AcFX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4412 |
+
parse/train/ryl8-3AcFX/ryl8-3AcFX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4413 |
+
parse/train/ryl8-3AcFX/ryl8-3AcFX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4414 |
+
parse/train/rJ5C67-C-/rJ5C67-C-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4415 |
+
parse/train/rJ5C67-C-/rJ5C67-C-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4416 |
+
parse/train/rJ5C67-C-/rJ5C67-C-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4417 |
+
parse/train/TgSVWXw22FQ/TgSVWXw22FQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4418 |
+
parse/train/TgSVWXw22FQ/TgSVWXw22FQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4419 |
+
parse/train/TgSVWXw22FQ/TgSVWXw22FQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4420 |
+
parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4421 |
+
parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4422 |
+
parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4423 |
+
parse/train/BkglSTNFDB/BkglSTNFDB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4424 |
+
parse/train/BkglSTNFDB/BkglSTNFDB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4425 |
+
parse/train/BkglSTNFDB/BkglSTNFDB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4426 |
+
parse/train/rJerHlrYwH/rJerHlrYwH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4427 |
+
parse/train/rJerHlrYwH/rJerHlrYwH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4428 |
+
parse/train/rJerHlrYwH/rJerHlrYwH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4429 |
+
parse/train/Xh5eMZVONGF/Xh5eMZVONGF_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4430 |
+
parse/train/Xh5eMZVONGF/Xh5eMZVONGF_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4431 |
+
parse/train/Xh5eMZVONGF/Xh5eMZVONGF_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4432 |
+
parse/train/LAKplpLMbP8/LAKplpLMbP8_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4433 |
+
parse/train/LAKplpLMbP8/LAKplpLMbP8_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4434 |
+
parse/train/LAKplpLMbP8/LAKplpLMbP8_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4435 |
+
parse/train/xQGYquca0gB/xQGYquca0gB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4436 |
+
parse/train/xQGYquca0gB/xQGYquca0gB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4437 |
+
parse/train/xQGYquca0gB/xQGYquca0gB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4438 |
+
parse/train/B1hdzd5lg/B1hdzd5lg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4439 |
+
parse/train/B1hdzd5lg/B1hdzd5lg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4440 |
+
parse/train/B1hdzd5lg/B1hdzd5lg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4441 |
+
parse/train/SJ3rcZcxl/SJ3rcZcxl_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4442 |
+
parse/train/SJ3rcZcxl/SJ3rcZcxl_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4443 |
+
parse/train/SJ3rcZcxl/SJ3rcZcxl_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4444 |
+
parse/train/ByxZX20qFQ/ByxZX20qFQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4445 |
+
parse/train/ByxZX20qFQ/ByxZX20qFQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4446 |
+
parse/train/ByxZX20qFQ/ByxZX20qFQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4447 |
+
parse/train/ryxdEkHtPS/ryxdEkHtPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4448 |
+
parse/train/ryxdEkHtPS/ryxdEkHtPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4449 |
+
parse/train/ryxdEkHtPS/ryxdEkHtPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4450 |
+
parse/train/r1xwA34KDB/r1xwA34KDB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4451 |
+
parse/train/r1xwA34KDB/r1xwA34KDB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4452 |
+
parse/train/r1xwA34KDB/r1xwA34KDB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4453 |
+
parse/train/HJxB5sRcFQ/HJxB5sRcFQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4454 |
+
parse/train/HJxB5sRcFQ/HJxB5sRcFQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4455 |
+
parse/train/HJxB5sRcFQ/HJxB5sRcFQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4456 |
+
parse/train/rkgz2aEKDr/rkgz2aEKDr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4457 |
+
parse/train/rkgz2aEKDr/rkgz2aEKDr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4458 |
+
parse/train/rkgz2aEKDr/rkgz2aEKDr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4459 |
+
parse/train/r1-4BLaQz/r1-4BLaQz_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4460 |
+
parse/train/r1-4BLaQz/r1-4BLaQz_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4461 |
+
parse/train/r1-4BLaQz/r1-4BLaQz_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4462 |
+
parse/train/SJyVzQ-C-/SJyVzQ-C-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4463 |
+
parse/train/SJyVzQ-C-/SJyVzQ-C-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4464 |
+
parse/train/SJyVzQ-C-/SJyVzQ-C-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4465 |
+
parse/train/S1Euwz-Rb/S1Euwz-Rb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4466 |
+
parse/train/S1Euwz-Rb/S1Euwz-Rb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4467 |
+
parse/train/S1Euwz-Rb/S1Euwz-Rb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4468 |
+
parse/train/S1sqHMZCb/S1sqHMZCb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4469 |
+
parse/train/S1sqHMZCb/S1sqHMZCb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4470 |
+
parse/train/S1sqHMZCb/S1sqHMZCb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4471 |
+
parse/train/r17lFgZ0Z/r17lFgZ0Z_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4472 |
+
parse/train/r17lFgZ0Z/r17lFgZ0Z_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4473 |
+
parse/train/r17lFgZ0Z/r17lFgZ0Z_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4474 |
+
parse/train/xGZG2kS5bFk/xGZG2kS5bFk_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4475 |
+
parse/train/xGZG2kS5bFk/xGZG2kS5bFk_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4476 |
+
parse/train/xGZG2kS5bFk/xGZG2kS5bFk_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4477 |
+
parse/train/r1l-VeSKwS/r1l-VeSKwS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4478 |
+
parse/train/r1l-VeSKwS/r1l-VeSKwS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4479 |
+
parse/train/r1l-VeSKwS/r1l-VeSKwS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4480 |
+
parse/train/rye7knCqK7/rye7knCqK7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4481 |
+
parse/train/rye7knCqK7/rye7knCqK7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4482 |
+
parse/train/rye7knCqK7/rye7knCqK7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4483 |
+
parse/train/03x6x6qNwJ3/03x6x6qNwJ3_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4484 |
+
parse/train/03x6x6qNwJ3/03x6x6qNwJ3_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4485 |
+
parse/train/03x6x6qNwJ3/03x6x6qNwJ3_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4486 |
+
parse/train/S1DWPP1A-/S1DWPP1A-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4487 |
+
parse/train/S1DWPP1A-/S1DWPP1A-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4488 |
+
parse/train/S1DWPP1A-/S1DWPP1A-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4489 |
+
parse/train/vOchfRdvPy7/vOchfRdvPy7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4490 |
+
parse/train/vOchfRdvPy7/vOchfRdvPy7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4491 |
+
parse/train/vOchfRdvPy7/vOchfRdvPy7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4492 |
+
parse/train/HkgsPhNYPS/HkgsPhNYPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4493 |
+
parse/train/HkgsPhNYPS/HkgsPhNYPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4494 |
+
parse/train/HkgsPhNYPS/HkgsPhNYPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4495 |
+
parse/train/RzYrn625bu8/RzYrn625bu8_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4496 |
+
parse/train/RzYrn625bu8/RzYrn625bu8_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4497 |
+
parse/train/RzYrn625bu8/RzYrn625bu8_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4498 |
+
parse/train/heqv8eIweMY/heqv8eIweMY_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4499 |
+
parse/train/heqv8eIweMY/heqv8eIweMY_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4500 |
+
parse/train/heqv8eIweMY/heqv8eIweMY_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4501 |
+
parse/train/Pk972F7J8Um/Pk972F7J8Um_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4502 |
+
parse/train/Pk972F7J8Um/Pk972F7J8Um_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4503 |
+
parse/train/Pk972F7J8Um/Pk972F7J8Um_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4504 |
+
parse/train/dvyUaK4neD0/dvyUaK4neD0_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4505 |
+
parse/train/dvyUaK4neD0/dvyUaK4neD0_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4506 |
+
parse/train/dvyUaK4neD0/dvyUaK4neD0_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4507 |
+
parse/train/o2tx_m7hK3t/o2tx_m7hK3t_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4508 |
+
parse/train/o2tx_m7hK3t/o2tx_m7hK3t_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4509 |
+
parse/train/o2tx_m7hK3t/o2tx_m7hK3t_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4510 |
+
parse/train/By-7dz-AZ/By-7dz-AZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4511 |
+
parse/train/By-7dz-AZ/By-7dz-AZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4512 |
+
parse/train/By-7dz-AZ/By-7dz-AZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4513 |
+
parse/train/Sklgs0NFvr/Sklgs0NFvr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4514 |
+
parse/train/Sklgs0NFvr/Sklgs0NFvr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4515 |
+
parse/train/Sklgs0NFvr/Sklgs0NFvr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4516 |
+
parse/train/BylQSxHFwr/BylQSxHFwr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4517 |
+
parse/train/BylQSxHFwr/BylQSxHFwr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4518 |
+
parse/train/BylQSxHFwr/BylQSxHFwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4519 |
+
parse/train/HysBZSqlx/HysBZSqlx_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4520 |
+
parse/train/HysBZSqlx/HysBZSqlx_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 4521 |
+
parse/train/HysBZSqlx/HysBZSqlx_span.pdf filter=lfs diff=lfs merge=lfs -text
|
parse/dev/2nJdh_C-UWe/2nJdh_C-UWe.md
ADDED
|
@@ -0,0 +1,413 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Towards Effective and Interpretable Human-AI Collaboration in MOBA Games
|
| 2 |
+
|
| 3 |
+
Anonymous Author(s)
|
| 4 |
+
Affiliation
|
| 5 |
+
Address
|
| 6 |
+
email
|
| 7 |
+
|
| 8 |
+
# Abstract
|
| 9 |
+
|
| 10 |
+
1 MOBA games, e.g., Dota2 and Honor of Kings, have been actively used as the
|
| 11 |
+
2 testbed for the recent AI research on games, and various AI systems have been
|
| 12 |
+
3 developed at the human level so far. However, these AI systems merely focus on
|
| 13 |
+
4 how to compete with humans, less exploring how to collaborate with humans. To
|
| 14 |
+
5 this end, this paper makes the first attempt to investigate human-AI collaboration in
|
| 15 |
+
6 MOBA games. In this paper, we propose to enable humans and agents to collaborate
|
| 16 |
+
7 through explicit communications by designing an efficient and interpretable Meta
|
| 17 |
+
8 Command Communication-based framework, dubbed MCC, for accomplishing
|
| 18 |
+
9 effective human-AI collaboration in MOBA games. The MCC framework consists
|
| 19 |
+
10 of two pivotal modules: 1) an interpretable communication protocol, i.e., the
|
| 20 |
+
11 Meta-Command, to bridge the communication gap between humans and agents;
|
| 21 |
+
12 2) a meta-command value estimation model, i.e., the Meta-Command Selector,
|
| 22 |
+
13 to select a valuable meta-command for each agent to achieve effective human-AI
|
| 23 |
+
14 collaboration. Experimental results in Honor of Kings demonstrate that MCC
|
| 24 |
+
15 agents can collaborate reasonably well with human teammates and even generalize
|
| 25 |
+
16 to collaborate with different levels and numbers of human teammates. Videos are
|
| 26 |
+
17 available at https://sites.google.com/view/mcc-demo.
|
| 27 |
+
|
| 28 |
+
# 18 1 Introduction
|
| 29 |
+
|
| 30 |
+
19 Games, as the microcosm of real-world problems, have been widely used as testbeds to evaluate
|
| 31 |
+
20 the performance of Artificial Intelligence (AI) techniques for decades. Recently, many researchers
|
| 32 |
+
21 focus on developing various human-level AI systems for complex games, such as board games like
|
| 33 |
+
22 Go [27, 28], First-Person Shooting (FPS) games like ViZDoom [14], Real-Time Strategy (RTS)
|
| 34 |
+
23 games like StarCraft 2 [34], and Multi-player Online Battle Arena (MOBA) games like Dota 2 [22].
|
| 35 |
+
24 However, these AI systems focus merely on how to compete instead of collaborating with humans,
|
| 36 |
+
25 leaving Human-AI Collaboration (HAC) in complex environments still to be investigated.
|
| 37 |
+
26 In this paper, we study the HAC problem in complex MOBA games, which is characterized by multi
|
| 38 |
+
27 agent cooperation and competition mechanisms, long time horizons, enormous state-action spaces
|
| 39 |
+
28 $( \bar { 1 0 } ^ { 2 0 0 0 0 } )$ , and imperfect information [22, 26, 38]. HAC requires the agent to collaborate reasonably
|
| 40 |
+
29 with various human teammates. One straightforward approach is to improve the generalization of
|
| 41 |
+
30 agents, that is, to collaborate with an enough diverse population of teammates during training. There
|
| 42 |
+
31 are some Population-Based Training (PBT) based algorithms and learning systems [1, 2, 10, 11,
|
| 43 |
+
32 31, 41] proposed to improve the generalization of agents in video games by constructing a diverse
|
| 44 |
+
33 population of agents in different ways. However, this approach requires a vast amount of diverse data
|
| 45 |
+
34 and massive computing resources, posing a big computational obstacle for complex MOBA games.
|
| 46 |
+
35 Human team success in MOBA games requires not only subtle individual micro-operations but also
|
| 47 |
+
36 excellent communications and collaborations among teammates on macro-strategies, i.e., long-term
|
| 48 |
+
37 intentions [8, 37]. Consequently, we focus on enabling humans and agents to collaborate through
|
| 49 |
+
38 explicit communications and propose an efficient and interpretable Meta-Command Communication
|
| 50 |
+
39 based human-AI collaboration framework, dubbed MCC, to solve the HAC problem in MOBA
|
| 51 |
+
40 games. First, we design an interpretable communication protocol, i.e., the Meta-Command, as a
|
| 52 |
+
41 general representation of macro-strategies to bridge the communication gap between agents and
|
| 53 |
+
42 humans. Both macro-strategies sent by humans and messages outputted by agents can be converted
|
| 54 |
+
43 into unified meta-commands (see Figure 1). Second, following Gao et al. [8], we construct a
|
| 55 |
+
44 hierarchical model that includes the command encoding network (macro-strategy layer) and the
|
| 56 |
+
45 meta-command conditioned action network (micro-action layer), used for agents to generate and
|
| 57 |
+
46 execute meta-commands, respectively. Third, we propose a meta-command value estimation model,
|
| 58 |
+
47 i.e., the Meta-Command Selector, to select the optimal meta-command for each agent to execute.
|
| 59 |
+
48 The training process of the MCC framework consists of three phases. We first train the command
|
| 60 |
+
49 encoding network to learn the distribution of meta-commands sent from humans. Afterward, we
|
| 61 |
+
50 train the meta-command conditioned action network to ensure that the agent has the near-human
|
| 62 |
+
51 completion rate for meta-commands. Finally, we train the meta-command selector to ensure that the
|
| 63 |
+
52 agent can select a valuable meta-command to achieve effective collaboration. We train and evaluate
|
| 64 |
+
53 the agent in Honor of Kings 5v5 mode with a full hero pool (over 100 heroes). Experimental results
|
| 65 |
+
54 demonstrate the effectiveness of the MCC framework. In general, our contributions are as follows:
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
Figure 1: MOBA game-related introduction. (a) Key elements of MOBA games such as Dota 2, Honor of Kings, etc. Players observe from the state of the environment, make micro-operations and macro-strategies decisions, and collaborate through explicit messages (e.g.,text and signals). (b) Example of collaboration via meta-commands. The Come And Kill The Dragon is more valuable for humans A and B and agent $_ \mathrm { D }$ t o collaborate, while the Clean Up Top-Lane Minions is more valuable for human C and agent E to collaborate.
|
| 69 |
+
|
| 70 |
+
• To the best of our knowledge, we are the first to investigate the HAC problem in MOBA games. We propose an efficient and interpretable Meta-Command Communication-based framework dubbed MCC to achieve effective human-AI collaboration in MOBA games.
|
| 71 |
+
|
| 72 |
+
• We design an interpretable communication protocol to bridge the communication gap between humans and agents. In addition, we propose a meta-command value estimation model to select a valuable meta-command for each agent to achieve effective human-AI collaboration.
|
| 73 |
+
|
| 74 |
+
• We introduce the training process of the MCC framework in a typical MOBA game Honor of Kings and evaluate it in practical human-AI game tests. Experimental results show that MCC agents can reasonably collaborate with different levels and numbers of human teammates.
|
| 75 |
+
|
| 76 |
+
# 64 2 Related Work
|
| 77 |
+
|
| 78 |
+
# 2.1 MOBA Games AI Research
|
| 79 |
+
|
| 80 |
+
66 MOBA games, such as Dota 2 and Honor of Kings, have attracted much attention from AI researchers
|
| 81 |
+
67 due to their multi-agent cooperative and competitive mechanics, long time horizons, partial observa
|
| 82 |
+
68 tion, and enormous state-action spaces [22, 38]. Recently, OpenAI et al. [22] introduced an AI system
|
| 83 |
+
69 named OpenAI-Five that defeated professional players in Dota 2 5v5 mode under the condition of
|
| 84 |
+
70 limited heroes. Ye et al. [38, 39, 40] proposed another learning system named WuKong that can
|
| 85 |
+
71 surpass top e-sport players in Honor of Kings with a full hero pool. Further, Wu [37] and Gao et
|
| 86 |
+
72 al. [8] proposed learning systems that enable the agent to learn human strategies to achieve policy
|
| 87 |
+
73 diversity. However, these AI systems can only defeat human players but cannot collaborate well due
|
| 88 |
+
74 to the communication gap between agents and humans, see Table 1. In most real-world scenarios, the
|
| 89 |
+
75 excellent collaboration between humans and agents may make more sense than the competition.
|
| 90 |
+
|
| 91 |
+
PBT is considered one way to solve the HAC problem [4]. Most PBT-based methods are devoted to training an agent which can be compatible with unseen partners by maintaining a population of agents with diverse behaviors in different ways [1, 2, 10, 11, 31, 41][6, 19, 20, 30]. These methods have been validated on both objective and subjective metrics in video games Overcooked and Capture the Flag and card game Hanabi. However, the main difference between these games and MOBA games is that these games do not provide explicit communication mechanics for collaboration on macrostrategies between agents and humans. Besides, MOBA AI agents usually need to learn billions of network parameters to cope with the enormous state-action spaces $( 1 0 ^ { 2 0 0 0 0 } )$ [38], which constitutes a prohibitive computational burden for learning. As a more realistic topic of HAC, human-robot interaction in manufacturing also attracts much attention [13, 17, 25]. However, these studies are mainly limited to collaboration between a robot and a human through one-way communication, i.e., humans give robots orders. Therefore, there is still a large room to study RL with the participation of humans. This work can be a stepping stone for broader real-world applications.
|
| 92 |
+
|
| 93 |
+
# 2.3 Multi-Agent Communication
|
| 94 |
+
|
| 95 |
+
Communication is often used in Multi-Agent Reinforcement Learning (MARL) to improve interagent collaboration. Most communication-based MARL methods are mainly focused on exploring communication protocols between multiple agents with an end-to-end RL framework [5, 7, 9, 23, 29, 32, 36]. Jiang and Lu [12] and Kim et al. [15] proposed to model the value of multi-agent communication for effective collaboration. Unfortunately, these methods all model communications in a latent space without considering human-AI interactions, making it less interpretable to humans. Instead, we focus on enabling humans and agents to collaborate through explicit communications.
|
| 96 |
+
|
| 97 |
+
# 3 Human-AI Collaboration
|
| 98 |
+
|
| 99 |
+
We consider an interpretable communicative human-AI collaboration task, which can be extended from Partially Observable Markov Decision Process (POMDP) and formulated as a tuple $< N , H , { \bf S } , { \bf A } ^ { N } , { \bf A } ^ { H } , { \bf O } , { \bf M } , r , P , \gamma >$ , where $N$ and $H$ represent the numbers of agents and humans, respectively. S is the space of global states. $\mathbf { A } ^ { N } = \{ A _ { i } ^ { N } \} _ { i = 1 , \dots , N }$ and $\mathbf { A } ^ { H } = \{ A _ { i } ^ { \breve { H } } \} _ { i = 1 , \dots , H }$ denote the spaces of actions of $N$ agents and $H$ humans, respectively. $\mathbf { O } = \{ O _ { i } \} _ { i = 1 , \dots , N + H }$ denotes the space of observations of $N$ agents and $H$ humans. $\mathbf { M }$ represents the space of interpretable messages, that is, the Meta-Commands in the MCC framework. $P : \mathbf { S } \times \bar { \mathbf { A } } ^ { N } \times \mathbf { A } ^ { H } \vec { \mathbf { \Lambda } } \vec { \mathbf { \Lambda } } \vec { \mathbf { \Lambda } } $ and $r : \mathbf { S } \times \mathbf { A } ^ { N } \times \mathbf { A } ^ { H } \to \mathbb { R }$ denote the shared state transition probability function and reward function of $N$ agents, respectively. Note that, $r$ includes both individual reward and team reward. $\gamma \in [ 0 , 1 )$ denotes the discount factor. For each agent $i$ in state $s _ { t } \in \mathbf { S }$ , it receives an observation $o _ { t } ^ { i } \in O _ { i }$ and a selected message $c _ { t } ^ { i } \in \mathbf { M }$ , and then outputs an action $a _ { t } ^ { i } = \pi _ { \theta } ( o _ { t } ^ { i } , c _ { t } ^ { i } ) \in A _ { i } ^ { N }$ and a new message $m _ { t + 1 } ^ { i } = \pi _ { \phi } \bar { ( } o _ { t } ^ { i } ) \bar { \ } \in \bf { M }$ , where $\pi _ { \theta }$ and $\pi _ { \phi }$ are action network and message encoding network, respectively. A message selector $c _ { t } ^ { i } = \pi _ { \omega } ( o _ { t } ^ { i } , C _ { t } )$ is introduced to receive a message set $C _ { t } = \{ m _ { t } ^ { i } \} _ { i = 1 , \dots , N + H } \subset \mathbf { M }$ from all agents and humans and select the optimal one to execute.
|
| 100 |
+
|
| 101 |
+
We divide the HAC problem in MOBA games into the Human-to-AI (H2A) and the AI-to-Human (A2H) scenarios. The H2A Scenario: Humans send macro-strategies as messages to agent teammates, and agents combine them with their own messages to select the optimal one based on their own message selector to execute, achieving effective collaboration with humans. The A2H Scenario: Agents send messages as macro-strategies to human teammates, and humans combine them with their own macro-strategies to select the optimal one based on their own value systems to execute, achieving effective collaboration with agents. The goal of both tasks is that agents and humans communicate macro-strategies with pre-defined communication protocols, and then select valuable macro-strategies for effective collaboration to win the game.
|
| 102 |
+
|
| 103 |
+
# 4 Meta-Command Communication-Based Framework
|
| 104 |
+
|
| 105 |
+
123 In this section, we present the proposed MCC framework in detail. We first briefly describe three key
|
| 106 |
+
124 stages of the MCC framework (see Section 4.1). Then we introduce the two pivotal modules in the
|
| 107 |
+
125 MCC framework: 1) an interpretable communication protocol, i.e., the Meta-Command, as a general
|
| 108 |
+
126 representation of macro-strategies to bridge the communication gap between agents and humans (see
|
| 109 |
+
127 Section 4.2); 2) a meta-command value estimation model, i.e., the Meta-Command Selector, to select
|
| 110 |
+
128 a valuable meta-command for each agent to achieve effective HAC in MOBA games(see Section 4.3).
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
Figure 2: The temporal process of the MCC framework. For each communication step $t$ and $T$ ), MCC first (I) converts messages from humans and agents into meta-commands, then (II) selects the optimal meta-command for each agent to execute, and (III) finally predicts a sequence of actions for each agent to perform. The selected meta-command is retained and executed for $_ n$ time steps. This process is repeated until the end of a game.
|
| 114 |
+
|
| 115 |
+
# 4.1 Overview
|
| 116 |
+
|
| 117 |
+
0 The flow of the MCC framework can be divided into three stages: the meta-command conversion stage,
|
| 118 |
+
the meta-command communication stage, and the human-AI collaboration stage, as plotted in Figure 2.
|
| 119 |
+
2 At the Meta-Command Conversion Stage, the MCC framework converts the macro-strategies sent
|
| 120 |
+
33 by humans and the messages outputted by the command encoding network of agents into unified
|
| 121 |
+
34 meta-commands and then broadcasts them to all agents and humans. At the Meta-Command
|
| 122 |
+
35 Communication Stage, the MCC framework uses the meta-command selector to estimate the values
|
| 123 |
+
36 of all received meta-commands and select the optimal one for each agent to execute. Note that
|
| 124 |
+
37 humans also select the optimal meta-command based on their value systems. At the Human-AI
|
| 125 |
+
38 Collaboration Stage, the MCC framework adopts the meta-command conditioned action network to
|
| 126 |
+
39 predict a sequence of actions for each agent to perform based on its selected meta-command. For
|
| 127 |
+
40 each game, humans and agents have to collaborate multiple times, that is, they need to perform the
|
| 128 |
+
above three stages multiple times to win the game.
|
| 129 |
+
|
| 130 |
+
# 142 4.2 Meta-Command
|
| 131 |
+
|
| 132 |
+
143 In MOBA games, we propose that a macro-strategy consists of three components: where to go, what
|
| 133 |
+
144 to do, and how long. For example, a macro-strategy can be Come And Kill The Dragon, which
|
| 134 |
+
145 consists of Come To The Dragon (where to go), Attack The Dragon (what to do), and Until The
|
| 135 |
+
146 Dragon Is Killed (how long). Thus, we propose a general representation of macro-strategies, i.e.,
|
| 136 |
+
147 the Meta-Command, as an interpretable communication protocol to bridge the communication gap
|
| 137 |
+
148 between agents and humans.
|
| 138 |
+
149 Meta-Command Definition. We formulate the Meta-Command as a tuple $< L , E , T ^ { m c } >$ , as shown
|
| 139 |
+
150 in Figure 1(b), where $L$ is the Location to go, $E$ is the Event to do after reaching $L$ , and $T ^ { m c }$ is the
|
| 140 |
+
151 Time Limit for executing the meta-command. Among them, $L$ is the key to the meta-command, which
|
| 141 |
+
152 contains the intention of the macro-strategy. $E$ can be thought of as human micro-operation, which is
|
| 142 |
+
153 implemented through a pre-trained micro-action network $\pi _ { \theta }$ in the MCC framework. $T ^ { m c }$ can be set
|
| 143 |
+
154 to how long it normally takes a human to complete a macro-strategy in MOBA games, usually 20
|
| 144 |
+
155 seconds corresponds to $80 \%$ completion rate for meta-commands, see Appendix A.12.1.
|
| 145 |
+
156 Meta-Command Conversion. To realize interpretable human-AI communication, we convert the
|
| 146 |
+
157 explicit messages from humans and the implicit messages from agents into unified meta-commands.
|
| 147 |
+
158 To achieve the former, a hand-crafted command converter function $f ^ { c c }$ is used to generate $L$ of meta
|
| 148 |
+
159 commands by extracting the location from explicit messages, such as text and signals, sent by humans.
|
| 149 |
+
160 To achieve the latter, we use a Command Encoding Network (CEN) $\pi _ { \phi } ( m | o )$ to generate $L$ of meta
|
| 150 |
+
161 commands. The CEN is trained via supervised learning (SL) with the goal of learning the distribution
|
| 151 |
+
162 of meta-commands sent from humans, as shown in Figure 3(a)(I). The training dataset $\{ < o , m > \}$
|
| 152 |
+
163 is obtained by extracting the observation $o$ and its corresponding meta-command $m$ from expert data.
|
| 153 |
+
164 After converting all messages into unified meta-commands, the MCC framework broadcasts them to
|
| 154 |
+
165 all agents and humans. Then, agents and humans receive an identical meta-command candidate set.
|
| 155 |
+
166 Meta-Command Execution. After receiving a meta-command candidate set, agents can se
|
| 156 |
+
167 lect one meta-command from it to execute. We adopt a Meta-Command Conditioned Ac
|
| 157 |
+
168 tion Network (MCCAN) $\pi _ { \boldsymbol { \theta } } ( a | o , m )$ for agents to perform actions based on the selected meta
|
| 158 |
+
169 command, as shown in Figure 3(a)(II). The MCCAN is trained via goal-conditioned RL with
|
| 159 |
+
170 the goal of achieving a near-human completion rate for the meta-commands generated by the
|
| 160 |
+
171 pre-trained CEN while ensuring that the win rate is not reduced. We adopt an intrinsic reward
|
| 161 |
+
172 $\begin{array} { r } { \dot { r } _ { t } ^ { i n t } ( s _ { t } , m _ { t } , s _ { t + 1 } ) = \left| f ^ { c e } ( s _ { t } ) - \check { m } _ { t } \right| - \left| f ^ { c e } ( s _ { t + 1 } ) - m _ { t } \right| } \end{array}$ to guide the process of executing the meta
|
| 162 |
+
173 command $m _ { t }$ , where $f ^ { c e }$ is a hand-crafted command extraction function. We train the MCCAN
|
| 163 |
+
174 with the objective of maximizing the expectation over extrinsic and intrinsic discounted total re
|
| 164 |
+
175 wards $\begin{array} { r } { G _ { t } = \mathbb { E } _ { s \sim d _ { \pi _ { \theta } } , a \sim \pi _ { \theta } } \left[ \sum _ { i = 0 } ^ { \infty } \gamma ^ { i } r _ { t + i } + \alpha \sum _ { j = 0 } ^ { T ^ { m c } } \gamma ^ { j } r _ { t + j } ^ { i n t } \right] , } \end{array}$ , where $\alpha$ is a trade-off parameter and
|
| 165 |
+
176 $\begin{array} { r } { d _ { \pi } ( s ) = \operatorname* { l i m } _ { t \infty } P ( s _ { t } = \bar { s } \mid s _ { 0 } , \pi ) } \end{array}$ is the probability when following $\pi$ for $t$ steps from $s _ { 0 }$ .
|
| 166 |
+
177 After training the CEN and MCCAN, we can achieve HAC by simply setting an agent to randomly
|
| 167 |
+
178 select a meta-command derived from humans to execute. However, such collaboration is non
|
| 168 |
+
179 intelligent and can even be a disaster for game victory because agents have no mechanism to
|
| 169 |
+
180 model the values of meta-commands and cannot choose the optimal meta-command to execute.
|
| 170 |
+
181 While humans usually choose the optimal one based on their value systems for achieving effective
|
| 171 |
+
182 collaboration to win the game. Thus, we further propose a meta-command value estimation model to
|
| 172 |
+
183 select a valuable meta-command for each agent, as described in the following subsection.
|
| 173 |
+
|
| 174 |
+
|
| 175 |
+
Figure 3: The training process and model structure of MCC. (a) The training process is divided into three phases: we first (I) train the CEN via supervised learning (SL), then (II) train the MCCAN via goal-conditioned RL, and finally (III) train the CS via RL. Among them, the dashed box represents the frozen model. (b) The detailed CS model structure, including CNN feature extraction, gating mechanism, target attention module, etc.
|
| 176 |
+
|
| 177 |
+
# 184 4.3 Meta-Command Selector
|
| 178 |
+
|
| 179 |
+
185 In real-world MOBA games, the same macro-strategy often has different values for different humans
|
| 180 |
+
186 in different situations. For example, a macro-strategy can be Come And Kill The Dragon, as shown in
|
| 181 |
+
187 Figure 1(b). It is more valuable for humans A and B to collaborate. While another macro-strategy can
|
| 182 |
+
188 be Clean Up Top-Lane Minions, which is more valuable for human C rather than humans A and B.
|
| 183 |
+
189 Therefore, it is important to select the most valuable meta-command from the received meta-command
|
| 184 |
+
190 candidate set $C$ to achieve effective human-AI collaboration. We propose a meta-command value
|
| 185 |
+
191 estimation model, i.e., the Meta-Command Selector (CS) $\pi _ { \omega } ( o , C )$ , to estimate the values of all
|
| 186 |
+
192 current meta-commands and select the most valuable one for each agent to execute.
|
| 187 |
+
|
| 188 |
+
193 CS Optimization Objective. Typically, the execution of a meta-command involves reaching location
|
| 189 |
+
194 $L$ and doing event $E$ , of which the latter is more important to the value of the meta-command.
|
| 190 |
+
195 For example, for the meta-command Come And Kill The Dragon, if Kill The Dragon event cannot
|
| 191 |
+
196 be done within $T ^ { m c }$ time steps, then it is pointless to Come To The Dragon. Thus, the long-term
|
| 192 |
+
reward 197 steps b198 $R ^ { m c }$ for executinracting with a mete en manent: ds withi, where $T ^ { m c }$ $\begin{array} { r } { R _ { t } ^ { m c } = \sum _ { i = 0 } ^ { T ^ { L } } r _ { t + i } + \beta \sum _ { j = T ^ { L } } ^ { T ^ { m c } } r _ { t + j } } \end{array}$ $T ^ { L } < T ^ { m c }$ is the time for reaching $L$ $\beta > 1$ is a trade-off parameter. Note that the reward function $r$
|
| 193 |
+
|
| 194 |
+
200 includes both individual rewards and team rewards. The optimization objective of CS is to select
|
| 195 |
+
201 the optimal meta-command $m _ { t } ^ { * } = \pi _ { \omega } ( o _ { t } , C _ { t } )$ for each agent to maximize the expected discounted
|
| 196 |
+
202 meta-command execution return Gmct = Es∼dπ ,m∼πω,a∼πθ $\begin{array} { r } { G _ { t } ^ { m c } = \mathbb { E } _ { s \sim d _ { \pi _ { \theta } } , m \sim \pi _ { \omega } , a \sim \pi _ { \theta } } \left[ \sum _ { i = 0 } ^ { \infty } \gamma _ { m c } ^ { i } R _ { t + i \cdot T ^ { m c } } ^ { m c } \right] } \end{array}$ , where $o _ { t } \in \mathbf { O }$ ,
|
| 197 |
+
203 $C _ { t }$ is the meta-command candidate set in state $s _ { t }$ , and $\gamma _ { m c } \in [ 0 , 1 )$ is the discount factor.
|
| 198 |
+
204 CS Training Process. We construct a self-play training environment for CS where agents can send
|
| 199 |
+
205 messages to each other. Specifically, three tricks in Figure 3(a)(III) are adopted to increase the
|
| 200 |
+
206 sample efficiency while ensuring efficient exploration. First, each sent meta-command $m$ is sampled
|
| 201 |
+
207 with the argmax rule from the results predicted by the pre-trained CEN. Second, each agent sends
|
| 202 |
+
208 its meta-command with a probability $p$ every $T ^ { m c }$ time steps. Finally, each agent selects the final
|
| 203 |
+
209 meta-command $c$ sampled with the softmax rule from its CS output results and hands it over to the
|
| 204 |
+
210 pre-trained MCCAN for execution. We use the multi-head value mechanism [38] to model the value
|
| 205 |
+
211 of the meta-command execution, and the corresponding value loss can be formulated as:
|
| 206 |
+
|
| 207 |
+
$$
|
| 208 |
+
L ^ { V } ( \omega ) = \mathbb { E } _ { S , C } \left[ \sum _ { h e a d _ { k } } \| G _ { k } ^ { m c } - V _ { \omega } ^ { k } ( S , C ) \| _ { 2 } \right] ,
|
| 209 |
+
$$
|
| 210 |
+
|
| 211 |
+
212 where 213 $V _ { \omega } ^ { k } ( S , C )$ is the value of the $k$ -th head. For DQN-based methods [21, 33, 35], the $Q$ loss is:
|
| 212 |
+
|
| 213 |
+
$$
|
| 214 |
+
L ^ { Q } ( \omega ) = \mathbb { E } _ { S , C , M } \left[ \Vert G _ { t o t a l } - Q _ { \omega } ^ { k } ( S , C , M ) \Vert _ { 2 } \right] , G _ { t o t a l } = \sum _ { h e a d _ { k } } w _ { k } G _ { k } ^ { m c } ,
|
| 215 |
+
$$
|
| 216 |
+
|
| 217 |
+
where $w _ { k }$ is the weight of the $k$ -th head and $G _ { k } ^ { m c }$ is the Temporal Difference (TD) estimated value error $R _ { k } ^ { m c } + \gamma _ { m c } V _ { \omega } ^ { k } ( S ^ { \prime } , C ^ { \prime } ) - V _ { \omega } ^ { k } ( S , C )$ .
|
| 218 |
+
|
| 219 |
+
CS Model Structure. We design a general network structure for CS towards MOBA games, as shown in Figure 3(b). In MOBA games, the meta-commands corresponding to adjacent regions usually have similar values. Thus, we divide the meta-commands in the map into grids, a common location description for MOBA games, and use the shared Convolutional Neural Network (CNN) to extract region-related information from the meta-commands to improve the generalization of CS to adjacent meta-commands. Besides, we use the gating mechanism [18] to fuse the map embedding of all received meta-commands and the state embedding of the observation information. Finally, to directly construct the relationship between the observation information and each meta-command, we introduce a target attention module, where the query is the fused embedding $h$ and the key is the map embedding $m ^ { \prime }$ of each meta-command. The fused embedding $h$ is used as the input into the subsequent Q network $Q ( h , m ^ { \prime } )$ and $\mathrm { v }$ network $V ( h )$ network of CS. In this way, the Q network can also be easily converted to the policy network $\pi ( \boldsymbol { m } | \boldsymbol { h } , \boldsymbol { m } ^ { \prime } )$ . Thus, the CS model structure can be easily applied to most popular RL algorithms, such as PPO [24], DQN [21], etc.
|
| 220 |
+
|
| 221 |
+
# 5 Experiments
|
| 222 |
+
|
| 223 |
+
We evaluate the proposed MCC framework in Honor of Kings, one of the most popular MOBA games worldwide, which has been actively used as the testbed for recent game AI research [8, 37–40]. We conduct all experiments in Honor of Kings 5v5 mode with a full hero pool (over 100 heroes), except ablation studies with a 20 hero pool for exploring the influence of different model components more sufficiently and efficiently.
|
| 224 |
+
|
| 225 |
+
# 5.1 Experimental Setup
|
| 226 |
+
|
| 227 |
+
# 5.1.1 Training Setup 1
|
| 228 |
+
|
| 229 |
+
Due to the complexity of MOBA games and limited resources, instead of training jointly, we train the CEN, MCCAN, and CS sequentially. For all model training, the location $L$ of meta-commands in the map is divided into 144 grids. The time limit $T ^ { m c }$ for the meta-command execution is set to 20s.
|
| 230 |
+
|
| 231 |
+
CEN Training Settings. We train the CEN via SL until it converges for 26 hours using 8 NVIDIA P40 GPUs. The batch size of each GPU is set to 512. Adam[16] is adopted as the optimizer with an initial learning rate of 0.0001.
|
| 232 |
+
|
| 233 |
+
MCCAN Training Settings. We train the MCCAN by finetuning a pre-trained micro-action network [38], the state-of-the-art (SOTA) model in Honor of Kings, which is conditioned on the meta-command sampled from the pre-trained CEN. The MCCAN is trained until it converges for 48
|
| 234 |
+
|
| 235 |
+
|
| 236 |
+
Figure 4: Communication environments in the experiment. The orange arrows indicate sending metacommands, and the blue arrows indicate receiving meta-commands. The dashed line denotes sending metacommands with probability $p$ .
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
Figure 5: AI performance in the testing environments. (a) and (b) show the win rate maps of different agents who play against each other. (c) shows the final Elo scores of these agents.
|
| 240 |
+
|
| 241 |
+
247 hours using a physical computer cluster with 63,000 CPUs and 560 NVIDIA V100 GPUs. The batch
|
| 242 |
+
248 size of each GPU is set to 256.
|
| 243 |
+
|
| 244 |
+
CS Training Settings. We train the CS via self-play until it converges for 24 hours using a physical computer cluster with 70,000 CPUs and 680 NVIDIA V100 GPUs. The batch size of each GPU is set to 256. The parameter $\beta$ is set to 2. Each agent sends a meta-command with a probability $p$ of 0.8 and an interval $T ^ { m c }$ of 20s, as shown in Figure 4(a).
|
| 245 |
+
|
| 246 |
+
# 5.1.2 Evaluating Setup
|
| 247 |
+
|
| 248 |
+
Our primary concern is whether the agents trained with the MCC framework, briefly called the MCC agents, can collaborate with humans well. However, evaluating agents with humans is expensive, which is not conducive to model selection and iteration. Therefore, we built two agent-only testing environments: Test I and Test II, for the model selection and iteration process, as shown in Figure 4(b). We also evaluate the MCC agents in practical human-AI game tests to examine the performance of collaborating with humans, as shown in Figure 4(c).
|
| 249 |
+
|
| 250 |
+
Compared Agents. We compare the MCC agent with three different types of agents: the MC-Base agent (agent only executes its own meta-command without communication), the MC-Rand agent (agent randomly selects a meta-command to execute), and the MC-Rule agent (agent selects the nearest meta-command to execute). We adopt the MC-Base agent-only team as the opponent for all tests. Note that the MC-Base agent-only team has the ability of the SOTA and is more stable than the human-only team. Results are reported over five random seeds.
|
| 251 |
+
|
| 252 |
+
Agent-Only Environmental Settings. Test I is the most complex environment where all agent teammates can send and receive meta-commands simultaneously with an interval of 20s. Test I is used to evaluate the agents’ performance under extremely complex situations as well as in ablation studies. Test $\mathrm { I I }$ is a simple environment to simulate practical game scenarios, where at most one human sends his macro-strategy at a time step. Thus, in Test II, only one agent is randomly selected to send its meta-command with an interval of 20s, and the other agents only receive meta-commands.
|
| 253 |
+
|
| 254 |
+
Human-AI Game Testing Settings. We had different types of agents team up with different levels and numbers of humans, including 15 strong humans $( \tan 1 \% )$ and 15 average humans $( \mathrm { t o p } 3 0 \% )$ ), in m $A I + n$ Human mode, where $m + n = 5$ . For fair comparisons, each tester was not told the type of agent teammates. To eliminate the effects of collaboration between agents, we prohibit agents from receiving meta-commands from their agent teammates, and the agent can only receive meta-commands from humans. In each game test, humans can send the converted meta-commands whenever they think their macro-strategies are important. To make the agent behave like humans (at most one human sends his macro-strategy at a time step), we restrict agents from sending their meta-commands. We randomly choose a human teammate and use his observation and all agents’ meta-commands as the CS input and select the final output of CS to send with an interval of 20s.
|
| 255 |
+
|
| 256 |
+
# 5.2 Results in Agent-Only Environment
|
| 257 |
+
|
| 258 |
+
# 5.2.1 AI Performance
|
| 259 |
+
|
| 260 |
+
The Kullback-Leibler (KL) divergence of the meta-command distribution between the CEN and humans decreased from 4.96 to 0.44 as training converges. The MCCAN is trained with the parameter $\alpha$ equal to 16. The win rate of the trained agent against the SOTA agent [8, 38] is close to $50 \%$ . The average completion rates of the trained agent and humans for meta-commands are $82 \%$ and $80 \%$ , respectively. Notably, we can train an agent with a higher completion rate by increasing $\alpha$ , but this will significantly reduce the win rate because the meta-command executed is not necessarily optimal and may result in the death of agents. We put the detailed experimental results of the CEN and MCCAN in the Appendix A.10.1 and A.10.2 due to space limitations.
|
| 261 |
+
|
| 262 |
+
Figure 5(a) and (b) show the win rates of four types of agents who play against each other for 600 matches in Test I and Test II, respectively. We see that the MCC agent achieves the highest win rate against all the other agents in both testing environments, indicating that the CS can select a valuable meta-command for each agent to collaborate, and such reasonable collaboration is conducive to winning the game. The MC-Rand and MC-Rule agents are worse than the MC-Base agent, confirming that agents executing low-value meta-commands can hurt performance. Notably, we find that the win rates of the MCC agent in Test I and Test II are close, suggesting that the MCC agent can generalize to different numbers of meta-commands. Figure 5(c) demonstrates the final Elo scores [3] of these agents. It clearly shows the effectiveness of CS in agent-only collaboration scenarios.
|
| 263 |
+
|
| 264 |
+
# 5.2.2 Ablation Studies
|
| 265 |
+
|
| 266 |
+
We further investigate the influence of different components, including CNN feature extraction with the gating mechanism (w/o CNN-GM), target attention module (w/o TA), and PPO optimization algorithm (MCC-PPO), on the performance of CS. We conduct ablation studies in Test I with a 20 hero pool. In practical games, meta-commands
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
Figure 6: Results of ablation studies. (a) The training curves of different CS ablation versions. (b) The converged WR-RR results of different CS ablation versions. The shadow indicates the standard deviation.
|
| 270 |
+
|
| 271 |
+
with adjacent regions often have similar intentions and values. Thus the response rate of the agent to adjacent meta-commands should be as close as possible. Besides, the higher the agent’s response rate to meta-commands, the more collaborative behaviors of the agent, thus we expect the response rate of CS as high as possible. Generally, we expect the Response Rate (RR) of CS as high as possible while ensuring that the Win Rate (WR) is not reduced.
|
| 272 |
+
|
| 273 |
+
321 Figure 6(a) demonstrates the WR of different CS ablation versions during the training process, and
|
| 274 |
+
322 Figure 6(b) shows the converged WR-RR results. We see that after ablating the TA module, the WR
|
| 275 |
+
323 and RR of CS are greatly reduced, indicating that the TA module can improve the accuracy of CS
|
| 276 |
+
324 to meta-commands. Besides, after ablating the CNN-GM module, the RR of CS is most affected,
|
| 277 |
+
325 which is reduced by $20 \%$ . It indicates that without the CNN-GM module, the value estimation of CS
|
| 278 |
+
326 to adjacent meta-commands is not accurate enough, resulting in missing some actual high valuable
|
| 279 |
+
327 meta-commands. We notice that the MCC and MCC-PPO in both metrics are close, confirming the
|
| 280 |
+
328 versatility of the CS model structure.
|
| 281 |
+
|
| 282 |
+
Table 1: The WR of different human-AI teams against MC-Base agents in $4 A I + I$ Human mode.
|
| 283 |
+
|
| 284 |
+
<table><tr><td rowspan="2">Teammate</td><td colspan="3">Type of Agent</td></tr><tr><td>MC-Base</td><td>MC-Rand</td><td>MCC</td></tr><tr><td>Average Human</td><td>23%</td><td>5%</td><td>37%</td></tr><tr><td>Strong Human</td><td>42%</td><td>28%</td><td>54%</td></tr></table>
|
| 285 |
+
|
| 286 |
+
Table 2: The RR of humans and agents to teammates.
|
| 287 |
+
|
| 288 |
+
<table><tr><td>Sender\Receiver</td><td>Average Human</td><td>Strong Human</td><td>MCC</td></tr><tr><td>MC-Rand</td><td>41.07%</td><td>35.69%</td><td>34.03%</td></tr><tr><td>Average Human</td><td>72.34%</td><td>-</td><td>61.17%</td></tr><tr><td>Strong Human</td><td>-</td><td>74.91%</td><td>73.05%</td></tr><tr><td>MCC</td><td>73.43%</td><td>78.50%</td><td>-</td></tr></table>
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
Average Rank of Strong Human
|
| 292 |
+
|
| 293 |
+
|
| 294 |
+
Figure 7: Case study on the value estimation of CS.
|
| 295 |
+
|
| 296 |
+
# 29 5.3 Results in Human-AI Game Test
|
| 297 |
+
|
| 298 |
+
Due to space limitations, we only show the objective results in $4 A I + I$ Human mode. Other modes results and the subjective preference results of testers can be found in the Appendix A.10.3 and A.11. Table 1 shows the WR of different human-AI teams who play against the MC-Base agent-only team. We see that the MCC agent significantly outperforms other agents, regardless of whether they pair with a strong or average human. To explain why humans have a higher WR when paired with the MCC agents, we count the RR of agents to the meta-commands sent from human teammates (H2A scenarios) and the RR of humans to the meta-commands sent from agent teammates (A2H scenarios), respectively, as shown in Table 2. In H2A scenarios, the RRs of the MCC agents to average humans and strong humans are $6 1 . 1 7 \%$ and $7 3 . 0 5 \%$ , respectively, indicating that the MCC agents are more willing to respond to valuable meta-commands sent from strong humans. We also notice that the RR of the MCC agents to strong humans $( 7 3 . 0 5 \% )$ is very close to the RR of strong humans themselves $( 7 4 . 9 1 \% )$ , suggesting that the CS is close to the value system of strong humans. In A2H scenarios, the RRs of average humans and strong humans to the MCC agents are $7 3 . 4 3 \%$ and $78 . 5 \%$ , respectively, which is significantly higher than that of MC-Rand agents $( 4 1 . 0 7 \%$ and $3 5 . 6 9 \%$ ), indicating that the meta-commands sent from the MCC agents are more valuable and reasonable to humans. Note that the RR of the MCC agents to the MC-Rand agents is $3 4 . 0 3 \%$ , which is close to that of strong humans $( 3 5 . 6 9 \% )$ , once again confirming that the CS is close to the value system of strong humans.
|
| 299 |
+
|
| 300 |
+
We also visualize the comparison of CS and strong human value systems on a game scene with three meta-commands existing, as shown in Figure 7. We see that the CS selects the meta-command B for the two heroes in the red dashed box to collaborate, selects the meta-command C for the two heroes in the purple dashed box to collaborate, and selects the meta-command A for the remaining hero to execute alone. The CS selection results are consistent with the ranking results of strong humans, confirming the effectiveness of CS and the interpretability of the collaboration behavior between MCC agents and humans.
|
| 301 |
+
|
| 302 |
+
# 6 Conclusion
|
| 303 |
+
|
| 304 |
+
In this paper, we proposed an efficient and interpretable Meta-Command Communication-based framework, dubbed MCC, to achieve effective human-AI collaboration in MOBA games. To bridge the communication gap between humans and agents, we designed an interpretable communication protocol, i.e., the Meta-Command, to convert the explicit messages from humans and the implicit messages from agents into unified meta-commands. To achieve effective collaboration, we constructed a meta-command value estimation model, i.e., the Meta-Command Selector, to select a valuable meta-command for each agent to execute. Finally, we introduced the training process of the MCC framework and conducted practical human-AI game tests in the typical MOBA game Honor of Kings. The experimental results show that the MCC agents can collaborate reasonably with human teammates and even generalize to collaborate with different levels and numbers of human teammates. We expect this work can be a foundation for future HAC research in complex environments.
|
| 305 |
+
|
| 306 |
+
366 References
|
| 307 |
+
367 [1] Lupu Andrei, Cui Brandon, Hu Hengyuan, and Foerster Jakob N. Trajectory diversity for
|
| 308 |
+
368 zero-shot coordination. 139:7204–7213, 2021.
|
| 309 |
+
369 [2] Micah Carroll, Rohin Shah, Mark K Ho, Tom Griffiths, Sanjit Seshia, Pieter Abbeel, and Anca
|
| 310 |
+
370 Dragan. On the utility of learning about humans for human-ai coordination. Advances in Neural
|
| 311 |
+
371 Information Processing Systems, 32, 2019.
|
| 312 |
+
372 [3] Rémi Coulom. Whole-history rating: A bayesian rating system for players of time-varying
|
| 313 |
+
373 strength. In Proceedings of the International Conference on Computers and Games, pages
|
| 314 |
+
374 113–124, 2008.
|
| 315 |
+
375 [4] Allan Dafoe, Edward Hughes, Yoram Bachrach, Tantum Collins, Kevin R McKee, Joel Z
|
| 316 |
+
376 Leibo, Kate Larson, and Thore Graepel. Open problems in cooperative ai. arXiv preprint
|
| 317 |
+
377 arXiv:2012.08630, 2020.
|
| 318 |
+
378 [5] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and
|
| 319 |
+
379 Joelle Pineau. Tarmac: Targeted multi-agent communication. In Proceedings of the International
|
| 320 |
+
380 Conference on Machine Learning, pages 1538–1546, 2019.
|
| 321 |
+
381 [6] Yuqing Du, Stas Tiomkin, Emre Kiciman, Daniel Polani, Pieter Abbeel, and Anca Dragan. Ave:
|
| 322 |
+
382 Assistance via empowerment. Advances in Neural Information Processing Systems, 33:4560–4571,
|
| 323 |
+
383 2020.
|
| 324 |
+
384 [7] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning
|
| 325 |
+
385 to communicate with deep multi-agent reinforcement learning. Advances in Neural Information
|
| 326 |
+
386 Processing Systems, 29, 2016.
|
| 327 |
+
387 [8] Yiming Gao, Bei Shi, Xueying Du, Liang Wang, Guangwei Chen, Zhenjie Lian, Fuhao Qiu,
|
| 328 |
+
388 Guoan Han, Weixuan Wang, Deheng Ye, et al. Learning diverse policies in moba games via
|
| 329 |
+
389 macro-goals. Advances in Neural Information Processing Systems, 34, 2021.
|
| 330 |
+
390 [9] Mohammad Ghavamzadeh and Sridhar Mahadevan. Learning to communicate and act using
|
| 331 |
+
391 hierarchical reinforcement learning. Computer Science Department Faculty Publication Series,
|
| 332 |
+
392 page 172, 2004.
|
| 333 |
+
393 [10] Hengyuan Hu, Adam Lerer, Alex Peysakhovich, and Jakob Foerster. “other-play” for zero
|
| 334 |
+
394 shot coordination. In Proceedings of the International Conference on Machine Learning, pages
|
| 335 |
+
395 4399–4410, 2020.
|
| 336 |
+
396 [11] Max Jaderberg, Wojciech M Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia
|
| 337 |
+
397 Castaneda, Charles Beattie, Neil C Rabinowitz, Ari S Morcos, Avraham Ruderman, et al. Human
|
| 338 |
+
398 level performance in 3D multiplayer games with population-based reinforcement learning. Science,
|
| 339 |
+
399 364(6443):859–865, 2019.
|
| 340 |
+
400 [12] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent coopera
|
| 341 |
+
401 tion. Advances in Neural Information Processing Systems, 31, 2018.
|
| 342 |
+
402 [13] Uri Kartoun, Helman Stern, and Yael Edan. A human-robot collaborative reinforcement learning
|
| 343 |
+
403 algorithm. Journal of Intelligent & Robotic Systems, 60(2):217–239, 2010.
|
| 344 |
+
404 [14] Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. ´
|
| 345 |
+
405 Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE
|
| 346 |
+
406 Conference on Computational Intelligence and Games, pages 1–8, 2016.
|
| 347 |
+
407 [15] Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan
|
| 348 |
+
408 Son, and Yung Yi. Learning to schedule communication in multi-agent reinforcement learning.
|
| 349 |
+
409 arXiv preprint arXiv:1902.01554, 2019.
|
| 350 |
+
410 [16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
|
| 351 |
+
411 arXiv:1412.6980, 2014.
|
| 352 |
+
|
| 353 |
+
[17] Quan Liu, Zhihao Liu, Wenjun Xu, Quan Tang, Zude Zhou, and Duc Truong Pham. Humanrobot collaboration in disassembly for sustainable manufacturing. International Journal of Production Research, 57(12):4027–4044, 2019.
|
| 354 |
+
[18] Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in Neural Information Processing Systems, 34:9204–9215, 2021.
|
| 355 |
+
[19] Kevin R McKee, Xuechunzi Bai, and Susan T Fiske. Warmth and competence in human-agent cooperation. arXiv preprint arXiv:2201.13448, 2022.
|
| 356 |
+
[20] Kevin R McKee, Joel Z Leibo, Charlie Beattie, and Richard Everett. Quantifying the effects of environment and population diversity in multi-agent reinforcement learning. Autonomous Agents and Multi-Agent Systems, 36(1):1–16, 2022.
|
| 357 |
+
[21] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
|
| 358 |
+
[22] OpenAI, Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemysław D˛ebiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, Rafal Józefowicz, Scott Gray, Catherine Olsson, Jakub Pachocki, Michael Petrov, Henrique Pondé de Oliveira Pinto, Jonathan Raiman, Tim Salimans, Jeremy Schlatter, Jonas Schneider, Szymon Sidor, Ilya Sutskever, Jie Tang, Filip Wolski, and Susan Zhang. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680, 2019.
|
| 359 |
+
[23] Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang. Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017.
|
| 360 |
+
[24] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
|
| 361 |
+
[25] Ali Shafti, Jonas Tjomsland, William Dudley, and A Aldo Faisal. Real-world human-robot collaborative reinforcement learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 11161–11166, 2020.
|
| 362 |
+
[26] Victor do Nascimento Silva and Luiz Chaimowicz. Moba: A new arena for game ai. arXiv preprint arXiv:1705.10443, 2017.
|
| 363 |
+
[27] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.
|
| 364 |
+
[28] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of Go without human knowledge. Nature, 550(7676):354–359, 2017.
|
| 365 |
+
[29] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multi-agent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018.
|
| 366 |
+
[30] Ho Chit Siu, Jaime Peña, Edenna Chen, Yutai Zhou, Victor Lopez, Kyle Palko, Kimberlee Chang, and Ross Allen. Evaluation of human-ai teams for learned and rule-based agents in hanabi. Advances in Neural Information Processing Systems, 34:16183–16195, 2021.
|
| 367 |
+
[31] DJ Strouse, Kevin McKee, Matt Botvinick, Edward Hughes, and Richard Everett. Collaborating with humans without human data. Advances in Neural Information Processing Systems, 34, 2021.
|
| 368 |
+
[32] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multi-agent communication with backpropagation. Advances in Neural Information Processing Systems, 29, 2016.
|
| 369 |
+
[33] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016.
|
| 370 |
+
[34] Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michaël Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):350–354, 2019.
|
| 371 |
+
[35] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas. Dueling network architectures for deep reinforcement learning. In Proceedings of the International Conference on Machine Learning, pages 1995–2003, 2016.
|
| 372 |
+
[36] Rundong Wang, Xu He, Runsheng Yu, Wei Qiu, Bo An, and Zinovi Rabinovich. Learning efficient multi-agent communication: An information bottleneck approach. In Proceedings of the International Conference on Machine Learning, pages 9908–9918, 2020.
|
| 373 |
+
[37] Bin Wu. Hierarchical macro strategy model for moba game ai. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 1206–1213, 2019.
|
| 374 |
+
[38] Deheng Ye, Guibin Chen, Wen Zhang, Sheng Chen, Bo Yuan, Bo Liu, Jia Chen, Zhao Liu, Fuhao Qiu, Hongsheng Yu, et al. Towards playing full moba games with deep reinforcement learning. Advances in Neural Information Processing Systems, 33:621–632, 2020.
|
| 375 |
+
[39] Deheng Ye, Guibin Chen, Peilin Zhao, Fuhao Qiu, Bo Yuan, Wen Zhang, Sheng Chen, Mingfei Sun, Xiaoqian Li, Siqin Li, et al. Supervised learning achieves human-level performance in moba games: A case study of honor of kings. IEEE Transactions on Neural Networks and Learning Systems, 2020.
|
| 376 |
+
[40] Deheng Ye, Zhao Liu, Mingfei Sun, Bei Shi, Peilin Zhao, Hao Wu, Hongsheng Yu, Shaojie Yang, Xipeng Wu, Qingwei Guo, et al. Mastering complex control in moba games with deep reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6672–6679, 2020.
|
| 377 |
+
[41] Rui Zhao, Jinming Song, Haifeng Hu, Yang Gao, Yi Wu, Zhongqian Sun, and Wei Yang. Maximum entropy population based training for zero-shot human-ai coordination. arXiv preprint arXiv:2112.11701, 2021.
|
| 378 |
+
|
| 379 |
+
# Checklist
|
| 380 |
+
|
| 381 |
+
1. For all authors...
|
| 382 |
+
|
| 383 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] See the abstract and the contributions in Section 1.
|
| 384 |
+
(b) Did you describe the limitations of your work? [Yes] For margin reasons, we will discuss limitations here. We have currently only verified the effectiveness of the MCC framework in MOBA games, and we will explore in more types of complex games, such as First-Person Shooting (FPS) and Massively Multiplayer Online (MMO) in the future.
|
| 385 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] The purpose of our method is only for academic research based on game environments, e.g., investigating the MOBA game-playing problems. Like AlphaGo or AlphaStar, the potential negative societal impacts of our work will be limited to the development of gaming AI applications. However, our research will contribute to the research community, the game industry, and the e-sports community.
|
| 386 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 387 |
+
|
| 388 |
+
2. If you are including theoretical results...
|
| 389 |
+
|
| 390 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
|
| 391 |
+
|
| 392 |
+
3. If you ran experiments...
|
| 393 |
+
|
| 394 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] The code and the data are proprietary.
|
| 395 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See the full experimental setup in Section 5.1.1 and Appendix.
|
| 396 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] All results are reported over five random seeds (see Section 5.1.2). For the convenience of presentation, the median is shown in AI Performance, while detailed error bars are presented in Ablation Studies. Due to the high cost of Human-AI Game Test, we only used the median model for testing.
|
| 397 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See the resources in Section 5.1.1.
|
| 398 |
+
|
| 399 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 400 |
+
|
| 401 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 402 |
+
(b) Did you mention the license of the assets? [Yes]
|
| 403 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [No]
|
| 404 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [Yes] Honor of Kings is a released and recognized "testbed" for MOBA-game-playing problems [40, 39, 38, 8, 37]. We contact the relevant author and obtain authorization from the game provider.
|
| 405 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] See Appendix. All data in this paper is gamerelated and has nothing to do with identity information.
|
| 406 |
+
|
| 407 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 408 |
+
|
| 409 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes] See Appendix. In the Human-AI Game Test, all participants have more than three years of experience in Honor of Kings and are familiar with all the information in the game. Before the test, we also inform the participants of the detailed test instructions, and the participants voluntarily choose whether to participate in the test. We have detailed ethics descriptions in Appendix A.9. As stated in Section A.9.2, participants were given instructions before testing. As mentioned in point 5 (Line 87-88), participants’ game statistics will be only used for academic research, and participants can choose whether to participate or not.
|
| 410 |
+
|
| 411 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [Yes] We had performed a process similar to IRB before the test was conducted in this paper. The institution and all participants have approved our research. First, we analyze the risks of these experiments to the participants. The risks mainly include the leakage of identity information and the time cost. Then, a series of measures are implemented to prevent these risks in Appendix A.9.2. We make a risk statement for participants and sign an identity information confidentiality agreement. We only use information related to the game state in our research without identity information. In addition, special equipment and accounts are provided to the participants to prevent leakage of equipment and account information during the test. The identity information of all participants is not disclosed to the public.
|
| 412 |
+
|
| 413 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [Yes] In the Human-AI Game Test, participants can get 5 dollars for each match, and each match is about 10 to 20 minutes.
|
parse/dev/2nJdh_C-UWe/2nJdh_C-UWe_content_list.json
ADDED
|
@@ -0,0 +1,1125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Towards Effective and Interpretable Human-AI Collaboration in MOBA Games ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
210,
|
| 8 |
+
122,
|
| 9 |
+
785,
|
| 10 |
+
171
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Anonymous Author(s) \nAffiliation \nAddress \nemail ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
423,
|
| 19 |
+
222,
|
| 20 |
+
578,
|
| 21 |
+
276
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Abstract ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
462,
|
| 31 |
+
313,
|
| 32 |
+
535,
|
| 33 |
+
329
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "1 MOBA games, e.g., Dota2 and Honor of Kings, have been actively used as the \n2 testbed for the recent AI research on games, and various AI systems have been \n3 developed at the human level so far. However, these AI systems merely focus on \n4 how to compete with humans, less exploring how to collaborate with humans. To \n5 this end, this paper makes the first attempt to investigate human-AI collaboration in \n6 MOBA games. In this paper, we propose to enable humans and agents to collaborate \n7 through explicit communications by designing an efficient and interpretable Meta \n8 Command Communication-based framework, dubbed MCC, for accomplishing \n9 effective human-AI collaboration in MOBA games. The MCC framework consists \n10 of two pivotal modules: 1) an interpretable communication protocol, i.e., the \n11 Meta-Command, to bridge the communication gap between humans and agents; \n12 2) a meta-command value estimation model, i.e., the Meta-Command Selector, \n13 to select a valuable meta-command for each agent to achieve effective human-AI \n14 collaboration. Experimental results in Honor of Kings demonstrate that MCC \n15 agents can collaborate reasonably well with human teammates and even generalize \n16 to collaborate with different levels and numbers of human teammates. Videos are \n17 available at https://sites.google.com/view/mcc-demo. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
148,
|
| 42 |
+
344,
|
| 43 |
+
766,
|
| 44 |
+
579
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "18 1 Introduction ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
148,
|
| 54 |
+
599,
|
| 55 |
+
312,
|
| 56 |
+
616
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "19 Games, as the microcosm of real-world problems, have been widely used as testbeds to evaluate \n20 the performance of Artificial Intelligence (AI) techniques for decades. Recently, many researchers \n21 focus on developing various human-level AI systems for complex games, such as board games like \n22 Go [27, 28], First-Person Shooting (FPS) games like ViZDoom [14], Real-Time Strategy (RTS) \n23 games like StarCraft 2 [34], and Multi-player Online Battle Arena (MOBA) games like Dota 2 [22]. \n24 However, these AI systems focus merely on how to compete instead of collaborating with humans, \n25 leaving Human-AI Collaboration (HAC) in complex environments still to be investigated. \n26 In this paper, we study the HAC problem in complex MOBA games, which is characterized by multi \n27 agent cooperation and competition mechanisms, long time horizons, enormous state-action spaces \n28 $( \\bar { 1 0 } ^ { 2 0 0 0 0 } )$ , and imperfect information [22, 26, 38]. HAC requires the agent to collaborate reasonably \n29 with various human teammates. One straightforward approach is to improve the generalization of \n30 agents, that is, to collaborate with an enough diverse population of teammates during training. There \n31 are some Population-Based Training (PBT) based algorithms and learning systems [1, 2, 10, 11, \n32 31, 41] proposed to improve the generalization of agents in video games by constructing a diverse \n33 population of agents in different ways. However, this approach requires a vast amount of diverse data \n34 and massive computing resources, posing a big computational obstacle for complex MOBA games. \n35 Human team success in MOBA games requires not only subtle individual micro-operations but also \n36 excellent communications and collaborations among teammates on macro-strategies, i.e., long-term \n37 intentions [8, 37]. Consequently, we focus on enabling humans and agents to collaborate through \n38 explicit communications and propose an efficient and interpretable Meta-Command Communication \n39 based human-AI collaboration framework, dubbed MCC, to solve the HAC problem in MOBA \n40 games. First, we design an interpretable communication protocol, i.e., the Meta-Command, as a \n41 general representation of macro-strategies to bridge the communication gap between agents and \n42 humans. Both macro-strategies sent by humans and messages outputted by agents can be converted \n43 into unified meta-commands (see Figure 1). Second, following Gao et al. [8], we construct a \n44 hierarchical model that includes the command encoding network (macro-strategy layer) and the \n45 meta-command conditioned action network (micro-action layer), used for agents to generate and \n46 execute meta-commands, respectively. Third, we propose a meta-command value estimation model, \n47 i.e., the Meta-Command Selector, to select the optimal meta-command for each agent to execute. \n48 The training process of the MCC framework consists of three phases. We first train the command \n49 encoding network to learn the distribution of meta-commands sent from humans. Afterward, we \n50 train the meta-command conditioned action network to ensure that the agent has the near-human \n51 completion rate for meta-commands. Finally, we train the meta-command selector to ensure that the \n52 agent can select a valuable meta-command to achieve effective collaboration. We train and evaluate \n53 the agent in Honor of Kings 5v5 mode with a full hero pool (over 100 heroes). Experimental results \n54 demonstrate the effectiveness of the MCC framework. In general, our contributions are as follows: ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
147,
|
| 65 |
+
625,
|
| 66 |
+
825,
|
| 67 |
+
722
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "",
|
| 74 |
+
"bbox": [
|
| 75 |
+
147,
|
| 76 |
+
728,
|
| 77 |
+
825,
|
| 78 |
+
853
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "",
|
| 85 |
+
"bbox": [
|
| 86 |
+
148,
|
| 87 |
+
858,
|
| 88 |
+
823,
|
| 89 |
+
901
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/a47d3c45884a2e38bf746cefd3d4fd4bb0b9d7d11daa81a3d4c35ac30fb17b41.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: MOBA game-related introduction. (a) Key elements of MOBA games such as Dota 2, Honor of Kings, etc. Players observe from the state of the environment, make micro-operations and macro-strategies decisions, and collaborate through explicit messages (e.g.,text and signals). (b) Example of collaboration via meta-commands. The Come And Kill The Dragon is more valuable for humans A and B and agent $_ \\mathrm { D }$ t o collaborate, while the Clean Up Top-Lane Minions is more valuable for human C and agent E to collaborate. "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
178,
|
| 102 |
+
92,
|
| 103 |
+
816,
|
| 104 |
+
250
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "",
|
| 111 |
+
"bbox": [
|
| 112 |
+
145,
|
| 113 |
+
335,
|
| 114 |
+
825,
|
| 115 |
+
570
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "• To the best of our knowledge, we are the first to investigate the HAC problem in MOBA games. We propose an efficient and interpretable Meta-Command Communication-based framework dubbed MCC to achieve effective human-AI collaboration in MOBA games. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
163,
|
| 124 |
+
577,
|
| 125 |
+
821,
|
| 126 |
+
619
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "• We design an interpretable communication protocol to bridge the communication gap between humans and agents. In addition, we propose a meta-command value estimation model to select a valuable meta-command for each agent to achieve effective human-AI collaboration. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
158,
|
| 135 |
+
626,
|
| 136 |
+
821,
|
| 137 |
+
667
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "• We introduce the training process of the MCC framework in a typical MOBA game Honor of Kings and evaluate it in practical human-AI game tests. Experimental results show that MCC agents can reasonably collaborate with different levels and numbers of human teammates. ",
|
| 144 |
+
"bbox": [
|
| 145 |
+
171,
|
| 146 |
+
674,
|
| 147 |
+
825,
|
| 148 |
+
715
|
| 149 |
+
],
|
| 150 |
+
"page_idx": 1
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"type": "text",
|
| 154 |
+
"text": "64 2 Related Work ",
|
| 155 |
+
"text_level": 1,
|
| 156 |
+
"bbox": [
|
| 157 |
+
153,
|
| 158 |
+
729,
|
| 159 |
+
321,
|
| 160 |
+
746
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 1
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "text",
|
| 166 |
+
"text": "2.1 MOBA Games AI Research ",
|
| 167 |
+
"text_level": 1,
|
| 168 |
+
"bbox": [
|
| 169 |
+
163,
|
| 170 |
+
756,
|
| 171 |
+
405,
|
| 172 |
+
770
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 1
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "66 MOBA games, such as Dota 2 and Honor of Kings, have attracted much attention from AI researchers \n67 due to their multi-agent cooperative and competitive mechanics, long time horizons, partial observa \n68 tion, and enormous state-action spaces [22, 38]. Recently, OpenAI et al. [22] introduced an AI system \n69 named OpenAI-Five that defeated professional players in Dota 2 5v5 mode under the condition of \n70 limited heroes. Ye et al. [38, 39, 40] proposed another learning system named WuKong that can \n71 surpass top e-sport players in Honor of Kings with a full hero pool. Further, Wu [37] and Gao et \n72 al. [8] proposed learning systems that enable the agent to learn human strategies to achieve policy \n73 diversity. However, these AI systems can only defeat human players but cannot collaborate well due \n74 to the communication gap between agents and humans, see Table 1. In most real-world scenarios, the \n75 excellent collaboration between humans and agents may make more sense than the competition. ",
|
| 179 |
+
"bbox": [
|
| 180 |
+
148,
|
| 181 |
+
780,
|
| 182 |
+
825,
|
| 183 |
+
917
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 1
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "PBT is considered one way to solve the HAC problem [4]. Most PBT-based methods are devoted to training an agent which can be compatible with unseen partners by maintaining a population of agents with diverse behaviors in different ways [1, 2, 10, 11, 31, 41][6, 19, 20, 30]. These methods have been validated on both objective and subjective metrics in video games Overcooked and Capture the Flag and card game Hanabi. However, the main difference between these games and MOBA games is that these games do not provide explicit communication mechanics for collaboration on macrostrategies between agents and humans. Besides, MOBA AI agents usually need to learn billions of network parameters to cope with the enormous state-action spaces $( 1 0 ^ { 2 0 0 0 0 } )$ [38], which constitutes a prohibitive computational burden for learning. As a more realistic topic of HAC, human-robot interaction in manufacturing also attracts much attention [13, 17, 25]. However, these studies are mainly limited to collaboration between a robot and a human through one-way communication, i.e., humans give robots orders. Therefore, there is still a large room to study RL with the participation of humans. This work can be a stepping stone for broader real-world applications. ",
|
| 190 |
+
"bbox": [
|
| 191 |
+
169,
|
| 192 |
+
114,
|
| 193 |
+
825,
|
| 194 |
+
295
|
| 195 |
+
],
|
| 196 |
+
"page_idx": 2
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"type": "text",
|
| 200 |
+
"text": "2.3 Multi-Agent Communication ",
|
| 201 |
+
"text_level": 1,
|
| 202 |
+
"bbox": [
|
| 203 |
+
163,
|
| 204 |
+
308,
|
| 205 |
+
415,
|
| 206 |
+
321
|
| 207 |
+
],
|
| 208 |
+
"page_idx": 2
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"type": "text",
|
| 212 |
+
"text": "Communication is often used in Multi-Agent Reinforcement Learning (MARL) to improve interagent collaboration. Most communication-based MARL methods are mainly focused on exploring communication protocols between multiple agents with an end-to-end RL framework [5, 7, 9, 23, 29, 32, 36]. Jiang and Lu [12] and Kim et al. [15] proposed to model the value of multi-agent communication for effective collaboration. Unfortunately, these methods all model communications in a latent space without considering human-AI interactions, making it less interpretable to humans. Instead, we focus on enabling humans and agents to collaborate through explicit communications. ",
|
| 213 |
+
"bbox": [
|
| 214 |
+
171,
|
| 215 |
+
332,
|
| 216 |
+
825,
|
| 217 |
+
429
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "3 Human-AI Collaboration ",
|
| 224 |
+
"text_level": 1,
|
| 225 |
+
"bbox": [
|
| 226 |
+
165,
|
| 227 |
+
438,
|
| 228 |
+
416,
|
| 229 |
+
455
|
| 230 |
+
],
|
| 231 |
+
"page_idx": 2
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"type": "text",
|
| 235 |
+
"text": "We consider an interpretable communicative human-AI collaboration task, which can be extended from Partially Observable Markov Decision Process (POMDP) and formulated as a tuple $< N , H , { \\bf S } , { \\bf A } ^ { N } , { \\bf A } ^ { H } , { \\bf O } , { \\bf M } , r , P , \\gamma >$ , where $N$ and $H$ represent the numbers of agents and humans, respectively. S is the space of global states. $\\mathbf { A } ^ { N } = \\{ A _ { i } ^ { N } \\} _ { i = 1 , \\dots , N }$ and $\\mathbf { A } ^ { H } = \\{ A _ { i } ^ { \\breve { H } } \\} _ { i = 1 , \\dots , H }$ denote the spaces of actions of $N$ agents and $H$ humans, respectively. $\\mathbf { O } = \\{ O _ { i } \\} _ { i = 1 , \\dots , N + H }$ denotes the space of observations of $N$ agents and $H$ humans. $\\mathbf { M }$ represents the space of interpretable messages, that is, the Meta-Commands in the MCC framework. $P : \\mathbf { S } \\times \\bar { \\mathbf { A } } ^ { N } \\times \\mathbf { A } ^ { H } \\vec { \\mathbf { \\Lambda } } \\vec { \\mathbf { \\Lambda } } \\vec { \\mathbf { \\Lambda } } $ and $r : \\mathbf { S } \\times \\mathbf { A } ^ { N } \\times \\mathbf { A } ^ { H } \\to \\mathbb { R }$ denote the shared state transition probability function and reward function of $N$ agents, respectively. Note that, $r$ includes both individual reward and team reward. $\\gamma \\in [ 0 , 1 )$ denotes the discount factor. For each agent $i$ in state $s _ { t } \\in \\mathbf { S }$ , it receives an observation $o _ { t } ^ { i } \\in O _ { i }$ and a selected message $c _ { t } ^ { i } \\in \\mathbf { M }$ , and then outputs an action $a _ { t } ^ { i } = \\pi _ { \\theta } ( o _ { t } ^ { i } , c _ { t } ^ { i } ) \\in A _ { i } ^ { N }$ and a new message $m _ { t + 1 } ^ { i } = \\pi _ { \\phi } \\bar { ( } o _ { t } ^ { i } ) \\bar { \\ } \\in \\bf { M }$ , where $\\pi _ { \\theta }$ and $\\pi _ { \\phi }$ are action network and message encoding network, respectively. A message selector $c _ { t } ^ { i } = \\pi _ { \\omega } ( o _ { t } ^ { i } , C _ { t } )$ is introduced to receive a message set $C _ { t } = \\{ m _ { t } ^ { i } \\} _ { i = 1 , \\dots , N + H } \\subset \\mathbf { M }$ from all agents and humans and select the optimal one to execute. ",
|
| 236 |
+
"bbox": [
|
| 237 |
+
171,
|
| 238 |
+
465,
|
| 239 |
+
825,
|
| 240 |
+
661
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 2
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "text",
|
| 246 |
+
"text": "We divide the HAC problem in MOBA games into the Human-to-AI (H2A) and the AI-to-Human (A2H) scenarios. The H2A Scenario: Humans send macro-strategies as messages to agent teammates, and agents combine them with their own messages to select the optimal one based on their own message selector to execute, achieving effective collaboration with humans. The A2H Scenario: Agents send messages as macro-strategies to human teammates, and humans combine them with their own macro-strategies to select the optimal one based on their own value systems to execute, achieving effective collaboration with agents. The goal of both tasks is that agents and humans communicate macro-strategies with pre-defined communication protocols, and then select valuable macro-strategies for effective collaboration to win the game. ",
|
| 247 |
+
"bbox": [
|
| 248 |
+
160,
|
| 249 |
+
666,
|
| 250 |
+
826,
|
| 251 |
+
792
|
| 252 |
+
],
|
| 253 |
+
"page_idx": 2
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"type": "text",
|
| 257 |
+
"text": "4 Meta-Command Communication-Based Framework ",
|
| 258 |
+
"text_level": 1,
|
| 259 |
+
"bbox": [
|
| 260 |
+
161,
|
| 261 |
+
801,
|
| 262 |
+
642,
|
| 263 |
+
819
|
| 264 |
+
],
|
| 265 |
+
"page_idx": 2
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"type": "text",
|
| 269 |
+
"text": "123 In this section, we present the proposed MCC framework in detail. We first briefly describe three key \n124 stages of the MCC framework (see Section 4.1). Then we introduce the two pivotal modules in the \n125 MCC framework: 1) an interpretable communication protocol, i.e., the Meta-Command, as a general \n126 representation of macro-strategies to bridge the communication gap between agents and humans (see \n127 Section 4.2); 2) a meta-command value estimation model, i.e., the Meta-Command Selector, to select \n128 a valuable meta-command for each agent to achieve effective HAC in MOBA games(see Section 4.3). ",
|
| 270 |
+
"bbox": [
|
| 271 |
+
147,
|
| 272 |
+
828,
|
| 273 |
+
825,
|
| 274 |
+
911
|
| 275 |
+
],
|
| 276 |
+
"page_idx": 2
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"type": "image",
|
| 280 |
+
"img_path": "images/c89c4ffb7d8cc5140248e1fc8c06e64f011a5a4fb8897abe7e8a4618c1bf202a.jpg",
|
| 281 |
+
"image_caption": [
|
| 282 |
+
"Figure 2: The temporal process of the MCC framework. For each communication step $t$ and $T$ ), MCC first (I) converts messages from humans and agents into meta-commands, then (II) selects the optimal meta-command for each agent to execute, and (III) finally predicts a sequence of actions for each agent to perform. The selected meta-command is retained and executed for $_ n$ time steps. This process is repeated until the end of a game. "
|
| 283 |
+
],
|
| 284 |
+
"image_footnote": [],
|
| 285 |
+
"bbox": [
|
| 286 |
+
174,
|
| 287 |
+
87,
|
| 288 |
+
818,
|
| 289 |
+
327
|
| 290 |
+
],
|
| 291 |
+
"page_idx": 3
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"type": "text",
|
| 295 |
+
"text": "4.1 Overview ",
|
| 296 |
+
"text_level": 1,
|
| 297 |
+
"bbox": [
|
| 298 |
+
163,
|
| 299 |
+
396,
|
| 300 |
+
279,
|
| 301 |
+
410
|
| 302 |
+
],
|
| 303 |
+
"page_idx": 3
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"type": "text",
|
| 307 |
+
"text": "0 The flow of the MCC framework can be divided into three stages: the meta-command conversion stage, \nthe meta-command communication stage, and the human-AI collaboration stage, as plotted in Figure 2. \n2 At the Meta-Command Conversion Stage, the MCC framework converts the macro-strategies sent \n33 by humans and the messages outputted by the command encoding network of agents into unified \n34 meta-commands and then broadcasts them to all agents and humans. At the Meta-Command \n35 Communication Stage, the MCC framework uses the meta-command selector to estimate the values \n36 of all received meta-commands and select the optimal one for each agent to execute. Note that \n37 humans also select the optimal meta-command based on their value systems. At the Human-AI \n38 Collaboration Stage, the MCC framework adopts the meta-command conditioned action network to \n39 predict a sequence of actions for each agent to perform based on its selected meta-command. For \n40 each game, humans and agents have to collaborate multiple times, that is, they need to perform the \nabove three stages multiple times to win the game. ",
|
| 308 |
+
"bbox": [
|
| 309 |
+
155,
|
| 310 |
+
416,
|
| 311 |
+
826,
|
| 312 |
+
582
|
| 313 |
+
],
|
| 314 |
+
"page_idx": 3
|
| 315 |
+
},
|
| 316 |
+
{
|
| 317 |
+
"type": "text",
|
| 318 |
+
"text": "142 4.2 Meta-Command ",
|
| 319 |
+
"text_level": 1,
|
| 320 |
+
"bbox": [
|
| 321 |
+
151,
|
| 322 |
+
593,
|
| 323 |
+
328,
|
| 324 |
+
608
|
| 325 |
+
],
|
| 326 |
+
"page_idx": 3
|
| 327 |
+
},
|
| 328 |
+
{
|
| 329 |
+
"type": "text",
|
| 330 |
+
"text": "143 In MOBA games, we propose that a macro-strategy consists of three components: where to go, what \n144 to do, and how long. For example, a macro-strategy can be Come And Kill The Dragon, which \n145 consists of Come To The Dragon (where to go), Attack The Dragon (what to do), and Until The \n146 Dragon Is Killed (how long). Thus, we propose a general representation of macro-strategies, i.e., \n147 the Meta-Command, as an interpretable communication protocol to bridge the communication gap \n148 between agents and humans. \n149 Meta-Command Definition. We formulate the Meta-Command as a tuple $< L , E , T ^ { m c } >$ , as shown \n150 in Figure 1(b), where $L$ is the Location to go, $E$ is the Event to do after reaching $L$ , and $T ^ { m c }$ is the \n151 Time Limit for executing the meta-command. Among them, $L$ is the key to the meta-command, which \n152 contains the intention of the macro-strategy. $E$ can be thought of as human micro-operation, which is \n153 implemented through a pre-trained micro-action network $\\pi _ { \\theta }$ in the MCC framework. $T ^ { m c }$ can be set \n154 to how long it normally takes a human to complete a macro-strategy in MOBA games, usually 20 \n155 seconds corresponds to $80 \\%$ completion rate for meta-commands, see Appendix A.12.1. \n156 Meta-Command Conversion. To realize interpretable human-AI communication, we convert the \n157 explicit messages from humans and the implicit messages from agents into unified meta-commands. \n158 To achieve the former, a hand-crafted command converter function $f ^ { c c }$ is used to generate $L$ of meta \n159 commands by extracting the location from explicit messages, such as text and signals, sent by humans. \n160 To achieve the latter, we use a Command Encoding Network (CEN) $\\pi _ { \\phi } ( m | o )$ to generate $L$ of meta \n161 commands. The CEN is trained via supervised learning (SL) with the goal of learning the distribution \n162 of meta-commands sent from humans, as shown in Figure 3(a)(I). The training dataset $\\{ < o , m > \\}$ \n163 is obtained by extracting the observation $o$ and its corresponding meta-command $m$ from expert data. \n164 After converting all messages into unified meta-commands, the MCC framework broadcasts them to \n165 all agents and humans. Then, agents and humans receive an identical meta-command candidate set. \n166 Meta-Command Execution. After receiving a meta-command candidate set, agents can se \n167 lect one meta-command from it to execute. We adopt a Meta-Command Conditioned Ac \n168 tion Network (MCCAN) $\\pi _ { \\boldsymbol { \\theta } } ( a | o , m )$ for agents to perform actions based on the selected meta \n169 command, as shown in Figure 3(a)(II). The MCCAN is trained via goal-conditioned RL with \n170 the goal of achieving a near-human completion rate for the meta-commands generated by the \n171 pre-trained CEN while ensuring that the win rate is not reduced. We adopt an intrinsic reward \n172 $\\begin{array} { r } { \\dot { r } _ { t } ^ { i n t } ( s _ { t } , m _ { t } , s _ { t + 1 } ) = \\left| f ^ { c e } ( s _ { t } ) - \\check { m } _ { t } \\right| - \\left| f ^ { c e } ( s _ { t + 1 } ) - m _ { t } \\right| } \\end{array}$ to guide the process of executing the meta \n173 command $m _ { t }$ , where $f ^ { c e }$ is a hand-crafted command extraction function. We train the MCCAN \n174 with the objective of maximizing the expectation over extrinsic and intrinsic discounted total re \n175 wards $\\begin{array} { r } { G _ { t } = \\mathbb { E } _ { s \\sim d _ { \\pi _ { \\theta } } , a \\sim \\pi _ { \\theta } } \\left[ \\sum _ { i = 0 } ^ { \\infty } \\gamma ^ { i } r _ { t + i } + \\alpha \\sum _ { j = 0 } ^ { T ^ { m c } } \\gamma ^ { j } r _ { t + j } ^ { i n t } \\right] , } \\end{array}$ , where $\\alpha$ is a trade-off parameter and \n176 $\\begin{array} { r } { d _ { \\pi } ( s ) = \\operatorname* { l i m } _ { t \\infty } P ( s _ { t } = \\bar { s } \\mid s _ { 0 } , \\pi ) } \\end{array}$ is the probability when following $\\pi$ for $t$ steps from $s _ { 0 }$ . \n177 After training the CEN and MCCAN, we can achieve HAC by simply setting an agent to randomly \n178 select a meta-command derived from humans to execute. However, such collaboration is non \n179 intelligent and can even be a disaster for game victory because agents have no mechanism to \n180 model the values of meta-commands and cannot choose the optimal meta-command to execute. \n181 While humans usually choose the optimal one based on their value systems for achieving effective \n182 collaboration to win the game. Thus, we further propose a meta-command value estimation model to \n183 select a valuable meta-command for each agent, as described in the following subsection. ",
|
| 331 |
+
"bbox": [
|
| 332 |
+
142,
|
| 333 |
+
613,
|
| 334 |
+
825,
|
| 335 |
+
695
|
| 336 |
+
],
|
| 337 |
+
"page_idx": 3
|
| 338 |
+
},
|
| 339 |
+
{
|
| 340 |
+
"type": "text",
|
| 341 |
+
"text": "",
|
| 342 |
+
"bbox": [
|
| 343 |
+
140,
|
| 344 |
+
699,
|
| 345 |
+
825,
|
| 346 |
+
796
|
| 347 |
+
],
|
| 348 |
+
"page_idx": 3
|
| 349 |
+
},
|
| 350 |
+
{
|
| 351 |
+
"type": "text",
|
| 352 |
+
"text": "",
|
| 353 |
+
"bbox": [
|
| 354 |
+
140,
|
| 355 |
+
800,
|
| 356 |
+
826,
|
| 357 |
+
911
|
| 358 |
+
],
|
| 359 |
+
"page_idx": 3
|
| 360 |
+
},
|
| 361 |
+
{
|
| 362 |
+
"type": "image",
|
| 363 |
+
"img_path": "images/11ef9ab63482191eb45580c4a980f761f010eb2b27918be64211bd35b8a522df.jpg",
|
| 364 |
+
"image_caption": [
|
| 365 |
+
"Figure 3: The training process and model structure of MCC. (a) The training process is divided into three phases: we first (I) train the CEN via supervised learning (SL), then (II) train the MCCAN via goal-conditioned RL, and finally (III) train the CS via RL. Among them, the dashed box represents the frozen model. (b) The detailed CS model structure, including CNN feature extraction, gating mechanism, target attention module, etc. "
|
| 366 |
+
],
|
| 367 |
+
"image_footnote": [],
|
| 368 |
+
"bbox": [
|
| 369 |
+
174,
|
| 370 |
+
87,
|
| 371 |
+
823,
|
| 372 |
+
275
|
| 373 |
+
],
|
| 374 |
+
"page_idx": 4
|
| 375 |
+
},
|
| 376 |
+
{
|
| 377 |
+
"type": "text",
|
| 378 |
+
"text": "",
|
| 379 |
+
"bbox": [
|
| 380 |
+
148,
|
| 381 |
+
353,
|
| 382 |
+
825,
|
| 383 |
+
382
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 4
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "",
|
| 390 |
+
"bbox": [
|
| 391 |
+
142,
|
| 392 |
+
386,
|
| 393 |
+
826,
|
| 394 |
+
551
|
| 395 |
+
],
|
| 396 |
+
"page_idx": 4
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"type": "text",
|
| 400 |
+
"text": "",
|
| 401 |
+
"bbox": [
|
| 402 |
+
140,
|
| 403 |
+
555,
|
| 404 |
+
826,
|
| 405 |
+
652
|
| 406 |
+
],
|
| 407 |
+
"page_idx": 4
|
| 408 |
+
},
|
| 409 |
+
{
|
| 410 |
+
"type": "text",
|
| 411 |
+
"text": "184 4.3 Meta-Command Selector ",
|
| 412 |
+
"text_level": 1,
|
| 413 |
+
"bbox": [
|
| 414 |
+
148,
|
| 415 |
+
669,
|
| 416 |
+
387,
|
| 417 |
+
684
|
| 418 |
+
],
|
| 419 |
+
"page_idx": 4
|
| 420 |
+
},
|
| 421 |
+
{
|
| 422 |
+
"type": "text",
|
| 423 |
+
"text": "185 In real-world MOBA games, the same macro-strategy often has different values for different humans \n186 in different situations. For example, a macro-strategy can be Come And Kill The Dragon, as shown in \n187 Figure 1(b). It is more valuable for humans A and B to collaborate. While another macro-strategy can \n188 be Clean Up Top-Lane Minions, which is more valuable for human C rather than humans A and B. \n189 Therefore, it is important to select the most valuable meta-command from the received meta-command \n190 candidate set $C$ to achieve effective human-AI collaboration. We propose a meta-command value \n191 estimation model, i.e., the Meta-Command Selector (CS) $\\pi _ { \\omega } ( o , C )$ , to estimate the values of all \n192 current meta-commands and select the most valuable one for each agent to execute. ",
|
| 424 |
+
"bbox": [
|
| 425 |
+
140,
|
| 426 |
+
693,
|
| 427 |
+
825,
|
| 428 |
+
804
|
| 429 |
+
],
|
| 430 |
+
"page_idx": 4
|
| 431 |
+
},
|
| 432 |
+
{
|
| 433 |
+
"type": "text",
|
| 434 |
+
"text": "193 CS Optimization Objective. Typically, the execution of a meta-command involves reaching location \n194 $L$ and doing event $E$ , of which the latter is more important to the value of the meta-command. \n195 For example, for the meta-command Come And Kill The Dragon, if Kill The Dragon event cannot \n196 be done within $T ^ { m c }$ time steps, then it is pointless to Come To The Dragon. Thus, the long-term \nreward 197 steps b198 $R ^ { m c }$ for executinracting with a mete en manent: ds withi, where $T ^ { m c }$ $\\begin{array} { r } { R _ { t } ^ { m c } = \\sum _ { i = 0 } ^ { T ^ { L } } r _ { t + i } + \\beta \\sum _ { j = T ^ { L } } ^ { T ^ { m c } } r _ { t + j } } \\end{array}$ $T ^ { L } < T ^ { m c }$ is the time for reaching $L$ $\\beta > 1$ is a trade-off parameter. Note that the reward function $r$ ",
|
| 435 |
+
"bbox": [
|
| 436 |
+
142,
|
| 437 |
+
808,
|
| 438 |
+
825,
|
| 439 |
+
912
|
| 440 |
+
],
|
| 441 |
+
"page_idx": 4
|
| 442 |
+
},
|
| 443 |
+
{
|
| 444 |
+
"type": "text",
|
| 445 |
+
"text": "200 includes both individual rewards and team rewards. The optimization objective of CS is to select \n201 the optimal meta-command $m _ { t } ^ { * } = \\pi _ { \\omega } ( o _ { t } , C _ { t } )$ for each agent to maximize the expected discounted \n202 meta-command execution return Gmct = Es∼dπ ,m∼πω,a∼πθ $\\begin{array} { r } { G _ { t } ^ { m c } = \\mathbb { E } _ { s \\sim d _ { \\pi _ { \\theta } } , m \\sim \\pi _ { \\omega } , a \\sim \\pi _ { \\theta } } \\left[ \\sum _ { i = 0 } ^ { \\infty } \\gamma _ { m c } ^ { i } R _ { t + i \\cdot T ^ { m c } } ^ { m c } \\right] } \\end{array}$ , where $o _ { t } \\in \\mathbf { O }$ , \n203 $C _ { t }$ is the meta-command candidate set in state $s _ { t }$ , and $\\gamma _ { m c } \\in [ 0 , 1 )$ is the discount factor. \n204 CS Training Process. We construct a self-play training environment for CS where agents can send \n205 messages to each other. Specifically, three tricks in Figure 3(a)(III) are adopted to increase the \n206 sample efficiency while ensuring efficient exploration. First, each sent meta-command $m$ is sampled \n207 with the argmax rule from the results predicted by the pre-trained CEN. Second, each agent sends \n208 its meta-command with a probability $p$ every $T ^ { m c }$ time steps. Finally, each agent selects the final \n209 meta-command $c$ sampled with the softmax rule from its CS output results and hands it over to the \n210 pre-trained MCCAN for execution. We use the multi-head value mechanism [38] to model the value \n211 of the meta-command execution, and the corresponding value loss can be formulated as: ",
|
| 446 |
+
"bbox": [
|
| 447 |
+
140,
|
| 448 |
+
90,
|
| 449 |
+
826,
|
| 450 |
+
159
|
| 451 |
+
],
|
| 452 |
+
"page_idx": 5
|
| 453 |
+
},
|
| 454 |
+
{
|
| 455 |
+
"type": "text",
|
| 456 |
+
"text": "",
|
| 457 |
+
"bbox": [
|
| 458 |
+
140,
|
| 459 |
+
162,
|
| 460 |
+
825,
|
| 461 |
+
275
|
| 462 |
+
],
|
| 463 |
+
"page_idx": 5
|
| 464 |
+
},
|
| 465 |
+
{
|
| 466 |
+
"type": "equation",
|
| 467 |
+
"img_path": "images/b949aa8735932a13a7fa2aa197625afa12473564e72aaa9cb41ad9b59d579dc0.jpg",
|
| 468 |
+
"text": "$$\nL ^ { V } ( \\omega ) = \\mathbb { E } _ { S , C } \\left[ \\sum _ { h e a d _ { k } } \\| G _ { k } ^ { m c } - V _ { \\omega } ^ { k } ( S , C ) \\| _ { 2 } \\right] ,\n$$",
|
| 469 |
+
"text_format": "latex",
|
| 470 |
+
"bbox": [
|
| 471 |
+
354,
|
| 472 |
+
276,
|
| 473 |
+
642,
|
| 474 |
+
323
|
| 475 |
+
],
|
| 476 |
+
"page_idx": 5
|
| 477 |
+
},
|
| 478 |
+
{
|
| 479 |
+
"type": "text",
|
| 480 |
+
"text": "212 where 213 $V _ { \\omega } ^ { k } ( S , C )$ is the value of the $k$ -th head. For DQN-based methods [21, 33, 35], the $Q$ loss is: ",
|
| 481 |
+
"bbox": [
|
| 482 |
+
147,
|
| 483 |
+
324,
|
| 484 |
+
812,
|
| 485 |
+
340
|
| 486 |
+
],
|
| 487 |
+
"page_idx": 5
|
| 488 |
+
},
|
| 489 |
+
{
|
| 490 |
+
"type": "equation",
|
| 491 |
+
"img_path": "images/da415b428f532b23a44269b0f357ce8300025472fb61cb4e170a78f923ca4e61.jpg",
|
| 492 |
+
"text": "$$\nL ^ { Q } ( \\omega ) = \\mathbb { E } _ { S , C , M } \\left[ \\Vert G _ { t o t a l } - Q _ { \\omega } ^ { k } ( S , C , M ) \\Vert _ { 2 } \\right] , G _ { t o t a l } = \\sum _ { h e a d _ { k } } w _ { k } G _ { k } ^ { m c } ,\n$$",
|
| 493 |
+
"text_format": "latex",
|
| 494 |
+
"bbox": [
|
| 495 |
+
277,
|
| 496 |
+
342,
|
| 497 |
+
718,
|
| 498 |
+
375
|
| 499 |
+
],
|
| 500 |
+
"page_idx": 5
|
| 501 |
+
},
|
| 502 |
+
{
|
| 503 |
+
"type": "text",
|
| 504 |
+
"text": "where $w _ { k }$ is the weight of the $k$ -th head and $G _ { k } ^ { m c }$ is the Temporal Difference (TD) estimated value error $R _ { k } ^ { m c } + \\gamma _ { m c } V _ { \\omega } ^ { k } ( S ^ { \\prime } , C ^ { \\prime } ) - V _ { \\omega } ^ { k } ( S , C )$ . ",
|
| 505 |
+
"bbox": [
|
| 506 |
+
158,
|
| 507 |
+
376,
|
| 508 |
+
821,
|
| 509 |
+
407
|
| 510 |
+
],
|
| 511 |
+
"page_idx": 5
|
| 512 |
+
},
|
| 513 |
+
{
|
| 514 |
+
"type": "text",
|
| 515 |
+
"text": "CS Model Structure. We design a general network structure for CS towards MOBA games, as shown in Figure 3(b). In MOBA games, the meta-commands corresponding to adjacent regions usually have similar values. Thus, we divide the meta-commands in the map into grids, a common location description for MOBA games, and use the shared Convolutional Neural Network (CNN) to extract region-related information from the meta-commands to improve the generalization of CS to adjacent meta-commands. Besides, we use the gating mechanism [18] to fuse the map embedding of all received meta-commands and the state embedding of the observation information. Finally, to directly construct the relationship between the observation information and each meta-command, we introduce a target attention module, where the query is the fused embedding $h$ and the key is the map embedding $m ^ { \\prime }$ of each meta-command. The fused embedding $h$ is used as the input into the subsequent Q network $Q ( h , m ^ { \\prime } )$ and $\\mathrm { v }$ network $V ( h )$ network of CS. In this way, the Q network can also be easily converted to the policy network $\\pi ( \\boldsymbol { m } | \\boldsymbol { h } , \\boldsymbol { m } ^ { \\prime } )$ . Thus, the CS model structure can be easily applied to most popular RL algorithms, such as PPO [24], DQN [21], etc. ",
|
| 516 |
+
"bbox": [
|
| 517 |
+
173,
|
| 518 |
+
412,
|
| 519 |
+
826,
|
| 520 |
+
592
|
| 521 |
+
],
|
| 522 |
+
"page_idx": 5
|
| 523 |
+
},
|
| 524 |
+
{
|
| 525 |
+
"type": "text",
|
| 526 |
+
"text": "5 Experiments ",
|
| 527 |
+
"text_level": 1,
|
| 528 |
+
"bbox": [
|
| 529 |
+
171,
|
| 530 |
+
603,
|
| 531 |
+
313,
|
| 532 |
+
619
|
| 533 |
+
],
|
| 534 |
+
"page_idx": 5
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"type": "text",
|
| 538 |
+
"text": "We evaluate the proposed MCC framework in Honor of Kings, one of the most popular MOBA games worldwide, which has been actively used as the testbed for recent game AI research [8, 37–40]. We conduct all experiments in Honor of Kings 5v5 mode with a full hero pool (over 100 heroes), except ablation studies with a 20 hero pool for exploring the influence of different model components more sufficiently and efficiently. ",
|
| 539 |
+
"bbox": [
|
| 540 |
+
174,
|
| 541 |
+
627,
|
| 542 |
+
825,
|
| 543 |
+
698
|
| 544 |
+
],
|
| 545 |
+
"page_idx": 5
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"type": "text",
|
| 549 |
+
"text": "5.1 Experimental Setup ",
|
| 550 |
+
"text_level": 1,
|
| 551 |
+
"bbox": [
|
| 552 |
+
176,
|
| 553 |
+
708,
|
| 554 |
+
351,
|
| 555 |
+
723
|
| 556 |
+
],
|
| 557 |
+
"page_idx": 5
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"type": "text",
|
| 561 |
+
"text": "5.1.1 Training Setup 1 ",
|
| 562 |
+
"text_level": 1,
|
| 563 |
+
"bbox": [
|
| 564 |
+
176,
|
| 565 |
+
731,
|
| 566 |
+
338,
|
| 567 |
+
746
|
| 568 |
+
],
|
| 569 |
+
"page_idx": 5
|
| 570 |
+
},
|
| 571 |
+
{
|
| 572 |
+
"type": "text",
|
| 573 |
+
"text": "Due to the complexity of MOBA games and limited resources, instead of training jointly, we train the CEN, MCCAN, and CS sequentially. For all model training, the location $L$ of meta-commands in the map is divided into 144 grids. The time limit $T ^ { m c }$ for the meta-command execution is set to 20s. ",
|
| 574 |
+
"bbox": [
|
| 575 |
+
173,
|
| 576 |
+
752,
|
| 577 |
+
823,
|
| 578 |
+
795
|
| 579 |
+
],
|
| 580 |
+
"page_idx": 5
|
| 581 |
+
},
|
| 582 |
+
{
|
| 583 |
+
"type": "text",
|
| 584 |
+
"text": "CEN Training Settings. We train the CEN via SL until it converges for 26 hours using 8 NVIDIA P40 GPUs. The batch size of each GPU is set to 512. Adam[16] is adopted as the optimizer with an initial learning rate of 0.0001. ",
|
| 585 |
+
"bbox": [
|
| 586 |
+
174,
|
| 587 |
+
797,
|
| 588 |
+
825,
|
| 589 |
+
840
|
| 590 |
+
],
|
| 591 |
+
"page_idx": 5
|
| 592 |
+
},
|
| 593 |
+
{
|
| 594 |
+
"type": "text",
|
| 595 |
+
"text": "MCCAN Training Settings. We train the MCCAN by finetuning a pre-trained micro-action network [38], the state-of-the-art (SOTA) model in Honor of Kings, which is conditioned on the meta-command sampled from the pre-trained CEN. The MCCAN is trained until it converges for 48 ",
|
| 596 |
+
"bbox": [
|
| 597 |
+
176,
|
| 598 |
+
844,
|
| 599 |
+
825,
|
| 600 |
+
887
|
| 601 |
+
],
|
| 602 |
+
"page_idx": 5
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"type": "image",
|
| 606 |
+
"img_path": "images/03a0d2fabc9647984c0c8f2039d13c7fd0c145ace28ab51b5984238876d58c56.jpg",
|
| 607 |
+
"image_caption": [
|
| 608 |
+
"Figure 4: Communication environments in the experiment. The orange arrows indicate sending metacommands, and the blue arrows indicate receiving meta-commands. The dashed line denotes sending metacommands with probability $p$ . "
|
| 609 |
+
],
|
| 610 |
+
"image_footnote": [],
|
| 611 |
+
"bbox": [
|
| 612 |
+
173,
|
| 613 |
+
88,
|
| 614 |
+
825,
|
| 615 |
+
250
|
| 616 |
+
],
|
| 617 |
+
"page_idx": 6
|
| 618 |
+
},
|
| 619 |
+
{
|
| 620 |
+
"type": "image",
|
| 621 |
+
"img_path": "images/687ec50018143f5fcc38b1e11bc5431b5a3e117f130b9cc3e456aa4ac3fccbbe.jpg",
|
| 622 |
+
"image_caption": [
|
| 623 |
+
"Figure 5: AI performance in the testing environments. (a) and (b) show the win rate maps of different agents who play against each other. (c) shows the final Elo scores of these agents. "
|
| 624 |
+
],
|
| 625 |
+
"image_footnote": [],
|
| 626 |
+
"bbox": [
|
| 627 |
+
192,
|
| 628 |
+
305,
|
| 629 |
+
805,
|
| 630 |
+
429
|
| 631 |
+
],
|
| 632 |
+
"page_idx": 6
|
| 633 |
+
},
|
| 634 |
+
{
|
| 635 |
+
"type": "text",
|
| 636 |
+
"text": "247 hours using a physical computer cluster with 63,000 CPUs and 560 NVIDIA V100 GPUs. The batch \n248 size of each GPU is set to 256. ",
|
| 637 |
+
"bbox": [
|
| 638 |
+
150,
|
| 639 |
+
479,
|
| 640 |
+
823,
|
| 641 |
+
508
|
| 642 |
+
],
|
| 643 |
+
"page_idx": 6
|
| 644 |
+
},
|
| 645 |
+
{
|
| 646 |
+
"type": "text",
|
| 647 |
+
"text": "CS Training Settings. We train the CS via self-play until it converges for 24 hours using a physical computer cluster with 70,000 CPUs and 680 NVIDIA V100 GPUs. The batch size of each GPU is set to 256. The parameter $\\beta$ is set to 2. Each agent sends a meta-command with a probability $p$ of 0.8 and an interval $T ^ { m c }$ of 20s, as shown in Figure 4(a). ",
|
| 648 |
+
"bbox": [
|
| 649 |
+
174,
|
| 650 |
+
515,
|
| 651 |
+
825,
|
| 652 |
+
570
|
| 653 |
+
],
|
| 654 |
+
"page_idx": 6
|
| 655 |
+
},
|
| 656 |
+
{
|
| 657 |
+
"type": "text",
|
| 658 |
+
"text": "5.1.2 Evaluating Setup ",
|
| 659 |
+
"text_level": 1,
|
| 660 |
+
"bbox": [
|
| 661 |
+
174,
|
| 662 |
+
585,
|
| 663 |
+
344,
|
| 664 |
+
601
|
| 665 |
+
],
|
| 666 |
+
"page_idx": 6
|
| 667 |
+
},
|
| 668 |
+
{
|
| 669 |
+
"type": "text",
|
| 670 |
+
"text": "Our primary concern is whether the agents trained with the MCC framework, briefly called the MCC agents, can collaborate with humans well. However, evaluating agents with humans is expensive, which is not conducive to model selection and iteration. Therefore, we built two agent-only testing environments: Test I and Test II, for the model selection and iteration process, as shown in Figure 4(b). We also evaluate the MCC agents in practical human-AI game tests to examine the performance of collaborating with humans, as shown in Figure 4(c). ",
|
| 671 |
+
"bbox": [
|
| 672 |
+
174,
|
| 673 |
+
608,
|
| 674 |
+
825,
|
| 675 |
+
691
|
| 676 |
+
],
|
| 677 |
+
"page_idx": 6
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"type": "text",
|
| 681 |
+
"text": "Compared Agents. We compare the MCC agent with three different types of agents: the MC-Base agent (agent only executes its own meta-command without communication), the MC-Rand agent (agent randomly selects a meta-command to execute), and the MC-Rule agent (agent selects the nearest meta-command to execute). We adopt the MC-Base agent-only team as the opponent for all tests. Note that the MC-Base agent-only team has the ability of the SOTA and is more stable than the human-only team. Results are reported over five random seeds. ",
|
| 682 |
+
"bbox": [
|
| 683 |
+
174,
|
| 684 |
+
695,
|
| 685 |
+
825,
|
| 686 |
+
779
|
| 687 |
+
],
|
| 688 |
+
"page_idx": 6
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"type": "text",
|
| 692 |
+
"text": "Agent-Only Environmental Settings. Test I is the most complex environment where all agent teammates can send and receive meta-commands simultaneously with an interval of 20s. Test I is used to evaluate the agents’ performance under extremely complex situations as well as in ablation studies. Test $\\mathrm { I I }$ is a simple environment to simulate practical game scenarios, where at most one human sends his macro-strategy at a time step. Thus, in Test II, only one agent is randomly selected to send its meta-command with an interval of 20s, and the other agents only receive meta-commands. ",
|
| 693 |
+
"bbox": [
|
| 694 |
+
174,
|
| 695 |
+
782,
|
| 696 |
+
825,
|
| 697 |
+
866
|
| 698 |
+
],
|
| 699 |
+
"page_idx": 6
|
| 700 |
+
},
|
| 701 |
+
{
|
| 702 |
+
"type": "text",
|
| 703 |
+
"text": "Human-AI Game Testing Settings. We had different types of agents team up with different levels and numbers of humans, including 15 strong humans $( \\tan 1 \\% )$ and 15 average humans $( \\mathrm { t o p } 3 0 \\% )$ ), in m $A I + n$ Human mode, where $m + n = 5$ . For fair comparisons, each tester was not told the type of agent teammates. To eliminate the effects of collaboration between agents, we prohibit agents from receiving meta-commands from their agent teammates, and the agent can only receive meta-commands from humans. In each game test, humans can send the converted meta-commands whenever they think their macro-strategies are important. To make the agent behave like humans (at most one human sends his macro-strategy at a time step), we restrict agents from sending their meta-commands. We randomly choose a human teammate and use his observation and all agents’ meta-commands as the CS input and select the final output of CS to send with an interval of 20s. ",
|
| 704 |
+
"bbox": [
|
| 705 |
+
169,
|
| 706 |
+
869,
|
| 707 |
+
825,
|
| 708 |
+
911
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 6
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "text",
|
| 714 |
+
"text": "",
|
| 715 |
+
"bbox": [
|
| 716 |
+
161,
|
| 717 |
+
92,
|
| 718 |
+
825,
|
| 719 |
+
189
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 7
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "text",
|
| 725 |
+
"text": "5.2 Results in Agent-Only Environment ",
|
| 726 |
+
"text_level": 1,
|
| 727 |
+
"bbox": [
|
| 728 |
+
165,
|
| 729 |
+
205,
|
| 730 |
+
460,
|
| 731 |
+
219
|
| 732 |
+
],
|
| 733 |
+
"page_idx": 7
|
| 734 |
+
},
|
| 735 |
+
{
|
| 736 |
+
"type": "text",
|
| 737 |
+
"text": "5.2.1 AI Performance ",
|
| 738 |
+
"text_level": 1,
|
| 739 |
+
"bbox": [
|
| 740 |
+
176,
|
| 741 |
+
229,
|
| 742 |
+
338,
|
| 743 |
+
243
|
| 744 |
+
],
|
| 745 |
+
"page_idx": 7
|
| 746 |
+
},
|
| 747 |
+
{
|
| 748 |
+
"type": "text",
|
| 749 |
+
"text": "The Kullback-Leibler (KL) divergence of the meta-command distribution between the CEN and humans decreased from 4.96 to 0.44 as training converges. The MCCAN is trained with the parameter $\\alpha$ equal to 16. The win rate of the trained agent against the SOTA agent [8, 38] is close to $50 \\%$ . The average completion rates of the trained agent and humans for meta-commands are $82 \\%$ and $80 \\%$ , respectively. Notably, we can train an agent with a higher completion rate by increasing $\\alpha$ , but this will significantly reduce the win rate because the meta-command executed is not necessarily optimal and may result in the death of agents. We put the detailed experimental results of the CEN and MCCAN in the Appendix A.10.1 and A.10.2 due to space limitations. ",
|
| 750 |
+
"bbox": [
|
| 751 |
+
173,
|
| 752 |
+
252,
|
| 753 |
+
825,
|
| 754 |
+
363
|
| 755 |
+
],
|
| 756 |
+
"page_idx": 7
|
| 757 |
+
},
|
| 758 |
+
{
|
| 759 |
+
"type": "text",
|
| 760 |
+
"text": "Figure 5(a) and (b) show the win rates of four types of agents who play against each other for 600 matches in Test I and Test II, respectively. We see that the MCC agent achieves the highest win rate against all the other agents in both testing environments, indicating that the CS can select a valuable meta-command for each agent to collaborate, and such reasonable collaboration is conducive to winning the game. The MC-Rand and MC-Rule agents are worse than the MC-Base agent, confirming that agents executing low-value meta-commands can hurt performance. Notably, we find that the win rates of the MCC agent in Test I and Test II are close, suggesting that the MCC agent can generalize to different numbers of meta-commands. Figure 5(c) demonstrates the final Elo scores [3] of these agents. It clearly shows the effectiveness of CS in agent-only collaboration scenarios. ",
|
| 761 |
+
"bbox": [
|
| 762 |
+
173,
|
| 763 |
+
369,
|
| 764 |
+
825,
|
| 765 |
+
494
|
| 766 |
+
],
|
| 767 |
+
"page_idx": 7
|
| 768 |
+
},
|
| 769 |
+
{
|
| 770 |
+
"type": "text",
|
| 771 |
+
"text": "5.2.2 Ablation Studies ",
|
| 772 |
+
"text_level": 1,
|
| 773 |
+
"bbox": [
|
| 774 |
+
176,
|
| 775 |
+
508,
|
| 776 |
+
339,
|
| 777 |
+
522
|
| 778 |
+
],
|
| 779 |
+
"page_idx": 7
|
| 780 |
+
},
|
| 781 |
+
{
|
| 782 |
+
"type": "text",
|
| 783 |
+
"text": "We further investigate the influence of different components, including CNN feature extraction with the gating mechanism (w/o CNN-GM), target attention module (w/o TA), and PPO optimization algorithm (MCC-PPO), on the performance of CS. We conduct ablation studies in Test I with a 20 hero pool. In practical games, meta-commands ",
|
| 784 |
+
"bbox": [
|
| 785 |
+
173,
|
| 786 |
+
532,
|
| 787 |
+
356,
|
| 788 |
+
724
|
| 789 |
+
],
|
| 790 |
+
"page_idx": 7
|
| 791 |
+
},
|
| 792 |
+
{
|
| 793 |
+
"type": "image",
|
| 794 |
+
"img_path": "images/a93f018c6e2b88df6ca939c1302af4814d6cd1836f9bf57043fac763bee3f8d5.jpg",
|
| 795 |
+
"image_caption": [
|
| 796 |
+
"Figure 6: Results of ablation studies. (a) The training curves of different CS ablation versions. (b) The converged WR-RR results of different CS ablation versions. The shadow indicates the standard deviation. "
|
| 797 |
+
],
|
| 798 |
+
"image_footnote": [],
|
| 799 |
+
"bbox": [
|
| 800 |
+
370,
|
| 801 |
+
525,
|
| 802 |
+
820,
|
| 803 |
+
665
|
| 804 |
+
],
|
| 805 |
+
"page_idx": 7
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "text",
|
| 809 |
+
"text": "with adjacent regions often have similar intentions and values. Thus the response rate of the agent to adjacent meta-commands should be as close as possible. Besides, the higher the agent’s response rate to meta-commands, the more collaborative behaviors of the agent, thus we expect the response rate of CS as high as possible. Generally, we expect the Response Rate (RR) of CS as high as possible while ensuring that the Win Rate (WR) is not reduced. ",
|
| 810 |
+
"bbox": [
|
| 811 |
+
174,
|
| 812 |
+
726,
|
| 813 |
+
823,
|
| 814 |
+
794
|
| 815 |
+
],
|
| 816 |
+
"page_idx": 7
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "text",
|
| 820 |
+
"text": "321 Figure 6(a) demonstrates the WR of different CS ablation versions during the training process, and \n322 Figure 6(b) shows the converged WR-RR results. We see that after ablating the TA module, the WR \n323 and RR of CS are greatly reduced, indicating that the TA module can improve the accuracy of CS \n324 to meta-commands. Besides, after ablating the CNN-GM module, the RR of CS is most affected, \n325 which is reduced by $20 \\%$ . It indicates that without the CNN-GM module, the value estimation of CS \n326 to adjacent meta-commands is not accurate enough, resulting in missing some actual high valuable \n327 meta-commands. We notice that the MCC and MCC-PPO in both metrics are close, confirming the \n328 versatility of the CS model structure. ",
|
| 821 |
+
"bbox": [
|
| 822 |
+
142,
|
| 823 |
+
800,
|
| 824 |
+
825,
|
| 825 |
+
911
|
| 826 |
+
],
|
| 827 |
+
"page_idx": 7
|
| 828 |
+
},
|
| 829 |
+
{
|
| 830 |
+
"type": "table",
|
| 831 |
+
"img_path": "images/334729e551ad5dfde0d22aaaf2d28a85cce7e095852f0f5fd51347c36f470e64.jpg",
|
| 832 |
+
"table_caption": [
|
| 833 |
+
"Table 1: The WR of different human-AI teams against MC-Base agents in $4 A I + I$ Human mode. "
|
| 834 |
+
],
|
| 835 |
+
"table_footnote": [],
|
| 836 |
+
"table_body": "<table><tr><td rowspan=\"2\">Teammate</td><td colspan=\"3\">Type of Agent</td></tr><tr><td>MC-Base</td><td>MC-Rand</td><td>MCC</td></tr><tr><td>Average Human</td><td>23%</td><td>5%</td><td>37%</td></tr><tr><td>Strong Human</td><td>42%</td><td>28%</td><td>54%</td></tr></table>",
|
| 837 |
+
"bbox": [
|
| 838 |
+
187,
|
| 839 |
+
125,
|
| 840 |
+
465,
|
| 841 |
+
178
|
| 842 |
+
],
|
| 843 |
+
"page_idx": 8
|
| 844 |
+
},
|
| 845 |
+
{
|
| 846 |
+
"type": "table",
|
| 847 |
+
"img_path": "images/b833b99a3592874da074593da9d9dc7b291df0798ebf6032e2a04a6ad2fa526a.jpg",
|
| 848 |
+
"table_caption": [
|
| 849 |
+
"Table 2: The RR of humans and agents to teammates. "
|
| 850 |
+
],
|
| 851 |
+
"table_footnote": [],
|
| 852 |
+
"table_body": "<table><tr><td>Sender\\Receiver</td><td>Average Human</td><td>Strong Human</td><td>MCC</td></tr><tr><td>MC-Rand</td><td>41.07%</td><td>35.69%</td><td>34.03%</td></tr><tr><td>Average Human</td><td>72.34%</td><td>-</td><td>61.17%</td></tr><tr><td>Strong Human</td><td>-</td><td>74.91%</td><td>73.05%</td></tr><tr><td>MCC</td><td>73.43%</td><td>78.50%</td><td>-</td></tr></table>",
|
| 853 |
+
"bbox": [
|
| 854 |
+
478,
|
| 855 |
+
111,
|
| 856 |
+
826,
|
| 857 |
+
178
|
| 858 |
+
],
|
| 859 |
+
"page_idx": 8
|
| 860 |
+
},
|
| 861 |
+
{
|
| 862 |
+
"type": "image",
|
| 863 |
+
"img_path": "images/11037341ffae27d4daf7386e092e7780793b1cee69cdcaba5a981953a5997e08.jpg",
|
| 864 |
+
"image_caption": [
|
| 865 |
+
"Average Rank of Strong Human "
|
| 866 |
+
],
|
| 867 |
+
"image_footnote": [],
|
| 868 |
+
"bbox": [
|
| 869 |
+
410,
|
| 870 |
+
202,
|
| 871 |
+
802,
|
| 872 |
+
339
|
| 873 |
+
],
|
| 874 |
+
"page_idx": 8
|
| 875 |
+
},
|
| 876 |
+
{
|
| 877 |
+
"type": "image",
|
| 878 |
+
"img_path": "images/ba5c6725660d0430fbb713315f508d10f2caf23f2a77b611c01cf7ae7b4b0d9a.jpg",
|
| 879 |
+
"image_caption": [
|
| 880 |
+
"Figure 7: Case study on the value estimation of CS. "
|
| 881 |
+
],
|
| 882 |
+
"image_footnote": [],
|
| 883 |
+
"bbox": [
|
| 884 |
+
199,
|
| 885 |
+
185,
|
| 886 |
+
398,
|
| 887 |
+
339
|
| 888 |
+
],
|
| 889 |
+
"page_idx": 8
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"type": "text",
|
| 893 |
+
"text": "29 5.3 Results in Human-AI Game Test ",
|
| 894 |
+
"text_level": 1,
|
| 895 |
+
"bbox": [
|
| 896 |
+
156,
|
| 897 |
+
368,
|
| 898 |
+
441,
|
| 899 |
+
383
|
| 900 |
+
],
|
| 901 |
+
"page_idx": 8
|
| 902 |
+
},
|
| 903 |
+
{
|
| 904 |
+
"type": "text",
|
| 905 |
+
"text": "Due to space limitations, we only show the objective results in $4 A I + I$ Human mode. Other modes results and the subjective preference results of testers can be found in the Appendix A.10.3 and A.11. Table 1 shows the WR of different human-AI teams who play against the MC-Base agent-only team. We see that the MCC agent significantly outperforms other agents, regardless of whether they pair with a strong or average human. To explain why humans have a higher WR when paired with the MCC agents, we count the RR of agents to the meta-commands sent from human teammates (H2A scenarios) and the RR of humans to the meta-commands sent from agent teammates (A2H scenarios), respectively, as shown in Table 2. In H2A scenarios, the RRs of the MCC agents to average humans and strong humans are $6 1 . 1 7 \\%$ and $7 3 . 0 5 \\%$ , respectively, indicating that the MCC agents are more willing to respond to valuable meta-commands sent from strong humans. We also notice that the RR of the MCC agents to strong humans $( 7 3 . 0 5 \\% )$ is very close to the RR of strong humans themselves $( 7 4 . 9 1 \\% )$ , suggesting that the CS is close to the value system of strong humans. In A2H scenarios, the RRs of average humans and strong humans to the MCC agents are $7 3 . 4 3 \\%$ and $78 . 5 \\%$ , respectively, which is significantly higher than that of MC-Rand agents $( 4 1 . 0 7 \\%$ and $3 5 . 6 9 \\%$ ), indicating that the meta-commands sent from the MCC agents are more valuable and reasonable to humans. Note that the RR of the MCC agents to the MC-Rand agents is $3 4 . 0 3 \\%$ , which is close to that of strong humans $( 3 5 . 6 9 \\% )$ , once again confirming that the CS is close to the value system of strong humans. ",
|
| 906 |
+
"bbox": [
|
| 907 |
+
173,
|
| 908 |
+
391,
|
| 909 |
+
825,
|
| 910 |
+
626
|
| 911 |
+
],
|
| 912 |
+
"page_idx": 8
|
| 913 |
+
},
|
| 914 |
+
{
|
| 915 |
+
"type": "text",
|
| 916 |
+
"text": "We also visualize the comparison of CS and strong human value systems on a game scene with three meta-commands existing, as shown in Figure 7. We see that the CS selects the meta-command B for the two heroes in the red dashed box to collaborate, selects the meta-command C for the two heroes in the purple dashed box to collaborate, and selects the meta-command A for the remaining hero to execute alone. The CS selection results are consistent with the ranking results of strong humans, confirming the effectiveness of CS and the interpretability of the collaboration behavior between MCC agents and humans. ",
|
| 917 |
+
"bbox": [
|
| 918 |
+
169,
|
| 919 |
+
633,
|
| 920 |
+
825,
|
| 921 |
+
729
|
| 922 |
+
],
|
| 923 |
+
"page_idx": 8
|
| 924 |
+
},
|
| 925 |
+
{
|
| 926 |
+
"type": "text",
|
| 927 |
+
"text": "6 Conclusion ",
|
| 928 |
+
"text_level": 1,
|
| 929 |
+
"bbox": [
|
| 930 |
+
173,
|
| 931 |
+
738,
|
| 932 |
+
299,
|
| 933 |
+
753
|
| 934 |
+
],
|
| 935 |
+
"page_idx": 8
|
| 936 |
+
},
|
| 937 |
+
{
|
| 938 |
+
"type": "text",
|
| 939 |
+
"text": "In this paper, we proposed an efficient and interpretable Meta-Command Communication-based framework, dubbed MCC, to achieve effective human-AI collaboration in MOBA games. To bridge the communication gap between humans and agents, we designed an interpretable communication protocol, i.e., the Meta-Command, to convert the explicit messages from humans and the implicit messages from agents into unified meta-commands. To achieve effective collaboration, we constructed a meta-command value estimation model, i.e., the Meta-Command Selector, to select a valuable meta-command for each agent to execute. Finally, we introduced the training process of the MCC framework and conducted practical human-AI game tests in the typical MOBA game Honor of Kings. The experimental results show that the MCC agents can collaborate reasonably with human teammates and even generalize to collaborate with different levels and numbers of human teammates. We expect this work can be a foundation for future HAC research in complex environments. ",
|
| 940 |
+
"bbox": [
|
| 941 |
+
173,
|
| 942 |
+
760,
|
| 943 |
+
825,
|
| 944 |
+
911
|
| 945 |
+
],
|
| 946 |
+
"page_idx": 8
|
| 947 |
+
},
|
| 948 |
+
{
|
| 949 |
+
"type": "text",
|
| 950 |
+
"text": "366 References \n367 [1] Lupu Andrei, Cui Brandon, Hu Hengyuan, and Foerster Jakob N. Trajectory diversity for \n368 zero-shot coordination. 139:7204–7213, 2021. \n369 [2] Micah Carroll, Rohin Shah, Mark K Ho, Tom Griffiths, Sanjit Seshia, Pieter Abbeel, and Anca \n370 Dragan. On the utility of learning about humans for human-ai coordination. Advances in Neural \n371 Information Processing Systems, 32, 2019. \n372 [3] Rémi Coulom. Whole-history rating: A bayesian rating system for players of time-varying \n373 strength. In Proceedings of the International Conference on Computers and Games, pages \n374 113–124, 2008. \n375 [4] Allan Dafoe, Edward Hughes, Yoram Bachrach, Tantum Collins, Kevin R McKee, Joel Z \n376 Leibo, Kate Larson, and Thore Graepel. Open problems in cooperative ai. arXiv preprint \n377 arXiv:2012.08630, 2020. \n378 [5] Abhishek Das, Théophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Mike Rabbat, and \n379 Joelle Pineau. Tarmac: Targeted multi-agent communication. In Proceedings of the International \n380 Conference on Machine Learning, pages 1538–1546, 2019. \n381 [6] Yuqing Du, Stas Tiomkin, Emre Kiciman, Daniel Polani, Pieter Abbeel, and Anca Dragan. Ave: \n382 Assistance via empowerment. Advances in Neural Information Processing Systems, 33:4560–4571, \n383 2020. \n384 [7] Jakob Foerster, Ioannis Alexandros Assael, Nando De Freitas, and Shimon Whiteson. Learning \n385 to communicate with deep multi-agent reinforcement learning. Advances in Neural Information \n386 Processing Systems, 29, 2016. \n387 [8] Yiming Gao, Bei Shi, Xueying Du, Liang Wang, Guangwei Chen, Zhenjie Lian, Fuhao Qiu, \n388 Guoan Han, Weixuan Wang, Deheng Ye, et al. Learning diverse policies in moba games via \n389 macro-goals. Advances in Neural Information Processing Systems, 34, 2021. \n390 [9] Mohammad Ghavamzadeh and Sridhar Mahadevan. Learning to communicate and act using \n391 hierarchical reinforcement learning. Computer Science Department Faculty Publication Series, \n392 page 172, 2004. \n393 [10] Hengyuan Hu, Adam Lerer, Alex Peysakhovich, and Jakob Foerster. “other-play” for zero \n394 shot coordination. In Proceedings of the International Conference on Machine Learning, pages \n395 4399–4410, 2020. \n396 [11] Max Jaderberg, Wojciech M Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia \n397 Castaneda, Charles Beattie, Neil C Rabinowitz, Ari S Morcos, Avraham Ruderman, et al. Human \n398 level performance in 3D multiplayer games with population-based reinforcement learning. Science, \n399 364(6443):859–865, 2019. \n400 [12] Jiechuan Jiang and Zongqing Lu. Learning attentional communication for multi-agent coopera \n401 tion. Advances in Neural Information Processing Systems, 31, 2018. \n402 [13] Uri Kartoun, Helman Stern, and Yael Edan. A human-robot collaborative reinforcement learning \n403 algorithm. Journal of Intelligent & Robotic Systems, 60(2):217–239, 2010. \n404 [14] Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. ´ \n405 Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE \n406 Conference on Computational Intelligence and Games, pages 1–8, 2016. \n407 [15] Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan \n408 Son, and Yung Yi. Learning to schedule communication in multi-agent reinforcement learning. \n409 arXiv preprint arXiv:1902.01554, 2019. \n410 [16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint \n411 arXiv:1412.6980, 2014. ",
|
| 951 |
+
"bbox": [
|
| 952 |
+
138,
|
| 953 |
+
66,
|
| 954 |
+
828,
|
| 955 |
+
931
|
| 956 |
+
],
|
| 957 |
+
"page_idx": 9
|
| 958 |
+
},
|
| 959 |
+
{
|
| 960 |
+
"type": "text",
|
| 961 |
+
"text": "[17] Quan Liu, Zhihao Liu, Wenjun Xu, Quan Tang, Zude Zhou, and Duc Truong Pham. Humanrobot collaboration in disassembly for sustainable manufacturing. International Journal of Production Research, 57(12):4027–4044, 2019. \n[18] Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. Pay attention to mlps. Advances in Neural Information Processing Systems, 34:9204–9215, 2021. \n[19] Kevin R McKee, Xuechunzi Bai, and Susan T Fiske. Warmth and competence in human-agent cooperation. arXiv preprint arXiv:2201.13448, 2022. \n[20] Kevin R McKee, Joel Z Leibo, Charlie Beattie, and Richard Everett. Quantifying the effects of environment and population diversity in multi-agent reinforcement learning. Autonomous Agents and Multi-Agent Systems, 36(1):1–16, 2022. \n[21] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. \n[22] OpenAI, Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemysław D˛ebiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, Rafal Józefowicz, Scott Gray, Catherine Olsson, Jakub Pachocki, Michael Petrov, Henrique Pondé de Oliveira Pinto, Jonathan Raiman, Tim Salimans, Jeremy Schlatter, Jonas Schneider, Szymon Sidor, Ilya Sutskever, Jie Tang, Filip Wolski, and Susan Zhang. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680, 2019. \n[23] Peng Peng, Ying Wen, Yaodong Yang, Quan Yuan, Zhenkun Tang, Haitao Long, and Jun Wang. Multiagent bidirectionally-coordinated nets: Emergence of human-level coordination in learning to play starcraft combat games. arXiv preprint arXiv:1703.10069, 2017. \n[24] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. \n[25] Ali Shafti, Jonas Tjomsland, William Dudley, and A Aldo Faisal. Real-world human-robot collaborative reinforcement learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 11161–11166, 2020. \n[26] Victor do Nascimento Silva and Luiz Chaimowicz. Moba: A new arena for game ai. arXiv preprint arXiv:1705.10443, 2017. \n[27] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. \n[28] David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of Go without human knowledge. Nature, 550(7676):354–359, 2017. \n[29] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multi-agent cooperative and competitive tasks. arXiv preprint arXiv:1812.09755, 2018. \n[30] Ho Chit Siu, Jaime Peña, Edenna Chen, Yutai Zhou, Victor Lopez, Kyle Palko, Kimberlee Chang, and Ross Allen. Evaluation of human-ai teams for learned and rule-based agents in hanabi. Advances in Neural Information Processing Systems, 34:16183–16195, 2021. \n[31] DJ Strouse, Kevin McKee, Matt Botvinick, Edward Hughes, and Richard Everett. Collaborating with humans without human data. Advances in Neural Information Processing Systems, 34, 2021. \n[32] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multi-agent communication with backpropagation. Advances in Neural Information Processing Systems, 29, 2016. \n[33] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016. \n[34] Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michaël Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, Petko Georgiev, et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):350–354, 2019. \n[35] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas. Dueling network architectures for deep reinforcement learning. In Proceedings of the International Conference on Machine Learning, pages 1995–2003, 2016. \n[36] Rundong Wang, Xu He, Runsheng Yu, Wei Qiu, Bo An, and Zinovi Rabinovich. Learning efficient multi-agent communication: An information bottleneck approach. In Proceedings of the International Conference on Machine Learning, pages 9908–9918, 2020. \n[37] Bin Wu. Hierarchical macro strategy model for moba game ai. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 1206–1213, 2019. \n[38] Deheng Ye, Guibin Chen, Wen Zhang, Sheng Chen, Bo Yuan, Bo Liu, Jia Chen, Zhao Liu, Fuhao Qiu, Hongsheng Yu, et al. Towards playing full moba games with deep reinforcement learning. Advances in Neural Information Processing Systems, 33:621–632, 2020. \n[39] Deheng Ye, Guibin Chen, Peilin Zhao, Fuhao Qiu, Bo Yuan, Wen Zhang, Sheng Chen, Mingfei Sun, Xiaoqian Li, Siqin Li, et al. Supervised learning achieves human-level performance in moba games: A case study of honor of kings. IEEE Transactions on Neural Networks and Learning Systems, 2020. \n[40] Deheng Ye, Zhao Liu, Mingfei Sun, Bei Shi, Peilin Zhao, Hao Wu, Hongsheng Yu, Shaojie Yang, Xipeng Wu, Qingwei Guo, et al. Mastering complex control in moba games with deep reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6672–6679, 2020. \n[41] Rui Zhao, Jinming Song, Haifeng Hu, Yang Gao, Yi Wu, Zhongqian Sun, and Wei Yang. Maximum entropy population based training for zero-shot human-ai coordination. arXiv preprint arXiv:2112.11701, 2021. ",
|
| 962 |
+
"bbox": [
|
| 963 |
+
158,
|
| 964 |
+
75,
|
| 965 |
+
828,
|
| 966 |
+
921
|
| 967 |
+
],
|
| 968 |
+
"page_idx": 10
|
| 969 |
+
},
|
| 970 |
+
{
|
| 971 |
+
"type": "text",
|
| 972 |
+
"text": "",
|
| 973 |
+
"bbox": [
|
| 974 |
+
158,
|
| 975 |
+
85,
|
| 976 |
+
826,
|
| 977 |
+
569
|
| 978 |
+
],
|
| 979 |
+
"page_idx": 11
|
| 980 |
+
},
|
| 981 |
+
{
|
| 982 |
+
"type": "text",
|
| 983 |
+
"text": "Checklist ",
|
| 984 |
+
"text_level": 1,
|
| 985 |
+
"bbox": [
|
| 986 |
+
169,
|
| 987 |
+
577,
|
| 988 |
+
253,
|
| 989 |
+
592
|
| 990 |
+
],
|
| 991 |
+
"page_idx": 11
|
| 992 |
+
},
|
| 993 |
+
{
|
| 994 |
+
"type": "text",
|
| 995 |
+
"text": "1. For all authors... ",
|
| 996 |
+
"bbox": [
|
| 997 |
+
214,
|
| 998 |
+
603,
|
| 999 |
+
339,
|
| 1000 |
+
617
|
| 1001 |
+
],
|
| 1002 |
+
"page_idx": 11
|
| 1003 |
+
},
|
| 1004 |
+
{
|
| 1005 |
+
"type": "text",
|
| 1006 |
+
"text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] See the abstract and the contributions in Section 1. \n(b) Did you describe the limitations of your work? [Yes] For margin reasons, we will discuss limitations here. We have currently only verified the effectiveness of the MCC framework in MOBA games, and we will explore in more types of complex games, such as First-Person Shooting (FPS) and Massively Multiplayer Online (MMO) in the future. \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] The purpose of our method is only for academic research based on game environments, e.g., investigating the MOBA game-playing problems. Like AlphaGo or AlphaStar, the potential negative societal impacts of our work will be limited to the development of gaming AI applications. However, our research will contribute to the research community, the game industry, and the e-sports community. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
|
| 1007 |
+
"bbox": [
|
| 1008 |
+
235,
|
| 1009 |
+
619,
|
| 1010 |
+
826,
|
| 1011 |
+
838
|
| 1012 |
+
],
|
| 1013 |
+
"page_idx": 11
|
| 1014 |
+
},
|
| 1015 |
+
{
|
| 1016 |
+
"type": "text",
|
| 1017 |
+
"text": "2. If you are including theoretical results... ",
|
| 1018 |
+
"bbox": [
|
| 1019 |
+
214,
|
| 1020 |
+
843,
|
| 1021 |
+
493,
|
| 1022 |
+
857
|
| 1023 |
+
],
|
| 1024 |
+
"page_idx": 11
|
| 1025 |
+
},
|
| 1026 |
+
{
|
| 1027 |
+
"type": "text",
|
| 1028 |
+
"text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ",
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
238,
|
| 1031 |
+
861,
|
| 1032 |
+
738,
|
| 1033 |
+
892
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 11
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "text",
|
| 1039 |
+
"text": "3. If you ran experiments... ",
|
| 1040 |
+
"bbox": [
|
| 1041 |
+
212,
|
| 1042 |
+
897,
|
| 1043 |
+
393,
|
| 1044 |
+
911
|
| 1045 |
+
],
|
| 1046 |
+
"page_idx": 11
|
| 1047 |
+
},
|
| 1048 |
+
{
|
| 1049 |
+
"type": "text",
|
| 1050 |
+
"text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] The code and the data are proprietary. \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See the full experimental setup in Section 5.1.1 and Appendix. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] All results are reported over five random seeds (see Section 5.1.2). For the convenience of presentation, the median is shown in AI Performance, while detailed error bars are presented in Ablation Studies. Due to the high cost of Human-AI Game Test, we only used the median model for testing. \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See the resources in Section 5.1.1. ",
|
| 1051 |
+
"bbox": [
|
| 1052 |
+
238,
|
| 1053 |
+
92,
|
| 1054 |
+
825,
|
| 1055 |
+
265
|
| 1056 |
+
],
|
| 1057 |
+
"page_idx": 12
|
| 1058 |
+
},
|
| 1059 |
+
{
|
| 1060 |
+
"type": "text",
|
| 1061 |
+
"text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
|
| 1062 |
+
"bbox": [
|
| 1063 |
+
218,
|
| 1064 |
+
270,
|
| 1065 |
+
825,
|
| 1066 |
+
284
|
| 1067 |
+
],
|
| 1068 |
+
"page_idx": 12
|
| 1069 |
+
},
|
| 1070 |
+
{
|
| 1071 |
+
"type": "text",
|
| 1072 |
+
"text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [Yes] \n(c) Did you include any new assets either in the supplemental material or as a URL? [No] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [Yes] Honor of Kings is a released and recognized \"testbed\" for MOBA-game-playing problems [40, 39, 38, 8, 37]. We contact the relevant author and obtain authorization from the game provider. \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] See Appendix. All data in this paper is gamerelated and has nothing to do with identity information. ",
|
| 1073 |
+
"bbox": [
|
| 1074 |
+
238,
|
| 1075 |
+
287,
|
| 1076 |
+
825,
|
| 1077 |
+
436
|
| 1078 |
+
],
|
| 1079 |
+
"page_idx": 12
|
| 1080 |
+
},
|
| 1081 |
+
{
|
| 1082 |
+
"type": "text",
|
| 1083 |
+
"text": "5. If you used crowdsourcing or conducted research with human subjects... ",
|
| 1084 |
+
"bbox": [
|
| 1085 |
+
215,
|
| 1086 |
+
440,
|
| 1087 |
+
705,
|
| 1088 |
+
455
|
| 1089 |
+
],
|
| 1090 |
+
"page_idx": 12
|
| 1091 |
+
},
|
| 1092 |
+
{
|
| 1093 |
+
"type": "text",
|
| 1094 |
+
"text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes] See Appendix. In the Human-AI Game Test, all participants have more than three years of experience in Honor of Kings and are familiar with all the information in the game. Before the test, we also inform the participants of the detailed test instructions, and the participants voluntarily choose whether to participate in the test. We have detailed ethics descriptions in Appendix A.9. As stated in Section A.9.2, participants were given instructions before testing. As mentioned in point 5 (Line 87-88), participants’ game statistics will be only used for academic research, and participants can choose whether to participate or not. ",
|
| 1095 |
+
"bbox": [
|
| 1096 |
+
238,
|
| 1097 |
+
459,
|
| 1098 |
+
825,
|
| 1099 |
+
599
|
| 1100 |
+
],
|
| 1101 |
+
"page_idx": 12
|
| 1102 |
+
},
|
| 1103 |
+
{
|
| 1104 |
+
"type": "text",
|
| 1105 |
+
"text": "(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [Yes] We had performed a process similar to IRB before the test was conducted in this paper. The institution and all participants have approved our research. First, we analyze the risks of these experiments to the participants. The risks mainly include the leakage of identity information and the time cost. Then, a series of measures are implemented to prevent these risks in Appendix A.9.2. We make a risk statement for participants and sign an identity information confidentiality agreement. We only use information related to the game state in our research without identity information. In addition, special equipment and accounts are provided to the participants to prevent leakage of equipment and account information during the test. The identity information of all participants is not disclosed to the public. ",
|
| 1106 |
+
"bbox": [
|
| 1107 |
+
238,
|
| 1108 |
+
602,
|
| 1109 |
+
825,
|
| 1110 |
+
767
|
| 1111 |
+
],
|
| 1112 |
+
"page_idx": 12
|
| 1113 |
+
},
|
| 1114 |
+
{
|
| 1115 |
+
"type": "text",
|
| 1116 |
+
"text": "(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [Yes] In the Human-AI Game Test, participants can get 5 dollars for each match, and each match is about 10 to 20 minutes. ",
|
| 1117 |
+
"bbox": [
|
| 1118 |
+
238,
|
| 1119 |
+
768,
|
| 1120 |
+
823,
|
| 1121 |
+
811
|
| 1122 |
+
],
|
| 1123 |
+
"page_idx": 12
|
| 1124 |
+
}
|
| 1125 |
+
]
|
parse/dev/2nJdh_C-UWe/2nJdh_C-UWe_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/2nJdh_C-UWe/2nJdh_C-UWe_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/3tYvDb4dwab/3tYvDb4dwab.md
ADDED
|
@@ -0,0 +1,292 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# UNDERSTANDING SELF-SUPERVISED PRETRAINING WITH PART-AWARE REPRESENTATION LEARNING
|
| 2 |
+
|
| 3 |
+
Anonymous authors Paper under double-blind review
|
| 4 |
+
|
| 5 |
+
# ABSTRACT
|
| 6 |
+
|
| 7 |
+
In this paper, we are interested in understanding self-supervised pretraining through studying the capability that self-supervised representation pretraining methods learn part-aware representations. The study is mainly motivated by that random views, used in contrastive learning, and random masked (visible) patches, used in masked image modeling, are often about object parts.
|
| 8 |
+
|
| 9 |
+
We explain that masked image modeling is a part-to-part task: the masked patches of the object are hallucinated from the visible patches, and that contrastive learning is a part-to-whole task: the projection layer hallucinates the whole object representation from the object part representation learned from the encoder. The explanation suggests that the self-supervised pretrained encoder is required to understand the object part. We empirically compare the off-the-shelf encoders pretrained with several representative methods on object-level recognition and part-level recognition. The results show that the fully-supervised model outperforms self-supervised models for object-level recognition, and most self-supervised contrastive learning and masked image modeling methods outperform the fully-supervised method for part-level recognition. It is observed that the combination of contrastive learning and masked image modeling further improves the performance.
|
| 10 |
+
|
| 11 |
+
# 1 INTRODUCTION
|
| 12 |
+
|
| 13 |
+
Self-supervised representation pretraining has been attracting a lot of research efforts recently. The goal is to train an encoder that maps an image to a representation from visual contents without the necessity of human annotation, expecting that the encoder benefits the downstream tasks, e.g., segmentation and detection.
|
| 14 |
+
|
| 15 |
+
There are two main frameworks: contrastive learning1 and masked image modeling. Contrastive learning aims to maximize the agreement of the embeddings of random augmented views from the same image. Masked image modeling partitions an image into masked patches and visible patches, and makes predictions for masked patches from visible patches. Figure 1 gives examples of random views for contrastive learning and masked and visible patches for masked image modeling.
|
| 16 |
+
|
| 17 |
+
We observe that a random view and a set of masked (visible) patches usually contain a portion of an object. It is also reported in self-supervised learning methods, e.g., DINO (Caron et al., 2021) and iBOT (Zhou et al., 2021), that different attention heads in ViTs can attend to different semantic regions or parts of an object. In light of this, we attempt to understand self-supervised pretraining by studying the capability that the pretrained encoder learns part representations.
|
| 18 |
+
|
| 19 |
+
We present a part-to-whole explanation for typical contrastive learning methods (e.g., SimCLR (Chen et al., 2020), MoCo (Chen et al., 2021), and BYOL (Grill et al., 2020)): the embedding of the whole object is hallucinated from the embedding of the part of the object contained in the random crop through a projection layer. In this way, embeddings of random crops from the same image naturally agrees with each other. Masked image modeling is a part-to-part process: the embeddings of the masked patches of the object (a part of the object), are hallucinated from the visible patches (the other part of the object).
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
Figure 1: (a) original image, (b-c) two random crops, and (d-e) masked and visible patches.
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
Figure 2: Top-24 patch retrieval results with three frozen encoders of DeiT, MoCo v3, and CAE, by taking the patch in the red box as the query. It can be seen that the retrieved results from CAE and MoCo v3 are about the object part (wing and dog mouth) and more precise than DeiT (about the whole object) implying that self-supervised pretraining methods, CAE and MoCo v3 are stronger at learning part-aware representations than the fully-supervised method DeiT.
|
| 26 |
+
|
| 27 |
+
We empirically compare the supervised model DeiT (Touvron et al., 2020) and typical self-supervised representation pretraining methods, including MoCo v3 (Chen et al., 2021), DINO (Caron et al., 2021), CAE (Chen et al., 2022a), MAE (He et al., 2021), BEiT (Bao et al., 2021), and iBOT (Zhou et al., 2021), on object-level recognition (image classification and object segmentation) and part-level recognition (patch retrieval, patch classification, and part segmentation). Figure 2 presents patch retrieval results using the encoders learned through CAE, MoCo v3, and DeiT, implying that the encoders pretrained by CAE and MoCo v3 are able to learn part-aware representations.
|
| 28 |
+
|
| 29 |
+
Through extensive studies and comparisons, we make the following observations. 1) DeiT outperforms contrastive learning and MIM methods except iBOT in object-level recognition tasks, which may benefit from its explicit object-level supervision. 2) In contrast, self-supervised methods learn better part-aware representations than DeiT. For example, while DeiT is superior to DINO and CAE by $0 . 4 \%$ and $2 . 3 \%$ on ADE20K object segmentation, DINO and CAE outperform DeiT by $1 . 6 \%$ and $1 . 1 \%$ on ADE20K part segmentation, respectively. 3) In contrastive learning, the encoder can learn part-aware information, while the projected representation tends to be more about the whole object. The evidence could be found in part retrieval experiments on MoCo v3, DINO, and iBOT. 4) The MIM method CAE shows good potential in part-aware representation learning. Interestingly, the method combines contrastive learning and MIM is promising, e.g., iBOT learns better representations at both object and part levels.
|
| 30 |
+
|
| 31 |
+
To summarize, this paper presents the following contributions:
|
| 32 |
+
|
| 33 |
+
• We study the capability of learning part-aware representations as a way of understanding self-supervised representation pretraining.
|
| 34 |
+
• We explain masked image modeling as a part-to-part task and contrastive learning as a partto-whole task, and speculate that self-supervised pretraining has the potential for learning part-aware representations.
|
| 35 |
+
|
| 36 |
+
• We empirically compare several pretrained models on object-level and part-level recognition tasks, showing interesting findings with supporting evidence of the capability of part-aware representation learning for self-supervised learning.
|
| 37 |
+
|
| 38 |
+
# 2 RELATED WORK
|
| 39 |
+
|
| 40 |
+
Contrastive learning. Contrastive pretraining has been an intense academic field in the CNN era. In this work, we use it to refer to methods for comparing random views (Caron et al., 2020; Chen et al., 2020; Zbontar et al., 2021; Xie et al., 2021a; Chen et al., 2021; Caron et al., 2021), including some instance discrimination work such as (Grill et al., 2020; Chen & He, 2021; Bardes et al., 2021). As one of the representative works, SimCLR (Chen et al., 2020) learns representations through maximizing agreement between different views of the same image in the latent space. BYOL (Grill et al., 2020) uses two asymmetrical networks to bootstrap latent representation without negative samples involved during the interaction. As vision transformer (ViT) (Dosovitskiy et al., 2021) shows excellent performance via supervised learning, it is adopted subsequently in contrastive pertaining, and numerous outstanding works are proposed. For example, MoCo v3 (Chen et al., 2021) observes the hidden instability while training self-supervised ViT and solves it by using a fixed random patch projection. DINO (Caron et al., 2021) explores new properties derived from self-supervised ViT and accordingly designs a learning strategy interpreted as a form of self-distillation with no labels.
|
| 41 |
+
|
| 42 |
+
Masked image modeling (MIM). Masked image modeling is another self-supervised pretraining paradigm that attracts much attention recently. BEiT (Bao et al., 2021) follows masked language modeling in the natural language process (NLP) area and predicts tokens via mapping image patches by d-VAE (Ramesh et al., 2021). PeCo (Dong et al., 2021) boosts BEiT by taking into consideration more semantic information in visual tokens. MAE (He et al., 2021) learns rich hidden information by directly performing masked image reconstruction in RGB color space using ViT while SimMIM (Xie et al., 2021b) uses Swin-transformer (Liu et al., 2021). CAE (Chen et al., 2022a) adds a regressor between encoder and decoder, which is designed to align unmasked patches with masked ones, leading to a pure context encoder. Recently, a trend that combines MIM with siamese frameworks has surfaced and showed encouraging results including MST (Li et al., 2021), SplitMask (El-Nouby et al., 2021), iBOT (Zhou et al., 2021), dBOT (Liu et al., 2022), and SIM (Tao et al., 2022).
|
| 43 |
+
|
| 44 |
+
Understanding self-supervised contrastive pretraining. The studies on understanding contrastive pretraining (Saunshi et al., 2022; Chen et al., 2022b; Zhong et al., 2022; Wei et al., 2022) mainly focus on random augmentations (views), contrastive loss function and its variants under the assumption that: the augmentations of inputs from the same class have significant overlap in the representation space, but there is little overlap for inputs from different classes. Our work is complementary to these studies. Inspired by the observation that random views usually contain a portion of an object, and methods (Caron et al., 2021; Zhou et al., 2021) show that different attention heads in ViTs can attend to different semantic regions of an object, we investigate what the encoder and the projector do in typical self-supervised contrastive pretraining. We speculate that the pretraining task is a part-to-whole problem, predicting the representation of the whole object through the projector from the representation (obtained from the encoder) of the part of an object. We use empirical results to verify our analysis.
|
| 45 |
+
|
| 46 |
+
Understanding self-supervised masked image modeling. The comparison of attention in different layers between the pretrained models from MIM and the supervised approach is conducted: MIM pretraining brings locality to the trained model with sufficient diversity on the attention heads (Xie et al., 2022a). Consistent with the analysis in NLP, empirical studies are conducted in Xie et al. (2022b) to verify that MIM benefits from larger models, more data, and longer training. CAE (Chen et al., 2022a) gives the comparison between contrastive and MIM and shows MIM cares about all patches and thus achieves better results for fine-tuning. Cao et al. (2022) provides a mathematical understanding of MIM. Kong & Zhang (2022) points out that the learned occlusion invariant feature contributes to the success of MIM. In this work, we speculate that masked image modeling is a partto-part process: the embeddings of the masked part of the object are hallucinated from the visible part using the position information of the masked patches, leading to better part-aware representation than the supervised model DeiT (Touvron et al., 2020).
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
Figure 3: The pipeline of a typical contrastive learning approach. Two augmented views, red box and blue box, are generated from the original image. The augmented view in red is fed into the encoder and the projector, and then the predictor (which does not appear in earlier works like MoCo (Chen et al., 2021) and SimCLR (Chen et al., 2020)), and the view in blue is fed into the encoder and the projector. The two outputs are expected to be aligned. The gradient is stopped for the bottom stream.
|
| 50 |
+
|
| 51 |
+
# 3 UNDERSTANDING CONTRASTIVE LEARNING AND MASK IMAGE MODELING
|
| 52 |
+
|
| 53 |
+
# 3.1 CONTRASTIVE LEARNING
|
| 54 |
+
|
| 55 |
+
Contrastive learning aims to learn the encoder through maximizing the agreement between differently augmented views of the same image in the representation space. An example pipeline is depicted in Figure 3. Given an image I, the augmentations, e.g., random cropping, random color distortion, and random Gaussian blur, are applied to generate a set of $N$ augmented views, $\{ \mathsf { V } _ { 1 } , \mathsf { V } _ { 2 } , \cdots , \mathsf { V } _ { N } \}$ . An augmented view $\mathsf { V } _ { n }$ is fed into an encoder Encoder, generating the encoded representation ${ \bf x } _ { n }$ , and followed by a projector, generating the projection $\mathbf { z } _ { n }$ . The basic goal is to maximize the agreement between the projections $\{ \mathbf { z } _ { 1 } , \mathbf { z } _ { 2 } , \cdots , \mathbf { z } _ { N } \}$ , i.e., minimize the loss
|
| 56 |
+
|
| 57 |
+
$$
|
| 58 |
+
{ \mathcal { L } } _ { \mathrm { C P T } } = \sum _ { i = 1 } ^ { N } \sum _ { j = 1 } ^ { N } { \ell } ( \mathrm { P r o j e c t o r ( E n c o d e r ( V } _ { i } ) ) , \mathrm { P r o j e c t o r ( E n c o d e r ( V } _ { j } ) ) ) .
|
| 59 |
+
$$
|
| 60 |
+
|
| 61 |
+
In the formulation with a contrastive loss, the agreement between the projections of random augmentations from different images is minimized.
|
| 62 |
+
|
| 63 |
+
Part-to-whole prediction explanation. Let us consider two crops randomly sampled from the original image (see the examples given in Figure 1(b-c)). The encoded representation of the first crop is expected to describe a part of the object dog; the encoded representation of the second crop is expected to describe another part of the object $\mathrm { d o g ^ { 2 } }$ . The two representations are related but different. Contrastive learning methods project the two encoded representations into two projected representations that are expected to agree. We hypothesize that the projection process maps the encoded part representation to the representation of the whole object3. Through this way, the projected representations will agree to different views from the same image. It is assumed that the part-to-whole projection is more reliable if the encoded representation is semantically richer and is able to describe the part information. The part-to-whole process suggests that the encoder pretrained by contrastive learning methods is potentially capable of learning part-aware representations.
|
| 64 |
+
|
| 65 |
+
Figure 4 provides patch search results of a representative contrastive learning method MoCo v3 (Chen et al., 2021) based on the encoded representations before and after the projections. One can see that the results through the encoded representations are mainly about the local part, and the results through the projections tend to include the other parts of the same object. In other words, the projections tend to be about the whole object. Similar observations are also shown in Chen et al. (2022b). The search results verify the part-to-whole hypothesis.
|
| 66 |
+
|
| 67 |
+
# 3.2 MASKED IMAGE MODELING
|
| 68 |
+
|
| 69 |
+
Mask image modeling is the task of predicting some parts of an image from the remaining parts. An augmented view of an image is partitioned into patches, $\mathcal { R } = \{ \mathsf { R } _ { 1 } , \mathsf { R } _ { 2 } , \ldots , \mathsf { R } _ { M } \}$ . The task is to predict a subset of patches $\mathcal { R } _ { m }$ , named masked patches, from the remaining patches $\mathcal { R } _ { v }$ , named visible patches. Considering contrastive learning that explicitly compares representations of random views, we take context autoencoder (CAE) (Chen et al., 2022a) as an example that explicitly predicts the encoded representations of the masked patches from the encoded representations of the visible patches4.
|
| 70 |
+
|
| 71 |
+
|
| 72 |
+
Figure 4: Illustration of patch search results using encoded representations and projections (pretrained with MoCo v3 as). Left: patch search results with encoded representations. Right: patch search results with projections. In each result, the small patch encircled by the red box is taken as the query. It can be seen that for encoded representations, the returned patches are about the same part, and for projections, the result patches are about the same object, verifying the part-to-whole hypothesis.
|
| 73 |
+
|
| 74 |
+
One goal of CAE (illustrated in Figure 5), which we call masked representation modeling (MRM), is to maximize the agreement between the predictions of the representations of masked patches (through a regressor) and the representation of masked patches computed from the encoder by minimizing the loss
|
| 75 |
+
|
| 76 |
+
$$
|
| 77 |
+
\ell _ { \mathrm { M R M } } ( \mathrm { R e g r e s s o r } ( \operatorname { E n c o d e r } ( \mathcal { P } _ { v } ) ) , \operatorname { E n c o d e r } ( \mathcal { P } _ { m } ) ) .
|
| 78 |
+
$$
|
| 79 |
+
|
| 80 |
+
Here, we do not include the positional embeddings of masked and visible patches for clarity. It is noted that MRM differs from contrastive learning: MRM does not compare multiple random views, but compares the regressed representations for masked patches and the encoded representations of masked patches. In addition, there is another loss for target prediction (reconstruction) for the masked patches, which is commonly used in masked image modeling (MIM) methods:
|
| 81 |
+
|
| 82 |
+
$$
|
| 83 |
+
\ell _ { \mathrm { M I M } } \big ( \mathrm { D e c o d e r } \big ( \mathrm { R e g r e s s o r } \big ( \mathrm { E n c o d e r } ( \mathcal { P } _ { v } ) \big ) \big ) , \mathrm { T a r g e t } ( \mathcal { P } _ { m } ) \big ) ,
|
| 84 |
+
$$
|
| 85 |
+
|
| 86 |
+
where ${ \mathrm { T a r g e t } } ( { \mathcal { P } } _ { m } )$ is a function to map the masked patches to the targets, e.g., d-VAE (Ramesh et al., 2021) token used in CAE and BeiT (Bao et al., 2021), or normalized RGB values used in MAE (He et al., 2021).
|
| 87 |
+
|
| 88 |
+
Part-to-part prediction explanation. The masked image modeling approaches, including CAE, MAE, and BEiT, make use of the positions of masked patches for making predictions for masked patches from visible patches. The visible patches and masked patches often contain different parts of an object. In other words, MIM aims to predict the masked part of an object from the visible part. We name this a part-to-part process. There are two part-to-part tasks: one is to reconstruct the part targets from the visible part representations (MAE and CAE) or from the visible part raw pixels (BEiT), and the other one is to regress the masked part representations (CAE). The part-to-part process suggests that the encoder pretrained by MIM methods is potentially capable of learning part-aware representations. Figure 2 illustrates the capability with the patch retrieval results.
|
| 89 |
+
|
| 90 |
+
|
| 91 |
+
Figure 5: The pipeline of an MIM approach, context autoencoder (CAE). An augmented view (in blue) of the image is partitioned into visible and masked patches. The CAE approach feeds visible patches into the encoder and extracts their representations $\mathbf { Z } _ { v }$ and then completes the pretext task by predicting the representations $\mathbf { Z } _ { m }$ of the masked patches from the visible patches in the encoded representation space with latent contextual regressor and alignment constraint, and mapping predicted representations $\mathbf { Z } _ { m }$ of masked patches to the targets. The pretrained encoder in (a) is applied to downstream tasks by simply replacing the pretext task part (b, c) with the downstream task completion part.
|
| 92 |
+
|
| 93 |
+
Table 1: Top-1 accuracy with linear probing, and attentive probing (Chen et al., 2022a), on the ImageNet classification benchmark (Deng et al., 2009).
|
| 94 |
+
|
| 95 |
+
<table><tr><td>Method</td><td>Linear</td><td>Attentive</td></tr><tr><td colspan="3">Supervised Model:</td></tr><tr><td>DeiT</td><td>81.8</td><td>81.8</td></tr><tr><td colspan="3">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>76.2</td><td>77.0</td></tr><tr><td>DINO</td><td>77.3</td><td>77.8</td></tr><tr><td colspan="3">Masked Image Modeling (MIM):</td></tr><tr><td>BEiT</td><td>41.8</td><td>51.9</td></tr><tr><td>MAE</td><td>67.8</td><td>74.2</td></tr><tr><td>CAE</td><td>70.4</td><td>77.1</td></tr><tr><td colspan="3">Contrastive Learning+MIM:</td></tr><tr><td>iBOT</td><td>79.5</td><td>79.8</td></tr></table>
|
| 96 |
+
|
| 97 |
+
Table 2: Linear evaluation of ADE20K (Zhou et al., 2019) object-level semantic segmentation (150 classes) using $4 \times$ upsampling and a single $1 \times 1$ convolutional layer on frozen backbones.
|
| 98 |
+
|
| 99 |
+
<table><tr><td>Method</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td>Supervised Model:</td><td></td><td></td><td></td></tr><tr><td>DeiT</td><td>34.9</td><td>44.2</td><td>75.4</td></tr><tr><td>Contrastive Learning:</td><td></td><td></td><td></td></tr><tr><td>MoCo v3</td><td>34.7</td><td>43.9</td><td>75.9</td></tr><tr><td>DINO</td><td>34.5</td><td>43.5</td><td>76.1</td></tr><tr><td colspan="4">Masked Image Modeling (MIM):</td></tr><tr><td>BEiT</td><td>17.8</td><td>23.7</td><td>64.9</td></tr><tr><td>MAE</td><td>27.1</td><td>34.8</td><td>71.6</td></tr><tr><td>CAE</td><td>32.6</td><td>42.2</td><td>75.2</td></tr><tr><td colspan="4">Contrastive Learning +MIM:</td></tr><tr><td>iBOT</td><td>38.3</td><td>47.4</td><td>78.1</td></tr></table>
|
| 100 |
+
|
| 101 |
+
# 4 EXPERIMENTS
|
| 102 |
+
|
| 103 |
+
We study seven representative methods with the same ViT-B encoder, including a supervised method DeiT (Touvron et al., 2020); contrastive learning methods MoCo v3 (Chen et al., 2021), DINO (Caron et al., 2021); masked image modeling (MIM) methods BEiT (Bao et al., 2021), MAE (He et al., 2021), and CAE (Chen et al., 2022a); and iBOT (Zhou et al., 2021) that combines contrastive learning and MIM. We take the training epochs specified in each work to ensure that all compared models are properly trained: 300 for DeiT, 300 $( 6 0 \dot { 0 } ^ { 5 } )$ for MoCo v3, 400 $( 1 6 0 0 ^ { 5 } )$ ) for DINO and iBOT, 800 for BEiT, and 1600 for MAE and CAE. Frozen encoders are used in all experiments to understand what these different representation pretraining methods learn. More details can be found in Appendix A.1.
|
| 104 |
+
|
| 105 |
+
# 4.1 OBJECT-LEVEL RECOGNITION
|
| 106 |
+
|
| 107 |
+
We benchmark two widely-studied object-level recognition, i.e., image classification and semantic segmentation to show the capability that the pretrained encoder learns object-level representations.
|
| 108 |
+
|
| 109 |
+
Table 3: Part retrieval (AP, $\%$ ) and classification (accuracy, $\%$ ) results on the cropped part patches of CUB-200-2011 and COCO. The “Encoded" and “Projected" refer to the encoded and projected representations. “Linear" and “Attentive" columns denote the linear probing and attentive probing accuracy, respectively.
|
| 110 |
+
|
| 111 |
+
<table><tr><td rowspan="3">Methods</td><td colspan="4">PartRetrieval</td><td colspan="4">Part Classification</td></tr><tr><td colspan="2">CUB-200-2011</td><td colspan="2">COCo</td><td colspan="2">CUB-200-2011</td><td colspan="2">COCO</td></tr><tr><td>Encoded</td><td>Projected</td><td>Encoded</td><td>Projected</td><td>Linear</td><td>Attentive</td><td>Linear</td><td>Attentive</td></tr><tr><td colspan="9">Supervised Model:</td></tr><tr><td>DeiT</td><td>35.0</td><td></td><td>44.1</td><td></td><td>90.9</td><td>92.9</td><td>88.5</td><td>91.4</td></tr><tr><td colspan="9">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>50.8</td><td>28.4</td><td>52.3</td><td>36.8</td><td>93.8</td><td>96.0</td><td>92.4</td><td>95.3</td></tr><tr><td>DINO</td><td>48.9</td><td>31.7</td><td>51.8</td><td>41.2</td><td>93.2</td><td>95.2</td><td>91.7</td><td>94.5</td></tr><tr><td colspan="9">Masked Image Modeling (MIM):</td></tr><tr><td>BEiT</td><td>27.9</td><td>1</td><td>35.3</td><td>1</td><td>55.4</td><td>86.5</td><td>69.3</td><td>86.5</td></tr><tr><td>MAE</td><td>28.5</td><td>1</td><td>37.1</td><td></td><td>86.9</td><td>92.8</td><td>88.0</td><td>93.9</td></tr><tr><td>CAE</td><td>58.0</td><td>1</td><td>57.0</td><td>1</td><td>89.5</td><td>95.8</td><td>91.1</td><td>95.5</td></tr><tr><td colspan="9">Contrastive Learning +MIM:</td></tr><tr><td>iBOT</td><td>49.3</td><td>31.2</td><td>59.2</td><td>41.5</td><td>93.8</td><td>95.8</td><td>92.1</td><td>95.1</td></tr></table>
|
| 112 |
+
|
| 113 |
+
Image classification. We report the linear probing, and attentive probing results of the selected models on ImageNet (Deng et al., 2009). For attentive probing, we follow the protocol in CAE (Chen et al., 2022a) that append a cross-attention layer together with a batch normalization layer and a linear classifier.
|
| 114 |
+
|
| 115 |
+
We have the following observations from Table 1. 1) The supervised model, DeiT performs better than self-supervised models at object-level recognition. 2) The models that leverage contrastive learning, i.e., MoCo, DINO, and iBOT, show superior linear probing performance than MIM-based models, demonstrating they contain more object-aware high-level semantics. 3) MIM-based models, e.g., CAE, show inferior results in linear probing while competitive results with contrastive-based methods in attentive probing. The reason might be that MIM is capable of attending to all the regions, including non-object regions in an image, thus needs a spatial feature selection step to attend to the object part, which is pointed out in Chen et al. (2022a). BEiT and MAE perform inferior, implying that the two methods are less capable of learning semantics.
|
| 116 |
+
|
| 117 |
+
Object-level semantic segmentation. We perform linear evaluation on ADE20K (Zhou et al., 2019) to show the object-level semantic capabilities of the pretrained models. A $4 \times$ bilinear interpolation and a single $1 \times 1$ convolutional layer for pixel labeling are attached to the frozen encoder.
|
| 118 |
+
|
| 119 |
+
We can see from Table 2 that the supervised model DeiT outperforms all self-supervised models except iBOT, including contrastive learning and MIM methods on ADE20K object-level segmentation. This implies that in general the self-supervised models are not strong at object-level understanding, which is consistent with the observations for image classification. iBOT (Zhou et al., 2021), as a combination of contrastive learning and MIM, shows surprisingly better performance than the supervised model DeiT on ADE20K, implying the power of combining contrastive learning and masked image modeling for downstream tasks.
|
| 120 |
+
|
| 121 |
+
# 4.2 PART-LEVEL RECOGNITION
|
| 122 |
+
|
| 123 |
+
Self-supervised methods like iBOT (Zhou et al., 2021) and DINO (Caron et al., 2021) qualitatively show that different attention heads in ViTs can attend to different semantic regions of an object. We conduct the quantitative evaluation for part-aware representation obtained by pretrained models that is not well explored before, through three part-level recognition tasks, part retrieval, part classification, and part segmentation.
|
| 124 |
+
|
| 125 |
+
Part retrieval. We conduct part retrieval experiments on two datasets, CUB-200-2011 (Wah et al., 2011) and COCO (Lin et al., 2014). We build the part patch databases by cropping the patches centered at the keypoint. We consider four and three keypoints from the two datasets, respectively. For each keypoint, we find the minimum L2 distance $( d )$ from the distances between it and all the other keypoints in the same image, then crop a $d \times d$ patch centered at this keypoint and resize it to $2 2 4 \times 2 2 4$ . We use the cosine distance as the patch distance and evaluate the retrieval performance using average precision (AP) as the retrieval metric.
|
| 126 |
+
|
| 127 |
+
Table 4: Part-level linear semantic segmentation results on ADE20K-Part, Pascal-Part, and LIP datasets.
|
| 128 |
+
|
| 129 |
+
<table><tr><td rowspan="3">Methods</td><td colspan="3">ADE20K-Part 209 Part Classes</td><td colspan="3">Pascal-Part 193 Part Classes</td><td colspan="3">LIP 19 Part Classes</td></tr><tr><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td colspan="10">Supervised Model:</td></tr><tr><td>DeiT</td><td>27.3</td><td>34.7</td><td>69.2</td><td>27.4</td><td>36.1</td><td>65.8</td><td>41.4</td><td>52.6</td><td>73.5</td></tr><tr><td colspan="10">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>27.1</td><td>34.7</td><td>70.1</td><td>27.1</td><td>35.8</td><td>66.0</td><td>41.9</td><td>53.0</td><td>74.5</td></tr><tr><td>DINO</td><td>28.9</td><td>36.8</td><td>70.3</td><td>27.8</td><td>36.5</td><td>66.4</td><td>41.0</td><td>51.9</td><td>74.0</td></tr><tr><td colspan="10">Masked ImageModeling ( (MIM):</td></tr><tr><td>BEiT</td><td>18.6</td><td>25.8</td><td>58.2</td><td>14.8</td><td>21.4</td><td>47.0</td><td>27.2</td><td>36.5</td><td>60.1</td></tr><tr><td>MAE</td><td>26.3</td><td>35.0</td><td>67.3</td><td>24.3</td><td>32.9</td><td>61.5</td><td>38.2</td><td>48.7</td><td>71.3</td></tr><tr><td>CAE</td><td>28.4</td><td>36.9</td><td>71.1</td><td>27.8</td><td>37.0</td><td>66.3</td><td>43.7</td><td>55.1</td><td>75.9</td></tr><tr><td colspan="10">Contrastive Learning + MIM:</td></tr><tr><td>iBOT</td><td>32.2</td><td>40.0</td><td>73.4</td><td>30.7</td><td>40.0</td><td>69.7</td><td>44.6</td><td>55.7</td><td>76.6</td></tr></table>
|
| 130 |
+
|
| 131 |
+
The results are provided in Table 3. We have the following observations. 1) Self-supervised models except BEiT and MAE outperform the supervised model DeiT, indicating the capability that contrastive learning and CAE learn part-aware representations. BEiT and MAE perform inferior, consistent to the observations in ImageNet classification in Table 1. 2) iBOT performs the best, and the reason might be that the capability of learning part-aware representations is boosted by making use of both contrastive learning and masked image modeling.
|
| 132 |
+
|
| 133 |
+
We also report the part retrieval performance of the projected representations of contrastive learning methods in Table 3. The performance is much lower than the encoded representations. This provides an extra evidence for the part-to-whole hypothesis of contrastive learning: the projected representations are more about the whole object.
|
| 134 |
+
|
| 135 |
+
Part classification. We further conduct part classification experiments on the datasets used for part retrieval. We consider two kinds of extra learnable layers, linear probing and attentive probing, for classification. The results in Table 3 show that: 1) While DeiT performs the best in the image classification task (see Table 1), for part classification, contrastive-based methods like MoCo v3, DINO, and iBOT outperform DeiT by more than $2 \%$ under both linear and attentive probing settings. 2) Though MIM-based models CAE and MAE are inferior to DeiT in object-level classification (e.g., more than $10 \%$ and $4 \%$ lower in linear and attentive probing), they show competitive performance in linear probing and higher results than DeiT in attentive probing, demonstrating they learn better partaware representations. 3) BEiT is inferior to other works, and iBOT has good performance, implying that the probing quality of pretrained encoders is a good indicator for downstream performance.
|
| 136 |
+
|
| 137 |
+
Part segmentation. We perform part-level linear semantic segmentation to study the finer-grained part representation modeling capability of different pretraining paradigms on three widely used datasets: ADE20K-Part (Zhou et al., 2019) containing 209 parts from the ADE20K dataset (Zhou et al., 2019), Pascal-Part (Chen et al., 2014) including 193 part categories, and LIP (Gong et al., 2017) consisting of 19 semantic human part labels. Similar to the object-level semantic segmentation experiments, linear evaluation is employed here. We maintain the same training protocols for all methods for fair comparisons. See Appendix A.2 and A.4 for dataset and training details.
|
| 138 |
+
|
| 139 |
+
The results are reported in Table 4 with the following observations. 1) Contrastive learning models, i.e., MoCo v3 and DINO, achieve competitive performance with the supervised model DeiT: DINO outperforms DeiT on ADE20K-Part and Pascal-Part, and MoCo v3 outperforms DeiT on LIP. 2) The MIM model CAE, outperforms DeiT by large margins on all three datasets, e.g., $1 . 1 \%$ on ADE20K-Part and $2 . 3 \%$ on LIP, indicating CAE learns good part-aware representations. Similar to part retrieval, possibly due to pretraining quality in representation encoding, BEiT and MAE perform inferior. 3) Compared with object-level segmentation results in Table 2, DeiT learns better object-level semantics by explicit supervision than both contrastive learning and MIM, however, it is generally inferior to self-supervised models on part segmentation. 4) The model iBOT, which leverages both contrastive learning and MIM, outperforms all other works on three datasets, demonstrating its powerful capability in learning finer part-level semantics. Combining the two self-supervised learning techniques is thus a promising direction.
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
Figure 6: Comparisons between object-level and part-level semantic segmentation on ADE20K and Pascal-Part datasets. Though the supervised DeiT is superior over self-supervised models (i.e., MoCo v3, DINO, MAE, CAE) on object-level segmentation, it is generally inferior to self-supervised models on part segmentation, demonstrating self-supervised methods learn good part-aware representations. iBOT enjoys the benefits of contrastive learning and MIM. See Appendix A.3 for detailed results.
|
| 143 |
+
|
| 144 |
+
In summary, we show that self-supervised methods are potentially capable of learning part-aware representations. Among them, CAE is a representative MIM work, showing good performance by explicitly predicting the encoded representations of the masked patches in the encoding space; contrastive learning methods MoCo v3 and DINO outperform BEiT and MAE; and iBOT performs the best by combining contrastive learning and MIM. The observations are evidenced by three part-based segmentation benchmarks consistently.
|
| 145 |
+
|
| 146 |
+
# 4.3 OBSERVATION SUMMARY BETWEEN OBJECT-LEVEL AND PART-LEVEL SEGMENTATION
|
| 147 |
+
|
| 148 |
+
We conduct both object-level and part-level linear semantic segmentation on different hierarchies of the same dataset. Considering that the 209 classes in ADE20K-Part are basically chosen from 59 object classes, we denote the 59-object dataset as ADE20K-Object. Similarly, Pascal-Object consists of 16 object categories, corresponding to the 193 part categories in Pascal-Part.
|
| 149 |
+
|
| 150 |
+
The results in Figure 6 show that: although the supervised DeiT is superior over contrastive learning and masked image modeling methods on ADE20K-Object and Pascal-Object except iBOT, it is generally inferior to self-supervised models on ADE20K-Part and Pascal-Part, demonstrating selfsupervised methods can learn good part-aware representations. Similar observations could be found from the object classification in Table 1 and part classification in Table 3.
|
| 151 |
+
|
| 152 |
+
In comparison to contrastive learning, CAE shows a stronger capability of learning part-aware representations, and a weaker capability of learning object-level semantics. The superiority of iBOT, a combination of contrastive learning and masked image modeling, demonstrates that it enjoys the benefits of contrastive learning and masked image modeling.
|
| 153 |
+
|
| 154 |
+
# 5 CONCLUSION
|
| 155 |
+
|
| 156 |
+
We attempt to study the capability of learning part-aware representations of self-supervised representation pretraining methods. We provide speculations for contrastive learning and masked image modeling: part-to-whole and part-to-part, with empirical results justifying the speculations. Our study presents an aspect to understand what self-supervised representation pretraining methods learn.
|
| 157 |
+
|
| 158 |
+
Future work. The strong capability of part-aware representation learning is one of the properties of self-supervised pretraining. There should be other characteristics that are leaved as the future work.
|
| 159 |
+
|
| 160 |
+
# REFERENCES
|
| 161 |
+
|
| 162 |
+
Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pre-training of image transformers. arXiv:2106.08254, 2021.
|
| 163 |
+
|
| 164 |
+
Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021.
|
| 165 |
+
|
| 166 |
+
Suzanna Becker and Geoffrey E. Hinton. A self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161–163, 1992.
|
| 167 |
+
|
| 168 |
+
Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Coco-stuff: Thing and stuff classes in context. In CVPR, pp. 1209–1218, 2018.
|
| 169 |
+
|
| 170 |
+
Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022.
|
| 171 |
+
|
| 172 |
+
Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. arXiv preprint arXiv:2006.09882, 2020.
|
| 173 |
+
|
| 174 |
+
Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.
|
| 175 |
+
|
| 176 |
+
Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119 of Proceedings of Machine Learning Research, pp. 1597–1607. PMLR, 2020.
|
| 177 |
+
|
| 178 |
+
Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, pp. 1971–1978, 2014.
|
| 179 |
+
|
| 180 |
+
Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022a.
|
| 181 |
+
|
| 182 |
+
Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pp. 15750–15758, 2021.
|
| 183 |
+
|
| 184 |
+
Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pp. 9640–9649, 2021.
|
| 185 |
+
|
| 186 |
+
Yubei Chen, Adrien Bardes, Zengyi Li, and Yann LeCun. Intra-instance vicreg: Bag of self-supervised image patch embedding. arXiv preprint arXiv:2206.08954, 2022b.
|
| 187 |
+
|
| 188 |
+
MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.
|
| 189 |
+
|
| 190 |
+
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pp. 248–255. Ieee, 2009.
|
| 191 |
+
|
| 192 |
+
Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021.
|
| 193 |
+
|
| 194 |
+
Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021.
|
| 195 |
+
|
| 196 |
+
Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021.
|
| 197 |
+
|
| 198 |
+
M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2010 (VOC2010) Results. http://www.pascalnetwork.org/challenges/VOC/voc2010/workshop/index.html, 2010.
|
| 199 |
+
|
| 200 |
+
Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Selfsupervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, July 2017.
|
| 201 |
+
|
| 202 |
+
Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020.
|
| 203 |
+
|
| 204 |
+
Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021.
|
| 205 |
+
|
| 206 |
+
Xiangwen Kong and Xiangyu Zhang. Understanding masked image modeling via learning occlusion invariant feature. arXiv preprint arXiv:2208.04164, 2022.
|
| 207 |
+
|
| 208 |
+
Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165–13176, 2021.
|
| 209 |
+
|
| 210 |
+
Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pp. 740–755. Springer, 2014.
|
| 211 |
+
|
| 212 |
+
Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022.
|
| 213 |
+
|
| 214 |
+
Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pp. 10012–10022, 2021.
|
| 215 |
+
|
| 216 |
+
Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pp. 8821–8831. PMLR, 2021.
|
| 217 |
+
|
| 218 |
+
Nikunj Saunshi, Jordan T. Ash, Surbhi Goel, Dipendra Misra, Cyril Zhang, Sanjeev Arora, Sham M. Kakade, and Akshay Krishnamurthy. Understanding contrastive learning requires incorporating inductive biases. CoRR, abs/2202.14037, 2022. URL https://arxiv.org/abs/2202. 14037.
|
| 219 |
+
|
| 220 |
+
Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022.
|
| 221 |
+
|
| 222 |
+
Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2020.
|
| 223 |
+
|
| 224 |
+
Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.
|
| 225 |
+
|
| 226 |
+
Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. arXiv preprint arXiv:2205.14141, 2022.
|
| 227 |
+
|
| 228 |
+
Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swin transformers. arXiv preprint arXiv:2105.04553, 2021a.
|
| 229 |
+
|
| 230 |
+
Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021b.
|
| 231 |
+
|
| 232 |
+
Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. CoRR, abs/2205.13543, 2022a. doi: 10.48550/arXiv.2205. 13543. URL https://doi.org/10.48550/arXiv.2205.13543.
|
| 233 |
+
|
| 234 |
+
Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Yixuan Wei, Qi Dai, and Han Hu. On data scaling in masked image modeling. CoRR, abs/2206.04664, 2022b. doi: 10.48550/arXiv.2206.04664. URL https://doi.org/10.48550/arXiv.2206.04664.
|
| 235 |
+
|
| 236 |
+
Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. arXiv preprint arXiv:2103.03230, 2021.
|
| 237 |
+
|
| 238 |
+
Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning? arXiv preprint arXiv:2206.05259, 2022.
|
| 239 |
+
|
| 240 |
+
Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127(3):302–321, 2019.
|
| 241 |
+
|
| 242 |
+
Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021.
|
| 243 |
+
|
| 244 |
+
# A APPENDIX
|
| 245 |
+
|
| 246 |
+
# A.1 MODEL DESCRIPTION
|
| 247 |
+
|
| 248 |
+
For all the models involved in the experiments including DeiT (Touvron et al., 2020), MoCo v3 (Chen et al., 2021), DINO (Caron et al., 2021), BEiT (Bao et al., 2021), MAE (He et al., 2021), CAE (Chen et al., 2022a), and iBOT (Zhou et al., 2021), we use their official code to implement the encoders. It is worth noticing that for DINO and iBOT, we choose the checkpoint of the teacher models as they have been reported to perform better than the student models in their papers (Caron et al., 2021; Zhou et al., 2021).
|
| 249 |
+
|
| 250 |
+
# A.2 DATASETS
|
| 251 |
+
|
| 252 |
+
ADE20K (Zhou et al., 2019) is one of the most challenging benchmarks, containing 150 fine-grained semantic concepts and a variety of scenes with 1,038 image-level labels. There are 20,210 images in the training set and 2,000 images in the validation set. We choose 59 out of total 150 semantic concepts that are concrete objects containing parts (Zhou et al., 2019), termed ADE20K-Object. We also select 209 part categories that emerge both in the training set and the validation set, called ADE20K-Part.
|
| 253 |
+
|
| 254 |
+
Pascal-Part (Chen et al., 2014) is a set of additional annotations for PASCAL VOC 2010 (Everingham et al., 2010), thereby holding the same statistics as those of PASCAL VOC 2010. It provides segmentation masks for each part of objects. Concretely, the dataset includes 20 object-level categories and 193 part-level categories. In our experiments, we remove 4 object categories that do not contain parts including boat, table, chair, and sofa.
|
| 255 |
+
|
| 256 |
+
LIP (Gong et al., 2017) is a large-scale benchmark for human parsing research, which includes 50,462 images with pixel-wise annotations on 19 semantic part labels. In detail, it includes 19,081 full-body images, 13,672 upper-body images, 403 lower-body images, 3,386 head-missed images, 2,778 back-view images and 21,028 images with occlusions. There are 30,462 images in the training set and 10,000 images in the validation set. The rest 10,000 images are served as the test set with missing labels for competition evaluation.
|
| 257 |
+
|
| 258 |
+
CUB-200-2011 (Wah et al., 2011) is a popular benchmark for fine-grained image classification, and also provides bounding box and part location annotations. It contains 11,788 images of 200 bird species and 15 part keypoint annotations per bird. In this work, we mainly leverage its part keypoint annotations. And only 4 part categories (right eye, right leg, left wing, and tail) are chosen to be considered in our experiments, to make sure that the selected keypoints are far enough away from each other and enough context information can be contained in the cropped patches. (We also tried using all keypoints and the conclusion is consistent.)
|
| 259 |
+
|
| 260 |
+
COCO (Caesar et al., 2018), as one of the most widely-used human pose estimation datasets, contains more than 200,000 images and 250,000 labeled person instances. Similar to CUB-200-2011 mentioned above, only 3 (nose, right wrist, and left ankle) of its 17 keypoint categories are considered in our experiments.
|
| 261 |
+
|
| 262 |
+
# A.3 DETAILED RESULTS FOR OBJECT-LEVEL AND PART-LEVEL SEGMENTATION
|
| 263 |
+
|
| 264 |
+
In this section, we provide detailed comparisons between object-level and part-level semantic segmentation in Table 5 and Table 6. Similar observations as in Figure 6 in the main paper are found: although the supervised DeiT is superior over self-supervised methods on ADE20K-Object and Pascal-Object except iBOT, it is generally inferior to self-supervised models on ADE20K-Part and Pascal-Part, demonstrating self-supervised methods can learn good part-aware representations. BEiT and MAE perform inferior, perhaps because the two methods do not have an explicit process to predict the encoded representations of masked patches, instead, directly reconstruct the targets.
|
| 265 |
+
|
| 266 |
+
# A.4 EXPERIMENT DETAILS
|
| 267 |
+
|
| 268 |
+
Part retrieval. In our part retrieval experiments, we directly use the pretrained encoders to extract features, without additional training process. For each method, we take the better one from the class token or the average embedding of all patch tokens as the extracted representation. With each patch as the query patch, we calculate the cosine similarity between its representation and all the other patches’ in the dataset and utilize the average precision (AP) as the retrieval metric. Finally, we average all the obtained AP scores (with all patches respectively taken as the query patch for retrieval) as the final retrieval score of the method.
|
| 269 |
+
|
| 270 |
+
Table 5: Linear semantic segmentation results on ADE20K-Object and ADE20K-Part.
|
| 271 |
+
|
| 272 |
+
<table><tr><td rowspan="3">Methods</td><td colspan="3">Object Seg on ADE20K-Object 59 Object Classes</td><td colspan="3">Part Seg on ADE20K-Part 209 Part Classes</td></tr><tr><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td colspan="7">Supervised Model:</td></tr><tr><td>DeiT</td><td>52.6</td><td>62.9</td><td>83.8</td><td>27.3</td><td>34.7</td><td>69.2</td></tr><tr><td colspan="7">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>50.2</td><td>60.4</td><td>83.6</td><td>27.1</td><td>34.7</td><td>70.1</td></tr><tr><td>DINO</td><td>50.8</td><td>60.8</td><td>83.9</td><td>28.9</td><td>36.8</td><td>70.3</td></tr><tr><td colspan="7">MaskedImageModeling (MIM):</td></tr><tr><td>BEiT</td><td>28.6</td><td>37.2</td><td>73.4</td><td>18.6</td><td>25.8</td><td>58.2</td></tr><tr><td>MAE</td><td>41.0</td><td>50.6</td><td>79.9</td><td>26.3</td><td>35.0</td><td>67.3</td></tr><tr><td>CAE</td><td>47.4</td><td>58.4</td><td>82.9</td><td>28.4</td><td>36.9</td><td>71.1</td></tr><tr><td colspan="7">Contrastive Learning+MIM:</td></tr><tr><td>iBOT</td><td>55.2</td><td>65.1</td><td>85.6</td><td>32.2</td><td>40.0</td><td>73.4</td></tr></table>
|
| 273 |
+
|
| 274 |
+
Table 6: Linear semantic segmentation results on Pascal-Object and Pascal-Part.
|
| 275 |
+
|
| 276 |
+
<table><tr><td rowspan="2">Methods</td><td colspan="3">Object Seg on Pascal-Object 16 Object Classes</td><td colspan="3">Part Seg on Pascal-Part 193 Part Classes</td></tr><tr><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td>Supervised Model: DeiT</td><td>92.2</td><td>95.3</td><td>96.8</td><td>27.4</td><td>36.2</td><td>65.8</td></tr><tr><td colspan="7">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>89.4</td><td>93.7</td><td>95.7</td><td>27.1</td><td>35.8</td><td>66.0</td></tr><tr><td>DINO</td><td>88.0</td><td>92.7</td><td>95.3</td><td>27.8</td><td>36.5</td><td>66.4</td></tr><tr><td colspan="3">Masked Image Modeling (MIM):</td><td></td><td></td><td></td><td></td></tr><tr><td>BEiT</td><td>56.4</td><td>69.0</td><td>76.8</td><td>14.8</td><td>21.4</td><td>47.0</td></tr><tr><td>MAE</td><td>76.1</td><td>84.6</td><td>89.5</td><td>24.3</td><td>32.9</td><td>61.5</td></tr><tr><td>CAE</td><td>83.3</td><td>89.7</td><td>93.2</td><td>27.8</td><td>37.0</td><td>66.3</td></tr><tr><td colspan="3">Contrastive Learning+ MIM:</td><td></td><td></td><td></td><td></td></tr><tr><td>iBOT</td><td>92.1</td><td>95.3</td><td>97.1</td><td>30.7</td><td>40.0</td><td>69.7</td></tr></table>
|
| 277 |
+
|
| 278 |
+
Apart from the part retrieval experiments shown in Table 3, the visualized patch retrieval results in Figures 2 and 4 are obtained based on ImageNet (Deng et al., 2009) validation set. Concretely, from each pre-processed $2 2 4 \times 2 2 4$ validation image in ImageNet, we uniformly crop 49 patches sized $5 6 \times 5 6$ using a stride of 28. With all the cropped patches from the validation set, we select one patch as a query and find top 24 patches with the highest cosine similarity with it.
|
| 279 |
+
|
| 280 |
+
Part classification. For linear probing, we learn a supervised linear classification layer on the extracted class token of the frozen encoders. While for attentive probing, following Chen et al. (2022a), a cross attention module and a batch normalization layer without affine transformation are additionally inserted between the encoder and the linear classifier. And a new learnable class token is taken as the query of the cross attention module, to replace the original class token extracted by the frozen encoder. We use SGD optimizer with a learning rate of 0.4 and 0.04 for linear probing and attentive probing, respectively. For both linear probing and attentive probing, the models are trained for 90 epochs. And the momentum of SGD is set to 0.9, the weight decay is set to 0, and the batch size is set to 1024.
|
| 281 |
+
|
| 282 |
+
|
| 283 |
+
Figure 7: Patch retrieval comparisons of encoded representations on cropped patches from ImageNet.
|
| 284 |
+
|
| 285 |
+
Segmentation. We use the same model structure that contains a parameter-fixed pretrained encoder (e.g., MAE and DeiT) and a simple learnable $1 \times 1$ convolutional layer for object-level and part-level segmentation tasks. Note that the learning rate $( 4 e - 4 )$ , training iterations $( 1 6 0 k )$ , and batch size (16) among all the experiments maintain the same during training for fair comparisons. For ADE20K, the input size is set to $5 1 2 \times 5 1 2$ following previous works (Bao et al., 2021; He et al., 2021; Chen et al., 2022a; Zhou et al., 2021). For Pascal-Part, we adopt $4 8 0 \times 4 8 0$ as image input resolution following Contributors (2020). As for LIP, we use the same input size $3 2 0 \times 3 2 0 )$ proposed in LIP (Gong et al., 2017).
|
| 286 |
+
|
| 287 |
+
# A.5 IMAGENET PATCH RETRIEVAL VISUALIZATION
|
| 288 |
+
|
| 289 |
+
We visualize more patch retrieval results of the encoded representations on the ImageNet validation set in Figures 7 and 8. It is observed that the retrieved patches of self-supervised methods are generally more about the semantics of the query part than that of DeiT. The results demonstrate that the encoded representations of DeiT focus more on object-level semantics, while the encoded representations of these self-supervised methods are more about part-level semantics. Among these methods, the retrieved patches of MAE have less semantic correlation but often share similar hues.
|
| 290 |
+
|
| 291 |
+
|
| 292 |
+
Figure 8: Patch retrieval comparisons of encoded representations on cropped patches from ImageNet.
|
parse/dev/3tYvDb4dwab/3tYvDb4dwab_content_list.json
ADDED
|
@@ -0,0 +1,1596 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "UNDERSTANDING SELF-SUPERVISED PRETRAINING WITH PART-AWARE REPRESENTATION LEARNING ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
176,
|
| 8 |
+
98,
|
| 9 |
+
823,
|
| 10 |
+
146
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Anonymous authors Paper under double-blind review ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
183,
|
| 19 |
+
170,
|
| 20 |
+
398,
|
| 21 |
+
198
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "ABSTRACT ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
454,
|
| 31 |
+
234,
|
| 32 |
+
544,
|
| 33 |
+
251
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "In this paper, we are interested in understanding self-supervised pretraining through studying the capability that self-supervised representation pretraining methods learn part-aware representations. The study is mainly motivated by that random views, used in contrastive learning, and random masked (visible) patches, used in masked image modeling, are often about object parts. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
233,
|
| 42 |
+
267,
|
| 43 |
+
766,
|
| 44 |
+
337
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "We explain that masked image modeling is a part-to-part task: the masked patches of the object are hallucinated from the visible patches, and that contrastive learning is a part-to-whole task: the projection layer hallucinates the whole object representation from the object part representation learned from the encoder. The explanation suggests that the self-supervised pretrained encoder is required to understand the object part. We empirically compare the off-the-shelf encoders pretrained with several representative methods on object-level recognition and part-level recognition. The results show that the fully-supervised model outperforms self-supervised models for object-level recognition, and most self-supervised contrastive learning and masked image modeling methods outperform the fully-supervised method for part-level recognition. It is observed that the combination of contrastive learning and masked image modeling further improves the performance. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
340,
|
| 54 |
+
766,
|
| 55 |
+
507
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 INTRODUCTION ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
176,
|
| 65 |
+
536,
|
| 66 |
+
334,
|
| 67 |
+
553
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Self-supervised representation pretraining has been attracting a lot of research efforts recently. The goal is to train an encoder that maps an image to a representation from visual contents without the necessity of human annotation, expecting that the encoder benefits the downstream tasks, e.g., segmentation and detection. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
570,
|
| 77 |
+
825,
|
| 78 |
+
625
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "There are two main frameworks: contrastive learning1 and masked image modeling. Contrastive learning aims to maximize the agreement of the embeddings of random augmented views from the same image. Masked image modeling partitions an image into masked patches and visible patches, and makes predictions for masked patches from visible patches. Figure 1 gives examples of random views for contrastive learning and masked and visible patches for masked image modeling. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
632,
|
| 88 |
+
825,
|
| 89 |
+
702
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "We observe that a random view and a set of masked (visible) patches usually contain a portion of an object. It is also reported in self-supervised learning methods, e.g., DINO (Caron et al., 2021) and iBOT (Zhou et al., 2021), that different attention heads in ViTs can attend to different semantic regions or parts of an object. In light of this, we attempt to understand self-supervised pretraining by studying the capability that the pretrained encoder learns part representations. ",
|
| 96 |
+
"bbox": [
|
| 97 |
+
174,
|
| 98 |
+
709,
|
| 99 |
+
825,
|
| 100 |
+
779
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 0
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "text",
|
| 106 |
+
"text": "We present a part-to-whole explanation for typical contrastive learning methods (e.g., SimCLR (Chen et al., 2020), MoCo (Chen et al., 2021), and BYOL (Grill et al., 2020)): the embedding of the whole object is hallucinated from the embedding of the part of the object contained in the random crop through a projection layer. In this way, embeddings of random crops from the same image naturally agrees with each other. Masked image modeling is a part-to-part process: the embeddings of the masked patches of the object (a part of the object), are hallucinated from the visible patches (the other part of the object). ",
|
| 107 |
+
"bbox": [
|
| 108 |
+
174,
|
| 109 |
+
786,
|
| 110 |
+
825,
|
| 111 |
+
883
|
| 112 |
+
],
|
| 113 |
+
"page_idx": 0
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"type": "image",
|
| 117 |
+
"img_path": "images/1dcba169f099594698e630dc560da9649431ed199ab88190ab27e2abe3d3ee16.jpg",
|
| 118 |
+
"image_caption": [
|
| 119 |
+
"Figure 1: (a) original image, (b-c) two random crops, and (d-e) masked and visible patches. "
|
| 120 |
+
],
|
| 121 |
+
"image_footnote": [],
|
| 122 |
+
"bbox": [
|
| 123 |
+
176,
|
| 124 |
+
113,
|
| 125 |
+
825,
|
| 126 |
+
218
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "image",
|
| 132 |
+
"img_path": "images/ce1cdaef1e5587eea23fae623a769166c8f264c809745fb83aa3642744d4b4d0.jpg",
|
| 133 |
+
"image_caption": [
|
| 134 |
+
"Figure 2: Top-24 patch retrieval results with three frozen encoders of DeiT, MoCo v3, and CAE, by taking the patch in the red box as the query. It can be seen that the retrieved results from CAE and MoCo v3 are about the object part (wing and dog mouth) and more precise than DeiT (about the whole object) implying that self-supervised pretraining methods, CAE and MoCo v3 are stronger at learning part-aware representations than the fully-supervised method DeiT. "
|
| 135 |
+
],
|
| 136 |
+
"image_footnote": [],
|
| 137 |
+
"bbox": [
|
| 138 |
+
179,
|
| 139 |
+
255,
|
| 140 |
+
815,
|
| 141 |
+
449
|
| 142 |
+
],
|
| 143 |
+
"page_idx": 1
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"type": "text",
|
| 147 |
+
"text": "We empirically compare the supervised model DeiT (Touvron et al., 2020) and typical self-supervised representation pretraining methods, including MoCo v3 (Chen et al., 2021), DINO (Caron et al., 2021), CAE (Chen et al., 2022a), MAE (He et al., 2021), BEiT (Bao et al., 2021), and iBOT (Zhou et al., 2021), on object-level recognition (image classification and object segmentation) and part-level recognition (patch retrieval, patch classification, and part segmentation). Figure 2 presents patch retrieval results using the encoders learned through CAE, MoCo v3, and DeiT, implying that the encoders pretrained by CAE and MoCo v3 are able to learn part-aware representations. ",
|
| 148 |
+
"bbox": [
|
| 149 |
+
173,
|
| 150 |
+
553,
|
| 151 |
+
825,
|
| 152 |
+
651
|
| 153 |
+
],
|
| 154 |
+
"page_idx": 1
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"type": "text",
|
| 158 |
+
"text": "Through extensive studies and comparisons, we make the following observations. 1) DeiT outperforms contrastive learning and MIM methods except iBOT in object-level recognition tasks, which may benefit from its explicit object-level supervision. 2) In contrast, self-supervised methods learn better part-aware representations than DeiT. For example, while DeiT is superior to DINO and CAE by $0 . 4 \\%$ and $2 . 3 \\%$ on ADE20K object segmentation, DINO and CAE outperform DeiT by $1 . 6 \\%$ and $1 . 1 \\%$ on ADE20K part segmentation, respectively. 3) In contrastive learning, the encoder can learn part-aware information, while the projected representation tends to be more about the whole object. The evidence could be found in part retrieval experiments on MoCo v3, DINO, and iBOT. 4) The MIM method CAE shows good potential in part-aware representation learning. Interestingly, the method combines contrastive learning and MIM is promising, e.g., iBOT learns better representations at both object and part levels. ",
|
| 159 |
+
"bbox": [
|
| 160 |
+
173,
|
| 161 |
+
657,
|
| 162 |
+
825,
|
| 163 |
+
810
|
| 164 |
+
],
|
| 165 |
+
"page_idx": 1
|
| 166 |
+
},
|
| 167 |
+
{
|
| 168 |
+
"type": "text",
|
| 169 |
+
"text": "To summarize, this paper presents the following contributions: ",
|
| 170 |
+
"bbox": [
|
| 171 |
+
174,
|
| 172 |
+
818,
|
| 173 |
+
583,
|
| 174 |
+
832
|
| 175 |
+
],
|
| 176 |
+
"page_idx": 1
|
| 177 |
+
},
|
| 178 |
+
{
|
| 179 |
+
"type": "text",
|
| 180 |
+
"text": "• We study the capability of learning part-aware representations as a way of understanding self-supervised representation pretraining. \n• We explain masked image modeling as a part-to-part task and contrastive learning as a partto-whole task, and speculate that self-supervised pretraining has the potential for learning part-aware representations. ",
|
| 181 |
+
"bbox": [
|
| 182 |
+
215,
|
| 183 |
+
843,
|
| 184 |
+
825,
|
| 185 |
+
924
|
| 186 |
+
],
|
| 187 |
+
"page_idx": 1
|
| 188 |
+
},
|
| 189 |
+
{
|
| 190 |
+
"type": "text",
|
| 191 |
+
"text": "• We empirically compare several pretrained models on object-level and part-level recognition tasks, showing interesting findings with supporting evidence of the capability of part-aware representation learning for self-supervised learning. ",
|
| 192 |
+
"bbox": [
|
| 193 |
+
220,
|
| 194 |
+
103,
|
| 195 |
+
823,
|
| 196 |
+
146
|
| 197 |
+
],
|
| 198 |
+
"page_idx": 2
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"type": "text",
|
| 202 |
+
"text": "2 RELATED WORK ",
|
| 203 |
+
"text_level": 1,
|
| 204 |
+
"bbox": [
|
| 205 |
+
176,
|
| 206 |
+
189,
|
| 207 |
+
343,
|
| 208 |
+
205
|
| 209 |
+
],
|
| 210 |
+
"page_idx": 2
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"type": "text",
|
| 214 |
+
"text": "Contrastive learning. Contrastive pretraining has been an intense academic field in the CNN era. In this work, we use it to refer to methods for comparing random views (Caron et al., 2020; Chen et al., 2020; Zbontar et al., 2021; Xie et al., 2021a; Chen et al., 2021; Caron et al., 2021), including some instance discrimination work such as (Grill et al., 2020; Chen & He, 2021; Bardes et al., 2021). As one of the representative works, SimCLR (Chen et al., 2020) learns representations through maximizing agreement between different views of the same image in the latent space. BYOL (Grill et al., 2020) uses two asymmetrical networks to bootstrap latent representation without negative samples involved during the interaction. As vision transformer (ViT) (Dosovitskiy et al., 2021) shows excellent performance via supervised learning, it is adopted subsequently in contrastive pertaining, and numerous outstanding works are proposed. For example, MoCo v3 (Chen et al., 2021) observes the hidden instability while training self-supervised ViT and solves it by using a fixed random patch projection. DINO (Caron et al., 2021) explores new properties derived from self-supervised ViT and accordingly designs a learning strategy interpreted as a form of self-distillation with no labels. ",
|
| 215 |
+
"bbox": [
|
| 216 |
+
173,
|
| 217 |
+
236,
|
| 218 |
+
825,
|
| 219 |
+
416
|
| 220 |
+
],
|
| 221 |
+
"page_idx": 2
|
| 222 |
+
},
|
| 223 |
+
{
|
| 224 |
+
"type": "text",
|
| 225 |
+
"text": "Masked image modeling (MIM). Masked image modeling is another self-supervised pretraining paradigm that attracts much attention recently. BEiT (Bao et al., 2021) follows masked language modeling in the natural language process (NLP) area and predicts tokens via mapping image patches by d-VAE (Ramesh et al., 2021). PeCo (Dong et al., 2021) boosts BEiT by taking into consideration more semantic information in visual tokens. MAE (He et al., 2021) learns rich hidden information by directly performing masked image reconstruction in RGB color space using ViT while SimMIM (Xie et al., 2021b) uses Swin-transformer (Liu et al., 2021). CAE (Chen et al., 2022a) adds a regressor between encoder and decoder, which is designed to align unmasked patches with masked ones, leading to a pure context encoder. Recently, a trend that combines MIM with siamese frameworks has surfaced and showed encouraging results including MST (Li et al., 2021), SplitMask (El-Nouby et al., 2021), iBOT (Zhou et al., 2021), dBOT (Liu et al., 2022), and SIM (Tao et al., 2022). ",
|
| 226 |
+
"bbox": [
|
| 227 |
+
174,
|
| 228 |
+
424,
|
| 229 |
+
825,
|
| 230 |
+
577
|
| 231 |
+
],
|
| 232 |
+
"page_idx": 2
|
| 233 |
+
},
|
| 234 |
+
{
|
| 235 |
+
"type": "text",
|
| 236 |
+
"text": "Understanding self-supervised contrastive pretraining. The studies on understanding contrastive pretraining (Saunshi et al., 2022; Chen et al., 2022b; Zhong et al., 2022; Wei et al., 2022) mainly focus on random augmentations (views), contrastive loss function and its variants under the assumption that: the augmentations of inputs from the same class have significant overlap in the representation space, but there is little overlap for inputs from different classes. Our work is complementary to these studies. Inspired by the observation that random views usually contain a portion of an object, and methods (Caron et al., 2021; Zhou et al., 2021) show that different attention heads in ViTs can attend to different semantic regions of an object, we investigate what the encoder and the projector do in typical self-supervised contrastive pretraining. We speculate that the pretraining task is a part-to-whole problem, predicting the representation of the whole object through the projector from the representation (obtained from the encoder) of the part of an object. We use empirical results to verify our analysis. ",
|
| 237 |
+
"bbox": [
|
| 238 |
+
174,
|
| 239 |
+
584,
|
| 240 |
+
825,
|
| 241 |
+
750
|
| 242 |
+
],
|
| 243 |
+
"page_idx": 2
|
| 244 |
+
},
|
| 245 |
+
{
|
| 246 |
+
"type": "text",
|
| 247 |
+
"text": "Understanding self-supervised masked image modeling. The comparison of attention in different layers between the pretrained models from MIM and the supervised approach is conducted: MIM pretraining brings locality to the trained model with sufficient diversity on the attention heads (Xie et al., 2022a). Consistent with the analysis in NLP, empirical studies are conducted in Xie et al. (2022b) to verify that MIM benefits from larger models, more data, and longer training. CAE (Chen et al., 2022a) gives the comparison between contrastive and MIM and shows MIM cares about all patches and thus achieves better results for fine-tuning. Cao et al. (2022) provides a mathematical understanding of MIM. Kong & Zhang (2022) points out that the learned occlusion invariant feature contributes to the success of MIM. In this work, we speculate that masked image modeling is a partto-part process: the embeddings of the masked part of the object are hallucinated from the visible part using the position information of the masked patches, leading to better part-aware representation than the supervised model DeiT (Touvron et al., 2020). ",
|
| 248 |
+
"bbox": [
|
| 249 |
+
173,
|
| 250 |
+
757,
|
| 251 |
+
825,
|
| 252 |
+
922
|
| 253 |
+
],
|
| 254 |
+
"page_idx": 2
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"type": "image",
|
| 258 |
+
"img_path": "images/dd5cea5e445400e9e9d24132dc8093c0b93046320f80e613aa8b0f5b1ba6d7bb.jpg",
|
| 259 |
+
"image_caption": [
|
| 260 |
+
"Figure 3: The pipeline of a typical contrastive learning approach. Two augmented views, red box and blue box, are generated from the original image. The augmented view in red is fed into the encoder and the projector, and then the predictor (which does not appear in earlier works like MoCo (Chen et al., 2021) and SimCLR (Chen et al., 2020)), and the view in blue is fed into the encoder and the projector. The two outputs are expected to be aligned. The gradient is stopped for the bottom stream. "
|
| 261 |
+
],
|
| 262 |
+
"image_footnote": [],
|
| 263 |
+
"bbox": [
|
| 264 |
+
222,
|
| 265 |
+
99,
|
| 266 |
+
771,
|
| 267 |
+
215
|
| 268 |
+
],
|
| 269 |
+
"page_idx": 3
|
| 270 |
+
},
|
| 271 |
+
{
|
| 272 |
+
"type": "text",
|
| 273 |
+
"text": "3 UNDERSTANDING CONTRASTIVE LEARNING AND MASK IMAGE MODELING ",
|
| 274 |
+
"text_level": 1,
|
| 275 |
+
"bbox": [
|
| 276 |
+
173,
|
| 277 |
+
314,
|
| 278 |
+
821,
|
| 279 |
+
332
|
| 280 |
+
],
|
| 281 |
+
"page_idx": 3
|
| 282 |
+
},
|
| 283 |
+
{
|
| 284 |
+
"type": "text",
|
| 285 |
+
"text": "3.1 CONTRASTIVE LEARNING ",
|
| 286 |
+
"text_level": 1,
|
| 287 |
+
"bbox": [
|
| 288 |
+
176,
|
| 289 |
+
345,
|
| 290 |
+
397,
|
| 291 |
+
359
|
| 292 |
+
],
|
| 293 |
+
"page_idx": 3
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"type": "text",
|
| 297 |
+
"text": "Contrastive learning aims to learn the encoder through maximizing the agreement between differently augmented views of the same image in the representation space. An example pipeline is depicted in Figure 3. Given an image I, the augmentations, e.g., random cropping, random color distortion, and random Gaussian blur, are applied to generate a set of $N$ augmented views, $\\{ \\mathsf { V } _ { 1 } , \\mathsf { V } _ { 2 } , \\cdots , \\mathsf { V } _ { N } \\}$ . An augmented view $\\mathsf { V } _ { n }$ is fed into an encoder Encoder, generating the encoded representation ${ \\bf x } _ { n }$ , and followed by a projector, generating the projection $\\mathbf { z } _ { n }$ . The basic goal is to maximize the agreement between the projections $\\{ \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } , \\cdots , \\mathbf { z } _ { N } \\}$ , i.e., minimize the loss ",
|
| 298 |
+
"bbox": [
|
| 299 |
+
173,
|
| 300 |
+
371,
|
| 301 |
+
825,
|
| 302 |
+
469
|
| 303 |
+
],
|
| 304 |
+
"page_idx": 3
|
| 305 |
+
},
|
| 306 |
+
{
|
| 307 |
+
"type": "equation",
|
| 308 |
+
"img_path": "images/b89908eb0ce99d0c54d465a1150513f70fc02b39de4360232dea9af7ee8a1d28.jpg",
|
| 309 |
+
"text": "$$\n{ \\mathcal { L } } _ { \\mathrm { C P T } } = \\sum _ { i = 1 } ^ { N } \\sum _ { j = 1 } ^ { N } { \\ell } ( \\mathrm { P r o j e c t o r ( E n c o d e r ( V } _ { i } ) ) , \\mathrm { P r o j e c t o r ( E n c o d e r ( V } _ { j } ) ) ) .\n$$",
|
| 310 |
+
"text_format": "latex",
|
| 311 |
+
"bbox": [
|
| 312 |
+
233,
|
| 313 |
+
476,
|
| 314 |
+
764,
|
| 315 |
+
506
|
| 316 |
+
],
|
| 317 |
+
"page_idx": 3
|
| 318 |
+
},
|
| 319 |
+
{
|
| 320 |
+
"type": "text",
|
| 321 |
+
"text": "In the formulation with a contrastive loss, the agreement between the projections of random augmentations from different images is minimized. ",
|
| 322 |
+
"bbox": [
|
| 323 |
+
173,
|
| 324 |
+
511,
|
| 325 |
+
825,
|
| 326 |
+
539
|
| 327 |
+
],
|
| 328 |
+
"page_idx": 3
|
| 329 |
+
},
|
| 330 |
+
{
|
| 331 |
+
"type": "text",
|
| 332 |
+
"text": "Part-to-whole prediction explanation. Let us consider two crops randomly sampled from the original image (see the examples given in Figure 1(b-c)). The encoded representation of the first crop is expected to describe a part of the object dog; the encoded representation of the second crop is expected to describe another part of the object $\\mathrm { d o g ^ { 2 } }$ . The two representations are related but different. Contrastive learning methods project the two encoded representations into two projected representations that are expected to agree. We hypothesize that the projection process maps the encoded part representation to the representation of the whole object3. Through this way, the projected representations will agree to different views from the same image. It is assumed that the part-to-whole projection is more reliable if the encoded representation is semantically richer and is able to describe the part information. The part-to-whole process suggests that the encoder pretrained by contrastive learning methods is potentially capable of learning part-aware representations. ",
|
| 333 |
+
"bbox": [
|
| 334 |
+
173,
|
| 335 |
+
545,
|
| 336 |
+
825,
|
| 337 |
+
699
|
| 338 |
+
],
|
| 339 |
+
"page_idx": 3
|
| 340 |
+
},
|
| 341 |
+
{
|
| 342 |
+
"type": "text",
|
| 343 |
+
"text": "Figure 4 provides patch search results of a representative contrastive learning method MoCo v3 (Chen et al., 2021) based on the encoded representations before and after the projections. One can see that the results through the encoded representations are mainly about the local part, and the results through the projections tend to include the other parts of the same object. In other words, the projections tend to be about the whole object. Similar observations are also shown in Chen et al. (2022b). The search results verify the part-to-whole hypothesis. ",
|
| 344 |
+
"bbox": [
|
| 345 |
+
173,
|
| 346 |
+
705,
|
| 347 |
+
825,
|
| 348 |
+
790
|
| 349 |
+
],
|
| 350 |
+
"page_idx": 3
|
| 351 |
+
},
|
| 352 |
+
{
|
| 353 |
+
"type": "text",
|
| 354 |
+
"text": "3.2 MASKED IMAGE MODELING ",
|
| 355 |
+
"text_level": 1,
|
| 356 |
+
"bbox": [
|
| 357 |
+
176,
|
| 358 |
+
806,
|
| 359 |
+
411,
|
| 360 |
+
820
|
| 361 |
+
],
|
| 362 |
+
"page_idx": 3
|
| 363 |
+
},
|
| 364 |
+
{
|
| 365 |
+
"type": "text",
|
| 366 |
+
"text": "Mask image modeling is the task of predicting some parts of an image from the remaining parts. An augmented view of an image is partitioned into patches, $\\mathcal { R } = \\{ \\mathsf { R } _ { 1 } , \\mathsf { R } _ { 2 } , \\ldots , \\mathsf { R } _ { M } \\}$ . The task is to predict a subset of patches $\\mathcal { R } _ { m }$ , named masked patches, from the remaining patches $\\mathcal { R } _ { v }$ , named visible patches. Considering contrastive learning that explicitly compares representations of random views, we take context autoencoder (CAE) (Chen et al., 2022a) as an example that explicitly predicts the encoded representations of the masked patches from the encoded representations of the visible patches4. ",
|
| 367 |
+
"bbox": [
|
| 368 |
+
174,
|
| 369 |
+
832,
|
| 370 |
+
823,
|
| 371 |
+
861
|
| 372 |
+
],
|
| 373 |
+
"page_idx": 3
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"type": "image",
|
| 377 |
+
"img_path": "images/d3390c81acfb9c93a356161138f9c0e7fc0270197613fca8c06924a2eac48b51.jpg",
|
| 378 |
+
"image_caption": [
|
| 379 |
+
"Figure 4: Illustration of patch search results using encoded representations and projections (pretrained with MoCo v3 as). Left: patch search results with encoded representations. Right: patch search results with projections. In each result, the small patch encircled by the red box is taken as the query. It can be seen that for encoded representations, the returned patches are about the same part, and for projections, the result patches are about the same object, verifying the part-to-whole hypothesis. "
|
| 380 |
+
],
|
| 381 |
+
"image_footnote": [],
|
| 382 |
+
"bbox": [
|
| 383 |
+
207,
|
| 384 |
+
99,
|
| 385 |
+
792,
|
| 386 |
+
342
|
| 387 |
+
],
|
| 388 |
+
"page_idx": 4
|
| 389 |
+
},
|
| 390 |
+
{
|
| 391 |
+
"type": "text",
|
| 392 |
+
"text": "",
|
| 393 |
+
"bbox": [
|
| 394 |
+
173,
|
| 395 |
+
443,
|
| 396 |
+
825,
|
| 397 |
+
512
|
| 398 |
+
],
|
| 399 |
+
"page_idx": 4
|
| 400 |
+
},
|
| 401 |
+
{
|
| 402 |
+
"type": "text",
|
| 403 |
+
"text": "One goal of CAE (illustrated in Figure 5), which we call masked representation modeling (MRM), is to maximize the agreement between the predictions of the representations of masked patches (through a regressor) and the representation of masked patches computed from the encoder by minimizing the loss ",
|
| 404 |
+
"bbox": [
|
| 405 |
+
173,
|
| 406 |
+
520,
|
| 407 |
+
825,
|
| 408 |
+
575
|
| 409 |
+
],
|
| 410 |
+
"page_idx": 4
|
| 411 |
+
},
|
| 412 |
+
{
|
| 413 |
+
"type": "equation",
|
| 414 |
+
"img_path": "images/03f93947969323b5c0de0e6cb078e7add1d87aba73c00ddd9d0fed7580bfb188.jpg",
|
| 415 |
+
"text": "$$\n\\ell _ { \\mathrm { M R M } } ( \\mathrm { R e g r e s s o r } ( \\operatorname { E n c o d e r } ( \\mathcal { P } _ { v } ) ) , \\operatorname { E n c o d e r } ( \\mathcal { P } _ { m } ) ) .\n$$",
|
| 416 |
+
"text_format": "latex",
|
| 417 |
+
"bbox": [
|
| 418 |
+
331,
|
| 419 |
+
582,
|
| 420 |
+
666,
|
| 421 |
+
598
|
| 422 |
+
],
|
| 423 |
+
"page_idx": 4
|
| 424 |
+
},
|
| 425 |
+
{
|
| 426 |
+
"type": "text",
|
| 427 |
+
"text": "Here, we do not include the positional embeddings of masked and visible patches for clarity. It is noted that MRM differs from contrastive learning: MRM does not compare multiple random views, but compares the regressed representations for masked patches and the encoded representations of masked patches. In addition, there is another loss for target prediction (reconstruction) for the masked patches, which is commonly used in masked image modeling (MIM) methods: ",
|
| 428 |
+
"bbox": [
|
| 429 |
+
174,
|
| 430 |
+
603,
|
| 431 |
+
825,
|
| 432 |
+
674
|
| 433 |
+
],
|
| 434 |
+
"page_idx": 4
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"type": "equation",
|
| 438 |
+
"img_path": "images/4df1cab25511bdd1636224ced86fcd6dcb3a5fc7454730bec53a32700677aec3.jpg",
|
| 439 |
+
"text": "$$\n\\ell _ { \\mathrm { M I M } } \\big ( \\mathrm { D e c o d e r } \\big ( \\mathrm { R e g r e s s o r } \\big ( \\mathrm { E n c o d e r } ( \\mathcal { P } _ { v } ) \\big ) \\big ) , \\mathrm { T a r g e t } ( \\mathcal { P } _ { m } ) \\big ) ,\n$$",
|
| 440 |
+
"text_format": "latex",
|
| 441 |
+
"bbox": [
|
| 442 |
+
303,
|
| 443 |
+
679,
|
| 444 |
+
692,
|
| 445 |
+
695
|
| 446 |
+
],
|
| 447 |
+
"page_idx": 4
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"type": "text",
|
| 451 |
+
"text": "where ${ \\mathrm { T a r g e t } } ( { \\mathcal { P } } _ { m } )$ is a function to map the masked patches to the targets, e.g., d-VAE (Ramesh et al., 2021) token used in CAE and BeiT (Bao et al., 2021), or normalized RGB values used in MAE (He et al., 2021). ",
|
| 452 |
+
"bbox": [
|
| 453 |
+
173,
|
| 454 |
+
700,
|
| 455 |
+
825,
|
| 456 |
+
742
|
| 457 |
+
],
|
| 458 |
+
"page_idx": 4
|
| 459 |
+
},
|
| 460 |
+
{
|
| 461 |
+
"type": "text",
|
| 462 |
+
"text": "Part-to-part prediction explanation. The masked image modeling approaches, including CAE, MAE, and BEiT, make use of the positions of masked patches for making predictions for masked patches from visible patches. The visible patches and masked patches often contain different parts of an object. In other words, MIM aims to predict the masked part of an object from the visible part. We name this a part-to-part process. There are two part-to-part tasks: one is to reconstruct the part targets from the visible part representations (MAE and CAE) or from the visible part raw pixels (BEiT), and the other one is to regress the masked part representations (CAE). The part-to-part process suggests that the encoder pretrained by MIM methods is potentially capable of learning part-aware representations. Figure 2 illustrates the capability with the patch retrieval results. ",
|
| 463 |
+
"bbox": [
|
| 464 |
+
173,
|
| 465 |
+
750,
|
| 466 |
+
825,
|
| 467 |
+
876
|
| 468 |
+
],
|
| 469 |
+
"page_idx": 4
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"type": "image",
|
| 473 |
+
"img_path": "images/c8e801f685cfaa84f3f5bf513d1799a5813f082308260a7621445ef9cf9e8e9e.jpg",
|
| 474 |
+
"image_caption": [
|
| 475 |
+
"Figure 5: The pipeline of an MIM approach, context autoencoder (CAE). An augmented view (in blue) of the image is partitioned into visible and masked patches. The CAE approach feeds visible patches into the encoder and extracts their representations $\\mathbf { Z } _ { v }$ and then completes the pretext task by predicting the representations $\\mathbf { Z } _ { m }$ of the masked patches from the visible patches in the encoded representation space with latent contextual regressor and alignment constraint, and mapping predicted representations $\\mathbf { Z } _ { m }$ of masked patches to the targets. The pretrained encoder in (a) is applied to downstream tasks by simply replacing the pretext task part (b, c) with the downstream task completion part. "
|
| 476 |
+
],
|
| 477 |
+
"image_footnote": [],
|
| 478 |
+
"bbox": [
|
| 479 |
+
178,
|
| 480 |
+
109,
|
| 481 |
+
818,
|
| 482 |
+
246
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 5
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "table",
|
| 488 |
+
"img_path": "images/229fae4bc3ce2aca7902b27bbeeb2d77d39549d2740e2bd87831a595eaf6ad99.jpg",
|
| 489 |
+
"table_caption": [
|
| 490 |
+
"Table 1: Top-1 accuracy with linear probing, and attentive probing (Chen et al., 2022a), on the ImageNet classification benchmark (Deng et al., 2009). "
|
| 491 |
+
],
|
| 492 |
+
"table_footnote": [],
|
| 493 |
+
"table_body": "<table><tr><td>Method</td><td>Linear</td><td>Attentive</td></tr><tr><td colspan=\"3\">Supervised Model:</td></tr><tr><td>DeiT</td><td>81.8</td><td>81.8</td></tr><tr><td colspan=\"3\">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>76.2</td><td>77.0</td></tr><tr><td>DINO</td><td>77.3</td><td>77.8</td></tr><tr><td colspan=\"3\">Masked Image Modeling (MIM):</td></tr><tr><td>BEiT</td><td>41.8</td><td>51.9</td></tr><tr><td>MAE</td><td>67.8</td><td>74.2</td></tr><tr><td>CAE</td><td>70.4</td><td>77.1</td></tr><tr><td colspan=\"3\">Contrastive Learning+MIM:</td></tr><tr><td>iBOT</td><td>79.5</td><td>79.8</td></tr></table>",
|
| 494 |
+
"bbox": [
|
| 495 |
+
178,
|
| 496 |
+
465,
|
| 497 |
+
486,
|
| 498 |
+
655
|
| 499 |
+
],
|
| 500 |
+
"page_idx": 5
|
| 501 |
+
},
|
| 502 |
+
{
|
| 503 |
+
"type": "table",
|
| 504 |
+
"img_path": "images/c35b5150147d0a2fb63f214b6b312cb8dafd196aff6841a64d15d325ac7f2de6.jpg",
|
| 505 |
+
"table_caption": [
|
| 506 |
+
"Table 2: Linear evaluation of ADE20K (Zhou et al., 2019) object-level semantic segmentation (150 classes) using $4 \\times$ upsampling and a single $1 \\times 1$ convolutional layer on frozen backbones. "
|
| 507 |
+
],
|
| 508 |
+
"table_footnote": [],
|
| 509 |
+
"table_body": "<table><tr><td>Method</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td>Supervised Model:</td><td></td><td></td><td></td></tr><tr><td>DeiT</td><td>34.9</td><td>44.2</td><td>75.4</td></tr><tr><td>Contrastive Learning:</td><td></td><td></td><td></td></tr><tr><td>MoCo v3</td><td>34.7</td><td>43.9</td><td>75.9</td></tr><tr><td>DINO</td><td>34.5</td><td>43.5</td><td>76.1</td></tr><tr><td colspan=\"4\">Masked Image Modeling (MIM):</td></tr><tr><td>BEiT</td><td>17.8</td><td>23.7</td><td>64.9</td></tr><tr><td>MAE</td><td>27.1</td><td>34.8</td><td>71.6</td></tr><tr><td>CAE</td><td>32.6</td><td>42.2</td><td>75.2</td></tr><tr><td colspan=\"4\">Contrastive Learning +MIM:</td></tr><tr><td>iBOT</td><td>38.3</td><td>47.4</td><td>78.1</td></tr></table>",
|
| 510 |
+
"bbox": [
|
| 511 |
+
506,
|
| 512 |
+
467,
|
| 513 |
+
821,
|
| 514 |
+
655
|
| 515 |
+
],
|
| 516 |
+
"page_idx": 5
|
| 517 |
+
},
|
| 518 |
+
{
|
| 519 |
+
"type": "text",
|
| 520 |
+
"text": "4 EXPERIMENTS ",
|
| 521 |
+
"text_level": 1,
|
| 522 |
+
"bbox": [
|
| 523 |
+
174,
|
| 524 |
+
681,
|
| 525 |
+
326,
|
| 526 |
+
696
|
| 527 |
+
],
|
| 528 |
+
"page_idx": 5
|
| 529 |
+
},
|
| 530 |
+
{
|
| 531 |
+
"type": "text",
|
| 532 |
+
"text": "We study seven representative methods with the same ViT-B encoder, including a supervised method DeiT (Touvron et al., 2020); contrastive learning methods MoCo v3 (Chen et al., 2021), DINO (Caron et al., 2021); masked image modeling (MIM) methods BEiT (Bao et al., 2021), MAE (He et al., 2021), and CAE (Chen et al., 2022a); and iBOT (Zhou et al., 2021) that combines contrastive learning and MIM. We take the training epochs specified in each work to ensure that all compared models are properly trained: 300 for DeiT, 300 $( 6 0 \\dot { 0 } ^ { 5 } )$ for MoCo v3, 400 $( 1 6 0 0 ^ { 5 } )$ ) for DINO and iBOT, 800 for BEiT, and 1600 for MAE and CAE. Frozen encoders are used in all experiments to understand what these different representation pretraining methods learn. More details can be found in Appendix A.1. ",
|
| 533 |
+
"bbox": [
|
| 534 |
+
173,
|
| 535 |
+
712,
|
| 536 |
+
825,
|
| 537 |
+
824
|
| 538 |
+
],
|
| 539 |
+
"page_idx": 5
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"type": "text",
|
| 543 |
+
"text": "4.1 OBJECT-LEVEL RECOGNITION ",
|
| 544 |
+
"text_level": 1,
|
| 545 |
+
"bbox": [
|
| 546 |
+
176,
|
| 547 |
+
843,
|
| 548 |
+
426,
|
| 549 |
+
857
|
| 550 |
+
],
|
| 551 |
+
"page_idx": 5
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"type": "text",
|
| 555 |
+
"text": "We benchmark two widely-studied object-level recognition, i.e., image classification and semantic segmentation to show the capability that the pretrained encoder learns object-level representations. ",
|
| 556 |
+
"bbox": [
|
| 557 |
+
178,
|
| 558 |
+
869,
|
| 559 |
+
823,
|
| 560 |
+
897
|
| 561 |
+
],
|
| 562 |
+
"page_idx": 5
|
| 563 |
+
},
|
| 564 |
+
{
|
| 565 |
+
"type": "table",
|
| 566 |
+
"img_path": "images/5115795ab160f25d9cbfb0223085807741606684a325f43762ba3e24a2158346.jpg",
|
| 567 |
+
"table_caption": [
|
| 568 |
+
"Table 3: Part retrieval (AP, $\\%$ ) and classification (accuracy, $\\%$ ) results on the cropped part patches of CUB-200-2011 and COCO. The “Encoded\" and “Projected\" refer to the encoded and projected representations. “Linear\" and “Attentive\" columns denote the linear probing and attentive probing accuracy, respectively. "
|
| 569 |
+
],
|
| 570 |
+
"table_footnote": [],
|
| 571 |
+
"table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"4\">PartRetrieval</td><td colspan=\"4\">Part Classification</td></tr><tr><td colspan=\"2\">CUB-200-2011</td><td colspan=\"2\">COCo</td><td colspan=\"2\">CUB-200-2011</td><td colspan=\"2\">COCO</td></tr><tr><td>Encoded</td><td>Projected</td><td>Encoded</td><td>Projected</td><td>Linear</td><td>Attentive</td><td>Linear</td><td>Attentive</td></tr><tr><td colspan=\"9\">Supervised Model:</td></tr><tr><td>DeiT</td><td>35.0</td><td></td><td>44.1</td><td></td><td>90.9</td><td>92.9</td><td>88.5</td><td>91.4</td></tr><tr><td colspan=\"9\">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>50.8</td><td>28.4</td><td>52.3</td><td>36.8</td><td>93.8</td><td>96.0</td><td>92.4</td><td>95.3</td></tr><tr><td>DINO</td><td>48.9</td><td>31.7</td><td>51.8</td><td>41.2</td><td>93.2</td><td>95.2</td><td>91.7</td><td>94.5</td></tr><tr><td colspan=\"9\">Masked Image Modeling (MIM):</td></tr><tr><td>BEiT</td><td>27.9</td><td>1</td><td>35.3</td><td>1</td><td>55.4</td><td>86.5</td><td>69.3</td><td>86.5</td></tr><tr><td>MAE</td><td>28.5</td><td>1</td><td>37.1</td><td></td><td>86.9</td><td>92.8</td><td>88.0</td><td>93.9</td></tr><tr><td>CAE</td><td>58.0</td><td>1</td><td>57.0</td><td>1</td><td>89.5</td><td>95.8</td><td>91.1</td><td>95.5</td></tr><tr><td colspan=\"9\">Contrastive Learning +MIM:</td></tr><tr><td>iBOT</td><td>49.3</td><td>31.2</td><td>59.2</td><td>41.5</td><td>93.8</td><td>95.8</td><td>92.1</td><td>95.1</td></tr></table>",
|
| 572 |
+
"bbox": [
|
| 573 |
+
194,
|
| 574 |
+
167,
|
| 575 |
+
803,
|
| 576 |
+
371
|
| 577 |
+
],
|
| 578 |
+
"page_idx": 6
|
| 579 |
+
},
|
| 580 |
+
{
|
| 581 |
+
"type": "text",
|
| 582 |
+
"text": "Image classification. We report the linear probing, and attentive probing results of the selected models on ImageNet (Deng et al., 2009). For attentive probing, we follow the protocol in CAE (Chen et al., 2022a) that append a cross-attention layer together with a batch normalization layer and a linear classifier. ",
|
| 583 |
+
"bbox": [
|
| 584 |
+
173,
|
| 585 |
+
400,
|
| 586 |
+
825,
|
| 587 |
+
455
|
| 588 |
+
],
|
| 589 |
+
"page_idx": 6
|
| 590 |
+
},
|
| 591 |
+
{
|
| 592 |
+
"type": "text",
|
| 593 |
+
"text": "We have the following observations from Table 1. 1) The supervised model, DeiT performs better than self-supervised models at object-level recognition. 2) The models that leverage contrastive learning, i.e., MoCo, DINO, and iBOT, show superior linear probing performance than MIM-based models, demonstrating they contain more object-aware high-level semantics. 3) MIM-based models, e.g., CAE, show inferior results in linear probing while competitive results with contrastive-based methods in attentive probing. The reason might be that MIM is capable of attending to all the regions, including non-object regions in an image, thus needs a spatial feature selection step to attend to the object part, which is pointed out in Chen et al. (2022a). BEiT and MAE perform inferior, implying that the two methods are less capable of learning semantics. ",
|
| 594 |
+
"bbox": [
|
| 595 |
+
174,
|
| 596 |
+
462,
|
| 597 |
+
825,
|
| 598 |
+
588
|
| 599 |
+
],
|
| 600 |
+
"page_idx": 6
|
| 601 |
+
},
|
| 602 |
+
{
|
| 603 |
+
"type": "text",
|
| 604 |
+
"text": "Object-level semantic segmentation. We perform linear evaluation on ADE20K (Zhou et al., 2019) to show the object-level semantic capabilities of the pretrained models. A $4 \\times$ bilinear interpolation and a single $1 \\times 1$ convolutional layer for pixel labeling are attached to the frozen encoder. ",
|
| 605 |
+
"bbox": [
|
| 606 |
+
174,
|
| 607 |
+
594,
|
| 608 |
+
825,
|
| 609 |
+
637
|
| 610 |
+
],
|
| 611 |
+
"page_idx": 6
|
| 612 |
+
},
|
| 613 |
+
{
|
| 614 |
+
"type": "text",
|
| 615 |
+
"text": "We can see from Table 2 that the supervised model DeiT outperforms all self-supervised models except iBOT, including contrastive learning and MIM methods on ADE20K object-level segmentation. This implies that in general the self-supervised models are not strong at object-level understanding, which is consistent with the observations for image classification. iBOT (Zhou et al., 2021), as a combination of contrastive learning and MIM, shows surprisingly better performance than the supervised model DeiT on ADE20K, implying the power of combining contrastive learning and masked image modeling for downstream tasks. ",
|
| 616 |
+
"bbox": [
|
| 617 |
+
174,
|
| 618 |
+
643,
|
| 619 |
+
825,
|
| 620 |
+
741
|
| 621 |
+
],
|
| 622 |
+
"page_idx": 6
|
| 623 |
+
},
|
| 624 |
+
{
|
| 625 |
+
"type": "text",
|
| 626 |
+
"text": "4.2 PART-LEVEL RECOGNITION ",
|
| 627 |
+
"text_level": 1,
|
| 628 |
+
"bbox": [
|
| 629 |
+
176,
|
| 630 |
+
763,
|
| 631 |
+
406,
|
| 632 |
+
777
|
| 633 |
+
],
|
| 634 |
+
"page_idx": 6
|
| 635 |
+
},
|
| 636 |
+
{
|
| 637 |
+
"type": "text",
|
| 638 |
+
"text": "Self-supervised methods like iBOT (Zhou et al., 2021) and DINO (Caron et al., 2021) qualitatively show that different attention heads in ViTs can attend to different semantic regions of an object. We conduct the quantitative evaluation for part-aware representation obtained by pretrained models that is not well explored before, through three part-level recognition tasks, part retrieval, part classification, and part segmentation. ",
|
| 639 |
+
"bbox": [
|
| 640 |
+
174,
|
| 641 |
+
790,
|
| 642 |
+
825,
|
| 643 |
+
861
|
| 644 |
+
],
|
| 645 |
+
"page_idx": 6
|
| 646 |
+
},
|
| 647 |
+
{
|
| 648 |
+
"type": "text",
|
| 649 |
+
"text": "Part retrieval. We conduct part retrieval experiments on two datasets, CUB-200-2011 (Wah et al., 2011) and COCO (Lin et al., 2014). We build the part patch databases by cropping the patches centered at the keypoint. We consider four and three keypoints from the two datasets, respectively. For each keypoint, we find the minimum L2 distance $( d )$ from the distances between it and all the other keypoints in the same image, then crop a $d \\times d$ patch centered at this keypoint and resize it to $2 2 4 \\times 2 2 4$ . We use the cosine distance as the patch distance and evaluate the retrieval performance using average precision (AP) as the retrieval metric. ",
|
| 650 |
+
"bbox": [
|
| 651 |
+
174,
|
| 652 |
+
867,
|
| 653 |
+
825,
|
| 654 |
+
924
|
| 655 |
+
],
|
| 656 |
+
"page_idx": 6
|
| 657 |
+
},
|
| 658 |
+
{
|
| 659 |
+
"type": "table",
|
| 660 |
+
"img_path": "images/7e6f2213a6770d670c42b0e69d8ff145da2edd21e6359304c407ca8b8d1befe4.jpg",
|
| 661 |
+
"table_caption": [
|
| 662 |
+
"Table 4: Part-level linear semantic segmentation results on ADE20K-Part, Pascal-Part, and LIP datasets. "
|
| 663 |
+
],
|
| 664 |
+
"table_footnote": [],
|
| 665 |
+
"table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"3\">ADE20K-Part 209 Part Classes</td><td colspan=\"3\">Pascal-Part 193 Part Classes</td><td colspan=\"3\">LIP 19 Part Classes</td></tr><tr><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td colspan=\"10\">Supervised Model:</td></tr><tr><td>DeiT</td><td>27.3</td><td>34.7</td><td>69.2</td><td>27.4</td><td>36.1</td><td>65.8</td><td>41.4</td><td>52.6</td><td>73.5</td></tr><tr><td colspan=\"10\">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>27.1</td><td>34.7</td><td>70.1</td><td>27.1</td><td>35.8</td><td>66.0</td><td>41.9</td><td>53.0</td><td>74.5</td></tr><tr><td>DINO</td><td>28.9</td><td>36.8</td><td>70.3</td><td>27.8</td><td>36.5</td><td>66.4</td><td>41.0</td><td>51.9</td><td>74.0</td></tr><tr><td colspan=\"10\">Masked ImageModeling ( (MIM):</td></tr><tr><td>BEiT</td><td>18.6</td><td>25.8</td><td>58.2</td><td>14.8</td><td>21.4</td><td>47.0</td><td>27.2</td><td>36.5</td><td>60.1</td></tr><tr><td>MAE</td><td>26.3</td><td>35.0</td><td>67.3</td><td>24.3</td><td>32.9</td><td>61.5</td><td>38.2</td><td>48.7</td><td>71.3</td></tr><tr><td>CAE</td><td>28.4</td><td>36.9</td><td>71.1</td><td>27.8</td><td>37.0</td><td>66.3</td><td>43.7</td><td>55.1</td><td>75.9</td></tr><tr><td colspan=\"10\">Contrastive Learning + MIM:</td></tr><tr><td>iBOT</td><td>32.2</td><td>40.0</td><td>73.4</td><td>30.7</td><td>40.0</td><td>69.7</td><td>44.6</td><td>55.7</td><td>76.6</td></tr></table>",
|
| 666 |
+
"bbox": [
|
| 667 |
+
197,
|
| 668 |
+
140,
|
| 669 |
+
802,
|
| 670 |
+
361
|
| 671 |
+
],
|
| 672 |
+
"page_idx": 7
|
| 673 |
+
},
|
| 674 |
+
{
|
| 675 |
+
"type": "text",
|
| 676 |
+
"text": "",
|
| 677 |
+
"bbox": [
|
| 678 |
+
174,
|
| 679 |
+
387,
|
| 680 |
+
823,
|
| 681 |
+
429
|
| 682 |
+
],
|
| 683 |
+
"page_idx": 7
|
| 684 |
+
},
|
| 685 |
+
{
|
| 686 |
+
"type": "text",
|
| 687 |
+
"text": "The results are provided in Table 3. We have the following observations. 1) Self-supervised models except BEiT and MAE outperform the supervised model DeiT, indicating the capability that contrastive learning and CAE learn part-aware representations. BEiT and MAE perform inferior, consistent to the observations in ImageNet classification in Table 1. 2) iBOT performs the best, and the reason might be that the capability of learning part-aware representations is boosted by making use of both contrastive learning and masked image modeling. ",
|
| 688 |
+
"bbox": [
|
| 689 |
+
174,
|
| 690 |
+
435,
|
| 691 |
+
825,
|
| 692 |
+
520
|
| 693 |
+
],
|
| 694 |
+
"page_idx": 7
|
| 695 |
+
},
|
| 696 |
+
{
|
| 697 |
+
"type": "text",
|
| 698 |
+
"text": "We also report the part retrieval performance of the projected representations of contrastive learning methods in Table 3. The performance is much lower than the encoded representations. This provides an extra evidence for the part-to-whole hypothesis of contrastive learning: the projected representations are more about the whole object. ",
|
| 699 |
+
"bbox": [
|
| 700 |
+
174,
|
| 701 |
+
526,
|
| 702 |
+
825,
|
| 703 |
+
582
|
| 704 |
+
],
|
| 705 |
+
"page_idx": 7
|
| 706 |
+
},
|
| 707 |
+
{
|
| 708 |
+
"type": "text",
|
| 709 |
+
"text": "Part classification. We further conduct part classification experiments on the datasets used for part retrieval. We consider two kinds of extra learnable layers, linear probing and attentive probing, for classification. The results in Table 3 show that: 1) While DeiT performs the best in the image classification task (see Table 1), for part classification, contrastive-based methods like MoCo v3, DINO, and iBOT outperform DeiT by more than $2 \\%$ under both linear and attentive probing settings. 2) Though MIM-based models CAE and MAE are inferior to DeiT in object-level classification (e.g., more than $10 \\%$ and $4 \\%$ lower in linear and attentive probing), they show competitive performance in linear probing and higher results than DeiT in attentive probing, demonstrating they learn better partaware representations. 3) BEiT is inferior to other works, and iBOT has good performance, implying that the probing quality of pretrained encoders is a good indicator for downstream performance. ",
|
| 710 |
+
"bbox": [
|
| 711 |
+
173,
|
| 712 |
+
589,
|
| 713 |
+
826,
|
| 714 |
+
728
|
| 715 |
+
],
|
| 716 |
+
"page_idx": 7
|
| 717 |
+
},
|
| 718 |
+
{
|
| 719 |
+
"type": "text",
|
| 720 |
+
"text": "Part segmentation. We perform part-level linear semantic segmentation to study the finer-grained part representation modeling capability of different pretraining paradigms on three widely used datasets: ADE20K-Part (Zhou et al., 2019) containing 209 parts from the ADE20K dataset (Zhou et al., 2019), Pascal-Part (Chen et al., 2014) including 193 part categories, and LIP (Gong et al., 2017) consisting of 19 semantic human part labels. Similar to the object-level semantic segmentation experiments, linear evaluation is employed here. We maintain the same training protocols for all methods for fair comparisons. See Appendix A.2 and A.4 for dataset and training details. ",
|
| 721 |
+
"bbox": [
|
| 722 |
+
174,
|
| 723 |
+
736,
|
| 724 |
+
825,
|
| 725 |
+
833
|
| 726 |
+
],
|
| 727 |
+
"page_idx": 7
|
| 728 |
+
},
|
| 729 |
+
{
|
| 730 |
+
"type": "text",
|
| 731 |
+
"text": "The results are reported in Table 4 with the following observations. 1) Contrastive learning models, i.e., MoCo v3 and DINO, achieve competitive performance with the supervised model DeiT: DINO outperforms DeiT on ADE20K-Part and Pascal-Part, and MoCo v3 outperforms DeiT on LIP. 2) The MIM model CAE, outperforms DeiT by large margins on all three datasets, e.g., $1 . 1 \\%$ on ADE20K-Part and $2 . 3 \\%$ on LIP, indicating CAE learns good part-aware representations. Similar to part retrieval, possibly due to pretraining quality in representation encoding, BEiT and MAE perform inferior. 3) Compared with object-level segmentation results in Table 2, DeiT learns better object-level semantics by explicit supervision than both contrastive learning and MIM, however, it is generally inferior to self-supervised models on part segmentation. 4) The model iBOT, which leverages both contrastive learning and MIM, outperforms all other works on three datasets, demonstrating its powerful capability in learning finer part-level semantics. Combining the two self-supervised learning techniques is thus a promising direction. ",
|
| 732 |
+
"bbox": [
|
| 733 |
+
173,
|
| 734 |
+
840,
|
| 735 |
+
825,
|
| 736 |
+
924
|
| 737 |
+
],
|
| 738 |
+
"page_idx": 7
|
| 739 |
+
},
|
| 740 |
+
{
|
| 741 |
+
"type": "image",
|
| 742 |
+
"img_path": "images/7e9e5bfed8e41154ea9b2ae7018ddd58380f4a7bda4793113c1ca5d795c92e22.jpg",
|
| 743 |
+
"image_caption": [
|
| 744 |
+
"Figure 6: Comparisons between object-level and part-level semantic segmentation on ADE20K and Pascal-Part datasets. Though the supervised DeiT is superior over self-supervised models (i.e., MoCo v3, DINO, MAE, CAE) on object-level segmentation, it is generally inferior to self-supervised models on part segmentation, demonstrating self-supervised methods learn good part-aware representations. iBOT enjoys the benefits of contrastive learning and MIM. See Appendix A.3 for detailed results. "
|
| 745 |
+
],
|
| 746 |
+
"image_footnote": [],
|
| 747 |
+
"bbox": [
|
| 748 |
+
176,
|
| 749 |
+
101,
|
| 750 |
+
821,
|
| 751 |
+
251
|
| 752 |
+
],
|
| 753 |
+
"page_idx": 8
|
| 754 |
+
},
|
| 755 |
+
{
|
| 756 |
+
"type": "text",
|
| 757 |
+
"text": "",
|
| 758 |
+
"bbox": [
|
| 759 |
+
174,
|
| 760 |
+
354,
|
| 761 |
+
825,
|
| 762 |
+
439
|
| 763 |
+
],
|
| 764 |
+
"page_idx": 8
|
| 765 |
+
},
|
| 766 |
+
{
|
| 767 |
+
"type": "text",
|
| 768 |
+
"text": "In summary, we show that self-supervised methods are potentially capable of learning part-aware representations. Among them, CAE is a representative MIM work, showing good performance by explicitly predicting the encoded representations of the masked patches in the encoding space; contrastive learning methods MoCo v3 and DINO outperform BEiT and MAE; and iBOT performs the best by combining contrastive learning and MIM. The observations are evidenced by three part-based segmentation benchmarks consistently. ",
|
| 769 |
+
"bbox": [
|
| 770 |
+
174,
|
| 771 |
+
445,
|
| 772 |
+
825,
|
| 773 |
+
529
|
| 774 |
+
],
|
| 775 |
+
"page_idx": 8
|
| 776 |
+
},
|
| 777 |
+
{
|
| 778 |
+
"type": "text",
|
| 779 |
+
"text": "4.3 OBSERVATION SUMMARY BETWEEN OBJECT-LEVEL AND PART-LEVEL SEGMENTATION ",
|
| 780 |
+
"text_level": 1,
|
| 781 |
+
"bbox": [
|
| 782 |
+
173,
|
| 783 |
+
550,
|
| 784 |
+
816,
|
| 785 |
+
564
|
| 786 |
+
],
|
| 787 |
+
"page_idx": 8
|
| 788 |
+
},
|
| 789 |
+
{
|
| 790 |
+
"type": "text",
|
| 791 |
+
"text": "We conduct both object-level and part-level linear semantic segmentation on different hierarchies of the same dataset. Considering that the 209 classes in ADE20K-Part are basically chosen from 59 object classes, we denote the 59-object dataset as ADE20K-Object. Similarly, Pascal-Object consists of 16 object categories, corresponding to the 193 part categories in Pascal-Part. ",
|
| 792 |
+
"bbox": [
|
| 793 |
+
174,
|
| 794 |
+
578,
|
| 795 |
+
825,
|
| 796 |
+
633
|
| 797 |
+
],
|
| 798 |
+
"page_idx": 8
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"type": "text",
|
| 802 |
+
"text": "The results in Figure 6 show that: although the supervised DeiT is superior over contrastive learning and masked image modeling methods on ADE20K-Object and Pascal-Object except iBOT, it is generally inferior to self-supervised models on ADE20K-Part and Pascal-Part, demonstrating selfsupervised methods can learn good part-aware representations. Similar observations could be found from the object classification in Table 1 and part classification in Table 3. ",
|
| 803 |
+
"bbox": [
|
| 804 |
+
174,
|
| 805 |
+
641,
|
| 806 |
+
825,
|
| 807 |
+
710
|
| 808 |
+
],
|
| 809 |
+
"page_idx": 8
|
| 810 |
+
},
|
| 811 |
+
{
|
| 812 |
+
"type": "text",
|
| 813 |
+
"text": "In comparison to contrastive learning, CAE shows a stronger capability of learning part-aware representations, and a weaker capability of learning object-level semantics. The superiority of iBOT, a combination of contrastive learning and masked image modeling, demonstrates that it enjoys the benefits of contrastive learning and masked image modeling. ",
|
| 814 |
+
"bbox": [
|
| 815 |
+
174,
|
| 816 |
+
718,
|
| 817 |
+
825,
|
| 818 |
+
773
|
| 819 |
+
],
|
| 820 |
+
"page_idx": 8
|
| 821 |
+
},
|
| 822 |
+
{
|
| 823 |
+
"type": "text",
|
| 824 |
+
"text": "5 CONCLUSION ",
|
| 825 |
+
"text_level": 1,
|
| 826 |
+
"bbox": [
|
| 827 |
+
176,
|
| 828 |
+
797,
|
| 829 |
+
318,
|
| 830 |
+
814
|
| 831 |
+
],
|
| 832 |
+
"page_idx": 8
|
| 833 |
+
},
|
| 834 |
+
{
|
| 835 |
+
"type": "text",
|
| 836 |
+
"text": "We attempt to study the capability of learning part-aware representations of self-supervised representation pretraining methods. We provide speculations for contrastive learning and masked image modeling: part-to-whole and part-to-part, with empirical results justifying the speculations. Our study presents an aspect to understand what self-supervised representation pretraining methods learn. ",
|
| 837 |
+
"bbox": [
|
| 838 |
+
174,
|
| 839 |
+
833,
|
| 840 |
+
825,
|
| 841 |
+
888
|
| 842 |
+
],
|
| 843 |
+
"page_idx": 8
|
| 844 |
+
},
|
| 845 |
+
{
|
| 846 |
+
"type": "text",
|
| 847 |
+
"text": "Future work. The strong capability of part-aware representation learning is one of the properties of self-supervised pretraining. There should be other characteristics that are leaved as the future work. ",
|
| 848 |
+
"bbox": [
|
| 849 |
+
173,
|
| 850 |
+
895,
|
| 851 |
+
823,
|
| 852 |
+
924
|
| 853 |
+
],
|
| 854 |
+
"page_idx": 8
|
| 855 |
+
},
|
| 856 |
+
{
|
| 857 |
+
"type": "text",
|
| 858 |
+
"text": "REFERENCES ",
|
| 859 |
+
"text_level": 1,
|
| 860 |
+
"bbox": [
|
| 861 |
+
174,
|
| 862 |
+
102,
|
| 863 |
+
287,
|
| 864 |
+
118
|
| 865 |
+
],
|
| 866 |
+
"page_idx": 9
|
| 867 |
+
},
|
| 868 |
+
{
|
| 869 |
+
"type": "text",
|
| 870 |
+
"text": "Hangbo Bao, Li Dong, and Furu Wei. BEiT: BERT pre-training of image transformers. arXiv:2106.08254, 2021. ",
|
| 871 |
+
"bbox": [
|
| 872 |
+
173,
|
| 873 |
+
126,
|
| 874 |
+
823,
|
| 875 |
+
155
|
| 876 |
+
],
|
| 877 |
+
"page_idx": 9
|
| 878 |
+
},
|
| 879 |
+
{
|
| 880 |
+
"type": "text",
|
| 881 |
+
"text": "Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. arXiv preprint arXiv:2105.04906, 2021. ",
|
| 882 |
+
"bbox": [
|
| 883 |
+
171,
|
| 884 |
+
165,
|
| 885 |
+
823,
|
| 886 |
+
194
|
| 887 |
+
],
|
| 888 |
+
"page_idx": 9
|
| 889 |
+
},
|
| 890 |
+
{
|
| 891 |
+
"type": "text",
|
| 892 |
+
"text": "Suzanna Becker and Geoffrey E. Hinton. A self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355(6356):161–163, 1992. ",
|
| 893 |
+
"bbox": [
|
| 894 |
+
171,
|
| 895 |
+
203,
|
| 896 |
+
823,
|
| 897 |
+
233
|
| 898 |
+
],
|
| 899 |
+
"page_idx": 9
|
| 900 |
+
},
|
| 901 |
+
{
|
| 902 |
+
"type": "text",
|
| 903 |
+
"text": "Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Coco-stuff: Thing and stuff classes in context. In CVPR, pp. 1209–1218, 2018. ",
|
| 904 |
+
"bbox": [
|
| 905 |
+
171,
|
| 906 |
+
242,
|
| 907 |
+
825,
|
| 908 |
+
271
|
| 909 |
+
],
|
| 910 |
+
"page_idx": 9
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"type": "text",
|
| 914 |
+
"text": "Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022. ",
|
| 915 |
+
"bbox": [
|
| 916 |
+
171,
|
| 917 |
+
280,
|
| 918 |
+
825,
|
| 919 |
+
310
|
| 920 |
+
],
|
| 921 |
+
"page_idx": 9
|
| 922 |
+
},
|
| 923 |
+
{
|
| 924 |
+
"type": "text",
|
| 925 |
+
"text": "Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. arXiv preprint arXiv:2006.09882, 2020. ",
|
| 926 |
+
"bbox": [
|
| 927 |
+
173,
|
| 928 |
+
319,
|
| 929 |
+
825,
|
| 930 |
+
362
|
| 931 |
+
],
|
| 932 |
+
"page_idx": 9
|
| 933 |
+
},
|
| 934 |
+
{
|
| 935 |
+
"type": "text",
|
| 936 |
+
"text": "Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. ",
|
| 937 |
+
"bbox": [
|
| 938 |
+
171,
|
| 939 |
+
372,
|
| 940 |
+
825,
|
| 941 |
+
401
|
| 942 |
+
],
|
| 943 |
+
"page_idx": 9
|
| 944 |
+
},
|
| 945 |
+
{
|
| 946 |
+
"type": "text",
|
| 947 |
+
"text": "Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, volume 119 of Proceedings of Machine Learning Research, pp. 1597–1607. PMLR, 2020. ",
|
| 948 |
+
"bbox": [
|
| 949 |
+
174,
|
| 950 |
+
410,
|
| 951 |
+
823,
|
| 952 |
+
454
|
| 953 |
+
],
|
| 954 |
+
"page_idx": 9
|
| 955 |
+
},
|
| 956 |
+
{
|
| 957 |
+
"type": "text",
|
| 958 |
+
"text": "Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, pp. 1971–1978, 2014. ",
|
| 959 |
+
"bbox": [
|
| 960 |
+
173,
|
| 961 |
+
463,
|
| 962 |
+
825,
|
| 963 |
+
506
|
| 964 |
+
],
|
| 965 |
+
"page_idx": 9
|
| 966 |
+
},
|
| 967 |
+
{
|
| 968 |
+
"type": "text",
|
| 969 |
+
"text": "Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, and Jingdong Wang. Context autoencoder for self-supervised representation learning. arXiv preprint arXiv:2202.03026, 2022a. ",
|
| 970 |
+
"bbox": [
|
| 971 |
+
176,
|
| 972 |
+
515,
|
| 973 |
+
825,
|
| 974 |
+
559
|
| 975 |
+
],
|
| 976 |
+
"page_idx": 9
|
| 977 |
+
},
|
| 978 |
+
{
|
| 979 |
+
"type": "text",
|
| 980 |
+
"text": "Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pp. 15750–15758, 2021. ",
|
| 981 |
+
"bbox": [
|
| 982 |
+
173,
|
| 983 |
+
568,
|
| 984 |
+
823,
|
| 985 |
+
598
|
| 986 |
+
],
|
| 987 |
+
"page_idx": 9
|
| 988 |
+
},
|
| 989 |
+
{
|
| 990 |
+
"type": "text",
|
| 991 |
+
"text": "Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pp. 9640–9649, 2021. ",
|
| 992 |
+
"bbox": [
|
| 993 |
+
173,
|
| 994 |
+
607,
|
| 995 |
+
821,
|
| 996 |
+
637
|
| 997 |
+
],
|
| 998 |
+
"page_idx": 9
|
| 999 |
+
},
|
| 1000 |
+
{
|
| 1001 |
+
"type": "text",
|
| 1002 |
+
"text": "Yubei Chen, Adrien Bardes, Zengyi Li, and Yann LeCun. Intra-instance vicreg: Bag of self-supervised image patch embedding. arXiv preprint arXiv:2206.08954, 2022b. ",
|
| 1003 |
+
"bbox": [
|
| 1004 |
+
173,
|
| 1005 |
+
645,
|
| 1006 |
+
823,
|
| 1007 |
+
676
|
| 1008 |
+
],
|
| 1009 |
+
"page_idx": 9
|
| 1010 |
+
},
|
| 1011 |
+
{
|
| 1012 |
+
"type": "text",
|
| 1013 |
+
"text": "MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020. ",
|
| 1014 |
+
"bbox": [
|
| 1015 |
+
173,
|
| 1016 |
+
684,
|
| 1017 |
+
825,
|
| 1018 |
+
714
|
| 1019 |
+
],
|
| 1020 |
+
"page_idx": 9
|
| 1021 |
+
},
|
| 1022 |
+
{
|
| 1023 |
+
"type": "text",
|
| 1024 |
+
"text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pp. 248–255. Ieee, 2009. ",
|
| 1025 |
+
"bbox": [
|
| 1026 |
+
173,
|
| 1027 |
+
723,
|
| 1028 |
+
823,
|
| 1029 |
+
753
|
| 1030 |
+
],
|
| 1031 |
+
"page_idx": 9
|
| 1032 |
+
},
|
| 1033 |
+
{
|
| 1034 |
+
"type": "text",
|
| 1035 |
+
"text": "Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Peco: Perceptual codebook for bert pre-training of vision transformers. arXiv preprint arXiv:2111.12710, 2021. ",
|
| 1036 |
+
"bbox": [
|
| 1037 |
+
174,
|
| 1038 |
+
762,
|
| 1039 |
+
825,
|
| 1040 |
+
805
|
| 1041 |
+
],
|
| 1042 |
+
"page_idx": 9
|
| 1043 |
+
},
|
| 1044 |
+
{
|
| 1045 |
+
"type": "text",
|
| 1046 |
+
"text": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. ",
|
| 1047 |
+
"bbox": [
|
| 1048 |
+
174,
|
| 1049 |
+
814,
|
| 1050 |
+
826,
|
| 1051 |
+
871
|
| 1052 |
+
],
|
| 1053 |
+
"page_idx": 9
|
| 1054 |
+
},
|
| 1055 |
+
{
|
| 1056 |
+
"type": "text",
|
| 1057 |
+
"text": "Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. ",
|
| 1058 |
+
"bbox": [
|
| 1059 |
+
176,
|
| 1060 |
+
882,
|
| 1061 |
+
825,
|
| 1062 |
+
924
|
| 1063 |
+
],
|
| 1064 |
+
"page_idx": 9
|
| 1065 |
+
},
|
| 1066 |
+
{
|
| 1067 |
+
"type": "text",
|
| 1068 |
+
"text": "M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2010 (VOC2010) Results. http://www.pascalnetwork.org/challenges/VOC/voc2010/workshop/index.html, 2010. ",
|
| 1069 |
+
"bbox": [
|
| 1070 |
+
174,
|
| 1071 |
+
103,
|
| 1072 |
+
821,
|
| 1073 |
+
146
|
| 1074 |
+
],
|
| 1075 |
+
"page_idx": 10
|
| 1076 |
+
},
|
| 1077 |
+
{
|
| 1078 |
+
"type": "text",
|
| 1079 |
+
"text": "Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Selfsupervised structure-sensitive learning and a new benchmark for human parsing. In CVPR, July 2017. ",
|
| 1080 |
+
"bbox": [
|
| 1081 |
+
176,
|
| 1082 |
+
155,
|
| 1083 |
+
823,
|
| 1084 |
+
198
|
| 1085 |
+
],
|
| 1086 |
+
"page_idx": 10
|
| 1087 |
+
},
|
| 1088 |
+
{
|
| 1089 |
+
"type": "text",
|
| 1090 |
+
"text": "Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. ",
|
| 1091 |
+
"bbox": [
|
| 1092 |
+
173,
|
| 1093 |
+
207,
|
| 1094 |
+
826,
|
| 1095 |
+
263
|
| 1096 |
+
],
|
| 1097 |
+
"page_idx": 10
|
| 1098 |
+
},
|
| 1099 |
+
{
|
| 1100 |
+
"type": "text",
|
| 1101 |
+
"text": "Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. ",
|
| 1102 |
+
"bbox": [
|
| 1103 |
+
171,
|
| 1104 |
+
272,
|
| 1105 |
+
823,
|
| 1106 |
+
303
|
| 1107 |
+
],
|
| 1108 |
+
"page_idx": 10
|
| 1109 |
+
},
|
| 1110 |
+
{
|
| 1111 |
+
"type": "text",
|
| 1112 |
+
"text": "Xiangwen Kong and Xiangyu Zhang. Understanding masked image modeling via learning occlusion invariant feature. arXiv preprint arXiv:2208.04164, 2022. ",
|
| 1113 |
+
"bbox": [
|
| 1114 |
+
171,
|
| 1115 |
+
311,
|
| 1116 |
+
823,
|
| 1117 |
+
340
|
| 1118 |
+
],
|
| 1119 |
+
"page_idx": 10
|
| 1120 |
+
},
|
| 1121 |
+
{
|
| 1122 |
+
"type": "text",
|
| 1123 |
+
"text": "Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Processing Systems, 34:13165–13176, 2021. ",
|
| 1124 |
+
"bbox": [
|
| 1125 |
+
174,
|
| 1126 |
+
349,
|
| 1127 |
+
823,
|
| 1128 |
+
392
|
| 1129 |
+
],
|
| 1130 |
+
"page_idx": 10
|
| 1131 |
+
},
|
| 1132 |
+
{
|
| 1133 |
+
"type": "text",
|
| 1134 |
+
"text": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pp. 740–755. Springer, 2014. ",
|
| 1135 |
+
"bbox": [
|
| 1136 |
+
176,
|
| 1137 |
+
401,
|
| 1138 |
+
823,
|
| 1139 |
+
444
|
| 1140 |
+
],
|
| 1141 |
+
"page_idx": 10
|
| 1142 |
+
},
|
| 1143 |
+
{
|
| 1144 |
+
"type": "text",
|
| 1145 |
+
"text": "Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. ",
|
| 1146 |
+
"bbox": [
|
| 1147 |
+
171,
|
| 1148 |
+
454,
|
| 1149 |
+
825,
|
| 1150 |
+
483
|
| 1151 |
+
],
|
| 1152 |
+
"page_idx": 10
|
| 1153 |
+
},
|
| 1154 |
+
{
|
| 1155 |
+
"type": "text",
|
| 1156 |
+
"text": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pp. 10012–10022, 2021. ",
|
| 1157 |
+
"bbox": [
|
| 1158 |
+
173,
|
| 1159 |
+
492,
|
| 1160 |
+
825,
|
| 1161 |
+
535
|
| 1162 |
+
],
|
| 1163 |
+
"page_idx": 10
|
| 1164 |
+
},
|
| 1165 |
+
{
|
| 1166 |
+
"type": "text",
|
| 1167 |
+
"text": "Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pp. 8821–8831. PMLR, 2021. ",
|
| 1168 |
+
"bbox": [
|
| 1169 |
+
173,
|
| 1170 |
+
544,
|
| 1171 |
+
825,
|
| 1172 |
+
574
|
| 1173 |
+
],
|
| 1174 |
+
"page_idx": 10
|
| 1175 |
+
},
|
| 1176 |
+
{
|
| 1177 |
+
"type": "text",
|
| 1178 |
+
"text": "Nikunj Saunshi, Jordan T. Ash, Surbhi Goel, Dipendra Misra, Cyril Zhang, Sanjeev Arora, Sham M. Kakade, and Akshay Krishnamurthy. Understanding contrastive learning requires incorporating inductive biases. CoRR, abs/2202.14037, 2022. URL https://arxiv.org/abs/2202. 14037. ",
|
| 1179 |
+
"bbox": [
|
| 1180 |
+
173,
|
| 1181 |
+
582,
|
| 1182 |
+
826,
|
| 1183 |
+
638
|
| 1184 |
+
],
|
| 1185 |
+
"page_idx": 10
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"type": "text",
|
| 1189 |
+
"text": "Chenxin Tao, Xizhou Zhu, Gao Huang, Yu Qiao, Xiaogang Wang, and Jifeng Dai. Siamese image modeling for self-supervised vision representation learning. arXiv preprint arXiv:2206.01204, 2022. ",
|
| 1190 |
+
"bbox": [
|
| 1191 |
+
173,
|
| 1192 |
+
648,
|
| 1193 |
+
825,
|
| 1194 |
+
690
|
| 1195 |
+
],
|
| 1196 |
+
"page_idx": 10
|
| 1197 |
+
},
|
| 1198 |
+
{
|
| 1199 |
+
"type": "text",
|
| 1200 |
+
"text": "Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. arXiv preprint arXiv:2012.12877, 2020. ",
|
| 1201 |
+
"bbox": [
|
| 1202 |
+
173,
|
| 1203 |
+
700,
|
| 1204 |
+
823,
|
| 1205 |
+
743
|
| 1206 |
+
],
|
| 1207 |
+
"page_idx": 10
|
| 1208 |
+
},
|
| 1209 |
+
{
|
| 1210 |
+
"type": "text",
|
| 1211 |
+
"text": "Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. ",
|
| 1212 |
+
"bbox": [
|
| 1213 |
+
169,
|
| 1214 |
+
752,
|
| 1215 |
+
825,
|
| 1216 |
+
781
|
| 1217 |
+
],
|
| 1218 |
+
"page_idx": 10
|
| 1219 |
+
},
|
| 1220 |
+
{
|
| 1221 |
+
"type": "text",
|
| 1222 |
+
"text": "Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. arXiv preprint arXiv:2205.14141, 2022. ",
|
| 1223 |
+
"bbox": [
|
| 1224 |
+
174,
|
| 1225 |
+
790,
|
| 1226 |
+
825,
|
| 1227 |
+
833
|
| 1228 |
+
],
|
| 1229 |
+
"page_idx": 10
|
| 1230 |
+
},
|
| 1231 |
+
{
|
| 1232 |
+
"type": "text",
|
| 1233 |
+
"text": "Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swin transformers. arXiv preprint arXiv:2105.04553, 2021a. ",
|
| 1234 |
+
"bbox": [
|
| 1235 |
+
169,
|
| 1236 |
+
842,
|
| 1237 |
+
825,
|
| 1238 |
+
872
|
| 1239 |
+
],
|
| 1240 |
+
"page_idx": 10
|
| 1241 |
+
},
|
| 1242 |
+
{
|
| 1243 |
+
"type": "text",
|
| 1244 |
+
"text": "Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021b. ",
|
| 1245 |
+
"bbox": [
|
| 1246 |
+
174,
|
| 1247 |
+
882,
|
| 1248 |
+
826,
|
| 1249 |
+
922
|
| 1250 |
+
],
|
| 1251 |
+
"page_idx": 10
|
| 1252 |
+
},
|
| 1253 |
+
{
|
| 1254 |
+
"type": "text",
|
| 1255 |
+
"text": "Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. CoRR, abs/2205.13543, 2022a. doi: 10.48550/arXiv.2205. 13543. URL https://doi.org/10.48550/arXiv.2205.13543. ",
|
| 1256 |
+
"bbox": [
|
| 1257 |
+
174,
|
| 1258 |
+
103,
|
| 1259 |
+
825,
|
| 1260 |
+
146
|
| 1261 |
+
],
|
| 1262 |
+
"page_idx": 11
|
| 1263 |
+
},
|
| 1264 |
+
{
|
| 1265 |
+
"type": "text",
|
| 1266 |
+
"text": "Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Yixuan Wei, Qi Dai, and Han Hu. On data scaling in masked image modeling. CoRR, abs/2206.04664, 2022b. doi: 10.48550/arXiv.2206.04664. URL https://doi.org/10.48550/arXiv.2206.04664. ",
|
| 1267 |
+
"bbox": [
|
| 1268 |
+
176,
|
| 1269 |
+
155,
|
| 1270 |
+
823,
|
| 1271 |
+
196
|
| 1272 |
+
],
|
| 1273 |
+
"page_idx": 11
|
| 1274 |
+
},
|
| 1275 |
+
{
|
| 1276 |
+
"type": "text",
|
| 1277 |
+
"text": "Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. arXiv preprint arXiv:2103.03230, 2021. ",
|
| 1278 |
+
"bbox": [
|
| 1279 |
+
171,
|
| 1280 |
+
207,
|
| 1281 |
+
823,
|
| 1282 |
+
234
|
| 1283 |
+
],
|
| 1284 |
+
"page_idx": 11
|
| 1285 |
+
},
|
| 1286 |
+
{
|
| 1287 |
+
"type": "text",
|
| 1288 |
+
"text": "Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning? arXiv preprint arXiv:2206.05259, 2022. ",
|
| 1289 |
+
"bbox": [
|
| 1290 |
+
173,
|
| 1291 |
+
243,
|
| 1292 |
+
825,
|
| 1293 |
+
273
|
| 1294 |
+
],
|
| 1295 |
+
"page_idx": 11
|
| 1296 |
+
},
|
| 1297 |
+
{
|
| 1298 |
+
"type": "text",
|
| 1299 |
+
"text": "Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127(3):302–321, 2019. ",
|
| 1300 |
+
"bbox": [
|
| 1301 |
+
174,
|
| 1302 |
+
281,
|
| 1303 |
+
825,
|
| 1304 |
+
310
|
| 1305 |
+
],
|
| 1306 |
+
"page_idx": 11
|
| 1307 |
+
},
|
| 1308 |
+
{
|
| 1309 |
+
"type": "text",
|
| 1310 |
+
"text": "Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. ",
|
| 1311 |
+
"bbox": [
|
| 1312 |
+
174,
|
| 1313 |
+
319,
|
| 1314 |
+
825,
|
| 1315 |
+
349
|
| 1316 |
+
],
|
| 1317 |
+
"page_idx": 11
|
| 1318 |
+
},
|
| 1319 |
+
{
|
| 1320 |
+
"type": "text",
|
| 1321 |
+
"text": "A APPENDIX ",
|
| 1322 |
+
"text_level": 1,
|
| 1323 |
+
"bbox": [
|
| 1324 |
+
176,
|
| 1325 |
+
102,
|
| 1326 |
+
297,
|
| 1327 |
+
117
|
| 1328 |
+
],
|
| 1329 |
+
"page_idx": 12
|
| 1330 |
+
},
|
| 1331 |
+
{
|
| 1332 |
+
"type": "text",
|
| 1333 |
+
"text": "A.1 MODEL DESCRIPTION ",
|
| 1334 |
+
"text_level": 1,
|
| 1335 |
+
"bbox": [
|
| 1336 |
+
176,
|
| 1337 |
+
133,
|
| 1338 |
+
370,
|
| 1339 |
+
148
|
| 1340 |
+
],
|
| 1341 |
+
"page_idx": 12
|
| 1342 |
+
},
|
| 1343 |
+
{
|
| 1344 |
+
"type": "text",
|
| 1345 |
+
"text": "For all the models involved in the experiments including DeiT (Touvron et al., 2020), MoCo v3 (Chen et al., 2021), DINO (Caron et al., 2021), BEiT (Bao et al., 2021), MAE (He et al., 2021), CAE (Chen et al., 2022a), and iBOT (Zhou et al., 2021), we use their official code to implement the encoders. It is worth noticing that for DINO and iBOT, we choose the checkpoint of the teacher models as they have been reported to perform better than the student models in their papers (Caron et al., 2021; Zhou et al., 2021). ",
|
| 1346 |
+
"bbox": [
|
| 1347 |
+
174,
|
| 1348 |
+
160,
|
| 1349 |
+
825,
|
| 1350 |
+
244
|
| 1351 |
+
],
|
| 1352 |
+
"page_idx": 12
|
| 1353 |
+
},
|
| 1354 |
+
{
|
| 1355 |
+
"type": "text",
|
| 1356 |
+
"text": "A.2 DATASETS ",
|
| 1357 |
+
"text_level": 1,
|
| 1358 |
+
"bbox": [
|
| 1359 |
+
176,
|
| 1360 |
+
262,
|
| 1361 |
+
290,
|
| 1362 |
+
277
|
| 1363 |
+
],
|
| 1364 |
+
"page_idx": 12
|
| 1365 |
+
},
|
| 1366 |
+
{
|
| 1367 |
+
"type": "text",
|
| 1368 |
+
"text": "ADE20K (Zhou et al., 2019) is one of the most challenging benchmarks, containing 150 fine-grained semantic concepts and a variety of scenes with 1,038 image-level labels. There are 20,210 images in the training set and 2,000 images in the validation set. We choose 59 out of total 150 semantic concepts that are concrete objects containing parts (Zhou et al., 2019), termed ADE20K-Object. We also select 209 part categories that emerge both in the training set and the validation set, called ADE20K-Part. ",
|
| 1369 |
+
"bbox": [
|
| 1370 |
+
174,
|
| 1371 |
+
289,
|
| 1372 |
+
825,
|
| 1373 |
+
372
|
| 1374 |
+
],
|
| 1375 |
+
"page_idx": 12
|
| 1376 |
+
},
|
| 1377 |
+
{
|
| 1378 |
+
"type": "text",
|
| 1379 |
+
"text": "Pascal-Part (Chen et al., 2014) is a set of additional annotations for PASCAL VOC 2010 (Everingham et al., 2010), thereby holding the same statistics as those of PASCAL VOC 2010. It provides segmentation masks for each part of objects. Concretely, the dataset includes 20 object-level categories and 193 part-level categories. In our experiments, we remove 4 object categories that do not contain parts including boat, table, chair, and sofa. ",
|
| 1380 |
+
"bbox": [
|
| 1381 |
+
174,
|
| 1382 |
+
380,
|
| 1383 |
+
825,
|
| 1384 |
+
449
|
| 1385 |
+
],
|
| 1386 |
+
"page_idx": 12
|
| 1387 |
+
},
|
| 1388 |
+
{
|
| 1389 |
+
"type": "text",
|
| 1390 |
+
"text": "LIP (Gong et al., 2017) is a large-scale benchmark for human parsing research, which includes 50,462 images with pixel-wise annotations on 19 semantic part labels. In detail, it includes 19,081 full-body images, 13,672 upper-body images, 403 lower-body images, 3,386 head-missed images, 2,778 back-view images and 21,028 images with occlusions. There are 30,462 images in the training set and 10,000 images in the validation set. The rest 10,000 images are served as the test set with missing labels for competition evaluation. ",
|
| 1391 |
+
"bbox": [
|
| 1392 |
+
174,
|
| 1393 |
+
457,
|
| 1394 |
+
825,
|
| 1395 |
+
540
|
| 1396 |
+
],
|
| 1397 |
+
"page_idx": 12
|
| 1398 |
+
},
|
| 1399 |
+
{
|
| 1400 |
+
"type": "text",
|
| 1401 |
+
"text": "CUB-200-2011 (Wah et al., 2011) is a popular benchmark for fine-grained image classification, and also provides bounding box and part location annotations. It contains 11,788 images of 200 bird species and 15 part keypoint annotations per bird. In this work, we mainly leverage its part keypoint annotations. And only 4 part categories (right eye, right leg, left wing, and tail) are chosen to be considered in our experiments, to make sure that the selected keypoints are far enough away from each other and enough context information can be contained in the cropped patches. (We also tried using all keypoints and the conclusion is consistent.) ",
|
| 1402 |
+
"bbox": [
|
| 1403 |
+
174,
|
| 1404 |
+
546,
|
| 1405 |
+
825,
|
| 1406 |
+
645
|
| 1407 |
+
],
|
| 1408 |
+
"page_idx": 12
|
| 1409 |
+
},
|
| 1410 |
+
{
|
| 1411 |
+
"type": "text",
|
| 1412 |
+
"text": "COCO (Caesar et al., 2018), as one of the most widely-used human pose estimation datasets, contains more than 200,000 images and 250,000 labeled person instances. Similar to CUB-200-2011 mentioned above, only 3 (nose, right wrist, and left ankle) of its 17 keypoint categories are considered in our experiments. ",
|
| 1413 |
+
"bbox": [
|
| 1414 |
+
174,
|
| 1415 |
+
651,
|
| 1416 |
+
825,
|
| 1417 |
+
708
|
| 1418 |
+
],
|
| 1419 |
+
"page_idx": 12
|
| 1420 |
+
},
|
| 1421 |
+
{
|
| 1422 |
+
"type": "text",
|
| 1423 |
+
"text": "A.3 DETAILED RESULTS FOR OBJECT-LEVEL AND PART-LEVEL SEGMENTATION ",
|
| 1424 |
+
"text_level": 1,
|
| 1425 |
+
"bbox": [
|
| 1426 |
+
176,
|
| 1427 |
+
726,
|
| 1428 |
+
741,
|
| 1429 |
+
741
|
| 1430 |
+
],
|
| 1431 |
+
"page_idx": 12
|
| 1432 |
+
},
|
| 1433 |
+
{
|
| 1434 |
+
"type": "text",
|
| 1435 |
+
"text": "In this section, we provide detailed comparisons between object-level and part-level semantic segmentation in Table 5 and Table 6. Similar observations as in Figure 6 in the main paper are found: although the supervised DeiT is superior over self-supervised methods on ADE20K-Object and Pascal-Object except iBOT, it is generally inferior to self-supervised models on ADE20K-Part and Pascal-Part, demonstrating self-supervised methods can learn good part-aware representations. BEiT and MAE perform inferior, perhaps because the two methods do not have an explicit process to predict the encoded representations of masked patches, instead, directly reconstruct the targets. ",
|
| 1436 |
+
"bbox": [
|
| 1437 |
+
174,
|
| 1438 |
+
752,
|
| 1439 |
+
825,
|
| 1440 |
+
851
|
| 1441 |
+
],
|
| 1442 |
+
"page_idx": 12
|
| 1443 |
+
},
|
| 1444 |
+
{
|
| 1445 |
+
"type": "text",
|
| 1446 |
+
"text": "A.4 EXPERIMENT DETAILS ",
|
| 1447 |
+
"text_level": 1,
|
| 1448 |
+
"bbox": [
|
| 1449 |
+
176,
|
| 1450 |
+
869,
|
| 1451 |
+
375,
|
| 1452 |
+
882
|
| 1453 |
+
],
|
| 1454 |
+
"page_idx": 12
|
| 1455 |
+
},
|
| 1456 |
+
{
|
| 1457 |
+
"type": "text",
|
| 1458 |
+
"text": "Part retrieval. In our part retrieval experiments, we directly use the pretrained encoders to extract features, without additional training process. For each method, we take the better one from the class token or the average embedding of all patch tokens as the extracted representation. With each patch as the query patch, we calculate the cosine similarity between its representation and all the other patches’ in the dataset and utilize the average precision (AP) as the retrieval metric. Finally, we average all the obtained AP scores (with all patches respectively taken as the query patch for retrieval) as the final retrieval score of the method. ",
|
| 1459 |
+
"bbox": [
|
| 1460 |
+
174,
|
| 1461 |
+
895,
|
| 1462 |
+
823,
|
| 1463 |
+
924
|
| 1464 |
+
],
|
| 1465 |
+
"page_idx": 12
|
| 1466 |
+
},
|
| 1467 |
+
{
|
| 1468 |
+
"type": "table",
|
| 1469 |
+
"img_path": "images/98df0ecb2a39d7473678e7af06bb277af58e24ae89c45913b8c222497211851a.jpg",
|
| 1470 |
+
"table_caption": [
|
| 1471 |
+
"Table 5: Linear semantic segmentation results on ADE20K-Object and ADE20K-Part. "
|
| 1472 |
+
],
|
| 1473 |
+
"table_footnote": [],
|
| 1474 |
+
"table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"3\">Object Seg on ADE20K-Object 59 Object Classes</td><td colspan=\"3\">Part Seg on ADE20K-Part 209 Part Classes</td></tr><tr><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td colspan=\"7\">Supervised Model:</td></tr><tr><td>DeiT</td><td>52.6</td><td>62.9</td><td>83.8</td><td>27.3</td><td>34.7</td><td>69.2</td></tr><tr><td colspan=\"7\">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>50.2</td><td>60.4</td><td>83.6</td><td>27.1</td><td>34.7</td><td>70.1</td></tr><tr><td>DINO</td><td>50.8</td><td>60.8</td><td>83.9</td><td>28.9</td><td>36.8</td><td>70.3</td></tr><tr><td colspan=\"7\">MaskedImageModeling (MIM):</td></tr><tr><td>BEiT</td><td>28.6</td><td>37.2</td><td>73.4</td><td>18.6</td><td>25.8</td><td>58.2</td></tr><tr><td>MAE</td><td>41.0</td><td>50.6</td><td>79.9</td><td>26.3</td><td>35.0</td><td>67.3</td></tr><tr><td>CAE</td><td>47.4</td><td>58.4</td><td>82.9</td><td>28.4</td><td>36.9</td><td>71.1</td></tr><tr><td colspan=\"7\">Contrastive Learning+MIM:</td></tr><tr><td>iBOT</td><td>55.2</td><td>65.1</td><td>85.6</td><td>32.2</td><td>40.0</td><td>73.4</td></tr></table>",
|
| 1475 |
+
"bbox": [
|
| 1476 |
+
223,
|
| 1477 |
+
127,
|
| 1478 |
+
774,
|
| 1479 |
+
347
|
| 1480 |
+
],
|
| 1481 |
+
"page_idx": 13
|
| 1482 |
+
},
|
| 1483 |
+
{
|
| 1484 |
+
"type": "table",
|
| 1485 |
+
"img_path": "images/3773c2164bde15831281f162890e83245185573890267b93301fdf753e50aada.jpg",
|
| 1486 |
+
"table_caption": [
|
| 1487 |
+
"Table 6: Linear semantic segmentation results on Pascal-Object and Pascal-Part. "
|
| 1488 |
+
],
|
| 1489 |
+
"table_footnote": [],
|
| 1490 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">Object Seg on Pascal-Object 16 Object Classes</td><td colspan=\"3\">Part Seg on Pascal-Part 193 Part Classes</td></tr><tr><td>mIoU</td><td>mAcc</td><td>aAcc</td><td>mIoU</td><td>mAcc</td><td>aAcc</td></tr><tr><td>Supervised Model: DeiT</td><td>92.2</td><td>95.3</td><td>96.8</td><td>27.4</td><td>36.2</td><td>65.8</td></tr><tr><td colspan=\"7\">Contrastive Learning:</td></tr><tr><td>MoCo v3</td><td>89.4</td><td>93.7</td><td>95.7</td><td>27.1</td><td>35.8</td><td>66.0</td></tr><tr><td>DINO</td><td>88.0</td><td>92.7</td><td>95.3</td><td>27.8</td><td>36.5</td><td>66.4</td></tr><tr><td colspan=\"3\">Masked Image Modeling (MIM):</td><td></td><td></td><td></td><td></td></tr><tr><td>BEiT</td><td>56.4</td><td>69.0</td><td>76.8</td><td>14.8</td><td>21.4</td><td>47.0</td></tr><tr><td>MAE</td><td>76.1</td><td>84.6</td><td>89.5</td><td>24.3</td><td>32.9</td><td>61.5</td></tr><tr><td>CAE</td><td>83.3</td><td>89.7</td><td>93.2</td><td>27.8</td><td>37.0</td><td>66.3</td></tr><tr><td colspan=\"3\">Contrastive Learning+ MIM:</td><td></td><td></td><td></td><td></td></tr><tr><td>iBOT</td><td>92.1</td><td>95.3</td><td>97.1</td><td>30.7</td><td>40.0</td><td>69.7</td></tr></table>",
|
| 1491 |
+
"bbox": [
|
| 1492 |
+
227,
|
| 1493 |
+
392,
|
| 1494 |
+
772,
|
| 1495 |
+
613
|
| 1496 |
+
],
|
| 1497 |
+
"page_idx": 13
|
| 1498 |
+
},
|
| 1499 |
+
{
|
| 1500 |
+
"type": "text",
|
| 1501 |
+
"text": "",
|
| 1502 |
+
"bbox": [
|
| 1503 |
+
174,
|
| 1504 |
+
645,
|
| 1505 |
+
825,
|
| 1506 |
+
714
|
| 1507 |
+
],
|
| 1508 |
+
"page_idx": 13
|
| 1509 |
+
},
|
| 1510 |
+
{
|
| 1511 |
+
"type": "text",
|
| 1512 |
+
"text": "Apart from the part retrieval experiments shown in Table 3, the visualized patch retrieval results in Figures 2 and 4 are obtained based on ImageNet (Deng et al., 2009) validation set. Concretely, from each pre-processed $2 2 4 \\times 2 2 4$ validation image in ImageNet, we uniformly crop 49 patches sized $5 6 \\times 5 6$ using a stride of 28. With all the cropped patches from the validation set, we select one patch as a query and find top 24 patches with the highest cosine similarity with it. ",
|
| 1513 |
+
"bbox": [
|
| 1514 |
+
174,
|
| 1515 |
+
722,
|
| 1516 |
+
825,
|
| 1517 |
+
791
|
| 1518 |
+
],
|
| 1519 |
+
"page_idx": 13
|
| 1520 |
+
},
|
| 1521 |
+
{
|
| 1522 |
+
"type": "text",
|
| 1523 |
+
"text": "Part classification. For linear probing, we learn a supervised linear classification layer on the extracted class token of the frozen encoders. While for attentive probing, following Chen et al. (2022a), a cross attention module and a batch normalization layer without affine transformation are additionally inserted between the encoder and the linear classifier. And a new learnable class token is taken as the query of the cross attention module, to replace the original class token extracted by the frozen encoder. We use SGD optimizer with a learning rate of 0.4 and 0.04 for linear probing and attentive probing, respectively. For both linear probing and attentive probing, the models are trained for 90 epochs. And the momentum of SGD is set to 0.9, the weight decay is set to 0, and the batch size is set to 1024. ",
|
| 1524 |
+
"bbox": [
|
| 1525 |
+
173,
|
| 1526 |
+
797,
|
| 1527 |
+
825,
|
| 1528 |
+
924
|
| 1529 |
+
],
|
| 1530 |
+
"page_idx": 13
|
| 1531 |
+
},
|
| 1532 |
+
{
|
| 1533 |
+
"type": "image",
|
| 1534 |
+
"img_path": "images/18b1c0330479d57d3e6da0d49c80f25f4cb1c1b6153f80750b9d818819fa367e.jpg",
|
| 1535 |
+
"image_caption": [
|
| 1536 |
+
"Figure 7: Patch retrieval comparisons of encoded representations on cropped patches from ImageNet. "
|
| 1537 |
+
],
|
| 1538 |
+
"image_footnote": [],
|
| 1539 |
+
"bbox": [
|
| 1540 |
+
179,
|
| 1541 |
+
87,
|
| 1542 |
+
818,
|
| 1543 |
+
493
|
| 1544 |
+
],
|
| 1545 |
+
"page_idx": 14
|
| 1546 |
+
},
|
| 1547 |
+
{
|
| 1548 |
+
"type": "text",
|
| 1549 |
+
"text": "Segmentation. We use the same model structure that contains a parameter-fixed pretrained encoder (e.g., MAE and DeiT) and a simple learnable $1 \\times 1$ convolutional layer for object-level and part-level segmentation tasks. Note that the learning rate $( 4 e - 4 )$ , training iterations $( 1 6 0 k )$ , and batch size (16) among all the experiments maintain the same during training for fair comparisons. For ADE20K, the input size is set to $5 1 2 \\times 5 1 2$ following previous works (Bao et al., 2021; He et al., 2021; Chen et al., 2022a; Zhou et al., 2021). For Pascal-Part, we adopt $4 8 0 \\times 4 8 0$ as image input resolution following Contributors (2020). As for LIP, we use the same input size $3 2 0 \\times 3 2 0 )$ proposed in LIP (Gong et al., 2017). ",
|
| 1550 |
+
"bbox": [
|
| 1551 |
+
173,
|
| 1552 |
+
546,
|
| 1553 |
+
826,
|
| 1554 |
+
659
|
| 1555 |
+
],
|
| 1556 |
+
"page_idx": 14
|
| 1557 |
+
},
|
| 1558 |
+
{
|
| 1559 |
+
"type": "text",
|
| 1560 |
+
"text": "A.5 IMAGENET PATCH RETRIEVAL VISUALIZATION ",
|
| 1561 |
+
"text_level": 1,
|
| 1562 |
+
"bbox": [
|
| 1563 |
+
174,
|
| 1564 |
+
675,
|
| 1565 |
+
545,
|
| 1566 |
+
689
|
| 1567 |
+
],
|
| 1568 |
+
"page_idx": 14
|
| 1569 |
+
},
|
| 1570 |
+
{
|
| 1571 |
+
"type": "text",
|
| 1572 |
+
"text": "We visualize more patch retrieval results of the encoded representations on the ImageNet validation set in Figures 7 and 8. It is observed that the retrieved patches of self-supervised methods are generally more about the semantics of the query part than that of DeiT. The results demonstrate that the encoded representations of DeiT focus more on object-level semantics, while the encoded representations of these self-supervised methods are more about part-level semantics. Among these methods, the retrieved patches of MAE have less semantic correlation but often share similar hues. ",
|
| 1573 |
+
"bbox": [
|
| 1574 |
+
174,
|
| 1575 |
+
702,
|
| 1576 |
+
825,
|
| 1577 |
+
785
|
| 1578 |
+
],
|
| 1579 |
+
"page_idx": 14
|
| 1580 |
+
},
|
| 1581 |
+
{
|
| 1582 |
+
"type": "image",
|
| 1583 |
+
"img_path": "images/4ab87d028fea0a8d5b3ec58fc792500037d785b2e1ab6904eaaded062f33fc7a.jpg",
|
| 1584 |
+
"image_caption": [
|
| 1585 |
+
"Figure 8: Patch retrieval comparisons of encoded representations on cropped patches from ImageNet. "
|
| 1586 |
+
],
|
| 1587 |
+
"image_footnote": [],
|
| 1588 |
+
"bbox": [
|
| 1589 |
+
179,
|
| 1590 |
+
59,
|
| 1591 |
+
818,
|
| 1592 |
+
902
|
| 1593 |
+
],
|
| 1594 |
+
"page_idx": 15
|
| 1595 |
+
}
|
| 1596 |
+
]
|
parse/dev/3tYvDb4dwab/3tYvDb4dwab_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/3tYvDb4dwab/3tYvDb4dwab_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/6HN7LHyzGgC/6HN7LHyzGgC.md
ADDED
|
@@ -0,0 +1,398 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# UNCERTAINTY MODELING FOR OUT-OF-DISTRIBUTION GENERALIZATION
|
| 2 |
+
|
| 3 |
+
Xiaotong $\mathbf { L i } ^ { 1 }$ , Yongxing $\mathbf { D a i } ^ { 1 }$ , Yixiao $\mathbf { G e ^ { 2 } }$ , $\mathbf { J u n L i u ^ { 3 } }$ , Ying Shan2, Ling-Yu Duan1,4∗
|
| 4 |
+
1Peking University, Beijing, China 2ARC Lab, Tencent PCG
|
| 5 |
+
3Singapore University of Technology and Design, Singapore
|
| 6 |
+
4Peng Cheng Laboratory, Shenzhen, China
|
| 7 |
+
lixiaotong@stu.pku.edu.cn, {yongxingdai, lingyu}@pku.edu.cn,
|
| 8 |
+
{yixiaoge, yingsshan}@tencent.com, jun liu@sutd.edu.sg
|
| 9 |
+
|
| 10 |
+
# ABSTRACT
|
| 11 |
+
|
| 12 |
+
Though remarkable progress has been achieved in various vision tasks, deep neural networks still suffer obvious performance degradation when tested in out-ofdistribution scenarios. We argue that the feature statistics (mean and standard deviation), which carry the domain characteristics of the training data, can be properly manipulated to improve the generalization ability of deep learning models. Common methods often consider the feature statistics as deterministic values measured from the learned features and do not explicitly consider the uncertain statistics discrepancy caused by potential domain shifts during testing. In this paper, we improve the network generalization ability by modeling the uncertainty of domain shifts with synthesized feature statistics during training. Specifically, we hypothesize that the feature statistic, after considering the potential uncertainties, follows a multivariate Gaussian distribution. Hence, each feature statistic is no longer a deterministic value, but a probabilistic point with diverse distribution possibilities. With the uncertain feature statistics, the models can be trained to alleviate the domain perturbations and achieve better robustness against potential domain shifts. Our method can be readily integrated into networks without additional parameters. Extensive experiments demonstrate that our proposed method consistently improves the network generalization ability on multiple vision tasks, including image classification, semantic segmentation, and instance retrieval. The code can be available at https://github.com/lixiaotong97/DSU.
|
| 13 |
+
|
| 14 |
+
# 1 INTRODUCTION
|
| 15 |
+
|
| 16 |
+
Deep neural networks have shown impressive success in computer vision, but with a severe reliance on the assumption that the training and testing domains follow an independent and identical distribution (Ben-David et al., 2010; Vapnik, 1992). This assumption, however, does not hold in many real-world applications. For instance, when employing segmentation models trained on sunny days for rainy and foggy environments (Choi et al., 2021), or recognizing art paintings with models that trained on photographs (Li et al., 2017), inevitable performance drop can often be observed in such out-of-distribution deployment scenarios. Therefore, the problem of domain generalization, aiming to improve the robustness of the network on various unseen testing domains, becomes quite important.
|
| 17 |
+
|
| 18 |
+
Previous works (Huang & Belongie, 2017; Li et al., 2021) demonstrate that feature statistics (mean and standard deviation), as the moments of the learned features, carry informative domain characteristics of the training data. Domain characteristics primarily refer to the information that is more specific to the individual domains but less relevant to the task objectives, such as the photo style and capturing environment information in object recognition. Consequently, domains with different data distributions generally have inconsistent feature statistics (Wang et al., 2020b; 2019a; Gao et al., 2021a). Most deep learning methods follow Empirical Risk Minimization principle (Vapnik, 1999) to minimize their average error over the training data (Shen et al., 2021). Despite the satisfactory performance on the training domain, these methods do not explicitly consider the uncertain statistics discrepancy caused by potential domain shifts during testing. As a result, the trained models tend to overfit the training domain and show vulnerability to the statistic changes at testing time, substantially limiting the generalization ability of the learned representations.
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
Figure 1: The visualization of reconstructed samples with synthesized feature statistics, using a pre-trained style transfer auto-encoder (Huang & Belongie, 2017). The illustration of the feature statistics shifts, which may vary in both intensity and direction (i.e., different offsets in the vector space of feature statistics). We also show images of “new” domains generated by manipulating feature statistic shifts with different direction and intensity. Note these images are for visualization only, rather than feeding into the network for training.
|
| 22 |
+
|
| 23 |
+
Intuitively, the test domains may bring uncertain statistics shifts with different potential directions and intensities compared to the training domain (as shown in Figure 1), implying the uncertain nature of domain shifts. Considering such “uncertainty” of potential domain shifts, synthesizing novel feature statistics variants to model diverse domain shifts can improve the robustness of the trained network to different testing distributions. Towards this end, we introduce a novel probabilistic method to improve the network generalization ability by properly modeling Domain Shifts with Uncertainty (DSU), i.e., characterizing the feature statistics as uncertain distributions.
|
| 24 |
+
|
| 25 |
+
In our method, instead of treating each feature statistic as a deterministic point measured from the feature, we hypothesize that the feature statistic, after considering potential uncertainties, follows a multi-variate Gaussian distribution. The distribution “center” is set as each feature’s original statistic value, and the distribution “scope” represents the variant intensity considering underlying domain shifts. Uncertainty estimation is adopted here to depict the distribution “scope” of probabilistic feature statistics. Specifically, we estimate the distribution “scope” based on the variances of the mini-batch statistics in an efficient non-parametric manner. Subsequently, feature statistics variants are randomly sampled from the estimated Gaussian distribution and then used to replace the original deterministic values for modeling diverse domain shifts, as illustrated in Figure 2. Due to the generated feature statistics with diverse distribution possibilities, the models can be trained to properly alleviate the domain perturbations and encode better domain-invariant features.
|
| 26 |
+
|
| 27 |
+
Our proposed method is simple yet fairly effective to alleviate performance drop caused by domain shifts, and can be readily integrated into existing networks without bringing additional model parameters or loss constraints. Comprehensive experiments on a wide range of vision tasks demonstrate the superiority of our proposed method, indicating that introducing uncertainty to feature statistics can well improve models’ generalization against domain shifts.
|
| 28 |
+
|
| 29 |
+
# 2 RELATED WORK
|
| 30 |
+
|
| 31 |
+
# 2.1 DOMAIN GENERALIZATION
|
| 32 |
+
|
| 33 |
+
Domain generalization (DG) has been attracting increasing attention in the past few years, which aims to achieve out-of-distribution generalization on unseen target domains using only single or multiple source domain data for training (Blanchard et al. (2011)). Research on addressing this problem has been extensively conducted in the literature (Zhou et al. (2021a); Wang et al. (2021); Shen et al. (2021)). Here some studies that are more related to our work are introduced below.
|
| 34 |
+
|
| 35 |
+
Data Augmentation: Data augmentation is an effective manner for improving generalization ability and relieving models from overfitting in training domains. Most augmentation methods adopt various transformations at the image level, such as AugMix (Hendrycks et al. (2020)) and CutMix (Yun et al. (2019)). Besides using handcraft transformations, mixup (Zhang et al. (2018)) trains the model by using pair-wise linearly interpolated samples in both the image and label spaces. Manifold Mixup (Verma et al. (2019)) further adopts this linear interpolation from image level to feature level. Some recent works extend the above transformations to feature statistics for improving model generalization. MixStyle (Zhou et al. (2021b)) adopts linear interpolation on feature statistics of two instances to generate synthesized samples. The pAdaIn (Nuriel et al. (2021)) swaps statistics between the samples applied with a random permutation of the batch.
|
| 36 |
+
|
| 37 |
+
Invariant Representation Learning: The main idea of invariant representation learning is to enable models to learn features that are invariant to domain shifts. Domain alignment-based approaches (Li et al. (2018c;b)) learn invariant features by minimizing the distances between different distributions. Instead of enforcing the entire features to be invariant, disentangled feature learning approaches (Chattopadhyay et al. (2020); Piratla et al. (2020)) decouple the features into domain-specific and domain-invariant parts and learn their representations simultaneously. In addition, normalizationbased methods (Pan et al. (2018); Choi et al. (2021)) can also be used to remove the style information to obtain invariant representations.
|
| 38 |
+
|
| 39 |
+
Learning Strategies: There are also some effective learning strategies that can be leveraged to improve generalization ability. Ensemble learning is an effective technique in boosting model performance. The ensemble predictions using a collection of diverse models (Zhou et al. (2020b)) or modules (Seo et al. (2020)) can be adopted to improve generalization and robustness. Metalearning-based methods (Finn et al. (2017); Li et al. (2018a); Dai et al. (2021)) learn to simulate the domain shifts following an episode training paradigm. Besides, self-challenging methods, such as RSC (Huang et al. (2020)), force the model to learn a general representation by discarding dominant features activated on the training data.
|
| 40 |
+
|
| 41 |
+
# 2.2 UNCERTAINTY IN DEEP LEARNING
|
| 42 |
+
|
| 43 |
+
Uncertainty capturing the “noise” and “randomness” inherent in the data has received increasing attention in deep representation learning. Variational Auto-encoder (Kingma & Welling (2013)), as an important method for learning generative models, can be regarded as a method to model the data uncertainty in the hidden space. Dropout (Srivastava et al. (2014)), which is widely used in many deep learning models to avoid over-fitting, can be interpreted to represent model uncertainty as a Bayesian approximation (Gal & Ghahramani (2016)). In some works, uncertainty is used to address the issues of low-quality training data. In person re-identification, DistributionNet (Gal & Ghahramani (2016)) adopts uncertainty to model the person images of noise-labels and outliers. In face recognition, DUL (Chang et al. (2020)) and PFE ((Shi & Jain, 2019)) apply data uncertainty to simultaneously learn the feature embedding and its uncertainty, where the uncertainty is learned through a learnable subnetwork to describe the quality of the image. Different from the aforementioned works, our proposed method is used to model the feature statistics uncertainty under potential domain shifts and acts as a feature augmentation method for handling our-of-distribution generalization problem.
|
| 44 |
+
|
| 45 |
+
# 3 METHOD
|
| 46 |
+
|
| 47 |
+
# 3.1 PRELIMINARIES
|
| 48 |
+
|
| 49 |
+
Given $x \in \mathbb { R } ^ { B \times C \times H \times W }$ to be the encoded features in the intermediate layers of the network, we denote $\mu \in \mathbb { R } ^ { B \times C }$ and $\sigma \in \mathbb { R } ^ { B \times C }$ as the channel-wise feature mean and standard deviation of each instance in a mini-batch, respectively, which can be formulated as:
|
| 50 |
+
|
| 51 |
+
$$
|
| 52 |
+
\begin{array} { l } { \displaystyle \mu ( x ) = \frac { 1 } { H W } \sum _ { h = 1 } ^ { H } \sum _ { w = 1 } ^ { W } { x _ { b , c , h , w } } , } \\ { \displaystyle \sigma ^ { 2 } ( x ) = \frac { 1 } { H W } \sum _ { h = 1 } ^ { H } \sum _ { w = 1 } ^ { W } ( { x _ { b , c , h , w } - \mu ( x ) } ) ^ { 2 } . } \end{array}
|
| 53 |
+
$$
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
Figure 2: Illustration of the proposed method. Feature statistic is assumed to follow a multi-variate Gaussian distribution during training. When passed through this module, the new feature statistics randomly drawn from the corresponding distribution will replace the original ones to model the diverse domain shifts.
|
| 57 |
+
|
| 58 |
+
As the abstraction of features, feature statistics can capture informative characteristics of the corresponding domain (such as color, texture, and contrast), according to previous works (Huang & Belongie, 2017; Li et al., 2021). In out-of-distribution scenarios, the feature statistics often show inconsistency with training domain due to different domain characteristics (Wang et al., $2 0 1 9 \mathrm { a }$ ; Gao et al., 2021a), which is ill-suited to deep learning modules like nonlinearity layers and normalization layer and degenerates the model’s generalization ability (Wang et al., 2020b). However, most of the deep learning methods only treat feature statistics as deterministic values measured from the features while lacking explicit consideration of the potential uncertain statistical discrepancy. Owing to the model’s inherent vulnerability to such discrepancy, the generalization ability of the learned representations is limited. Some recent methods (Nuriel et al., 2021; Zhou et al., 2021b) utilize feature statistics to tackle the domain generalization problem. Despite the success, they typically adopt linear manipulation (i.e., exchange and interpolation) on pairwise samples to generate new feature statistics, which limits the diversity of synthetic changes. Specifically, the direction of their variants is determined by the chosen reference sample and such internal operation restricts their variant intensity. Thus these methods are sub-optimal when handling the diverse and uncertain domain shifts in real world.
|
| 59 |
+
|
| 60 |
+
# 3.2 MODELING DOMAIN SHIFTS WITH UNCERTAINTY
|
| 61 |
+
|
| 62 |
+
Given the arbitrary testing domains with uncertain feature statistic shifts in both direction and intensity, properly modeling the domain shifts becomes an important task for tackling the challenge of domain generalization problem.
|
| 63 |
+
|
| 64 |
+
Considering the uncertainty and randomness of domain shifts, it is promising to employ the methods of “uncertainty” to treat the “uncertainty” of domain shifts. In this paper, we propose a novel method by modeling Domain Shifts with Uncertainty (DSU). Instead of treating each feature statistic as a deterministic value measured from the learned feature, we hypothesize that the distribution of each feature statistic, after considering potential uncertainties, follows a multi-variate Gaussian distribution. This means each feature statistic has a probabilistic representation drawn from a certain distribution, i.e., the feature statistics mean and standard deviation follow $\mathcal { N } ( \mu , \Sigma _ { \mu } ^ { 2 } )$ and $\textstyle \mathcal { N } ( \sigma , \Sigma _ { \sigma } ^ { 2 } )$ , respectively. Specifically, the corresponding Gaussian distribution’s center is set as each feature’s original statistics, while the Gaussian distribution’s standard deviation describes the uncertainty scope for different potential shifts. Through randomly sampling diverse synthesized feature statistics with the probabilistic approach, the models can be trained to improve the robustness of the network against statistics shifts.
|
| 65 |
+
|
| 66 |
+
# 3.2.1 UNCERTAINTY ESTIMATION
|
| 67 |
+
|
| 68 |
+
Taking the uncertainty of domain shifts into consideration, the uncertainty estimation in our method aims to depict the uncertainty scope of each probabilistic feature statistic. However, the testing domain is unknown, which makes it challenging to obtain an appropriate variant range.
|
| 69 |
+
|
| 70 |
+
Some generative-based studies (Shen & Zhou, 2021; Wang et al., 2019b) show that the variances between features contain implicit semantic meaning and the directions with larger variances can imply potentials of more valuable semantic changes. Inspired by this, we propose a simple yet effective non-parametric method for uncertainty estimation, utilizing the variance of the feature statistics to provide some instructions:
|
| 71 |
+
|
| 72 |
+
$$
|
| 73 |
+
\begin{array} { l } { { \Sigma _ { \mu } ^ { 2 } ( x ) = \displaystyle \frac { 1 } { B } \sum _ { b = 1 } ^ { B } ( \mu ( x ) - \mathbb { E } _ { b } [ \mu ( x ) ] ) ^ { 2 } , } } \\ { { \Sigma _ { \sigma } ^ { 2 } ( x ) = \displaystyle \frac { 1 } { B } \sum _ { b = 1 } ^ { B } ( \sigma ( x ) - \mathbb { E } _ { b } [ \sigma ( x ) ] ) ^ { 2 } . } } \end{array}
|
| 74 |
+
$$
|
| 75 |
+
|
| 76 |
+
where $\Sigma _ { \mu } \in \mathbb { R } ^ { C }$ and $\Sigma _ { \sigma } \in \mathbb { R } ^ { C }$ represent the uncertainty estimation of the feature mean $\mu$ and feature standard deviation $\sigma$ , respectively. The magnitudes of uncertainty estimation can reveal the possibility that the corresponding channel may change potentially. Although the underlying distribution of the domain shifts is unpredictable, the uncertainty estimation captured from the minibatch can provide an appropriate and meaningful variation range for each feature channel, which does not harm model training but can simulate diverse potential shifts.
|
| 77 |
+
|
| 78 |
+
# 3.2.2 PROBABILISTIC DISTRIBUTION OF FEATURE STATISTICS
|
| 79 |
+
|
| 80 |
+
Once the uncertainty estimation of each feature channel is obtained, the Gaussian distribution for probabilistic feature statistics can be established. To use randomness to model the uncertainty, we adopt the random sampling to further exploit the uncertainty in the probabilistic representations. The new feature statistics, mean $\beta ( x ) \sim { \mathrm { \bar { \mathcal { N } } } } ( \mu , \Sigma _ { \mu } ^ { 2 } )$ and standard deviation $\gamma ( x ) \sim \bar { \mathcal { N } } ( \sigma , \Sigma _ { \sigma } ^ { 2 } )$ , can be randomly drawn from the corresponding distributions as:
|
| 81 |
+
|
| 82 |
+
$$
|
| 83 |
+
\begin{array} { l l } { { \beta ( x ) = \mu ( x ) + \epsilon _ { \mu } \Sigma _ { \mu } ( x ) , } } & { { \epsilon _ { \mu } \sim \mathcal { N } ( \mathbf { 0 } , \mathbf { 1 } ) , } } \\ { { } } & { { } } \\ { { \gamma ( x ) = \sigma ( x ) + \epsilon _ { \sigma } \Sigma _ { \sigma } ( x ) , } } & { { \epsilon _ { \sigma } \sim \mathcal { N } ( \mathbf { 0 } , \mathbf { 1 } ) . } } \end{array}
|
| 84 |
+
$$
|
| 85 |
+
|
| 86 |
+
Here we use the re-parameterization trick (Kingma & Welling (2013)) to make the sampling operation differentiable, and $\epsilon _ { \mu }$ and $\epsilon _ { \sigma }$ both follow the standard Gaussian distribution. By exploiting the given Gaussian distribution, random sampling can generate various new feature statistics information with different combinations of directions and intensities.
|
| 87 |
+
|
| 88 |
+
# 3.2.3 IMPLEMENTATION
|
| 89 |
+
|
| 90 |
+
The implementation of our method is by the means of AdaIN (Huang & Belongie (2017)), and replaces the feature statistics with the randomly drawing ones to achieve the transformation. The final form of the proposed method can be formulated as:
|
| 91 |
+
|
| 92 |
+
$$
|
| 93 |
+
\begin{array} { r } { \mathrm { D S U } ( x ) = \underbrace { \left( \sigma ( x ) + \epsilon _ { \sigma } \Sigma _ { \sigma } ( x ) \right) } _ { \gamma ( x ) } \left( \frac { x - \mu ( x ) } { \sigma ( x ) } \right) + \underbrace { \left( \mu ( x ) + \epsilon _ { \mu } \Sigma _ { \mu } ( x ) \right) } _ { \beta ( x ) } . } \end{array}
|
| 94 |
+
$$
|
| 95 |
+
|
| 96 |
+
The above operation can be integrated at various positions of the network as a flexible module. Note that the module only works during model training and can be discarded while testing. To trade off the strength of this module, we set a hyperparameter $p$ that denotes the probability to apply it. The algorithm is described in the Appendix. Benefiting from the proposed method, the model trained with uncertain feature statistics will gain better robustness against potential statistics shifts, and thus acquires a better generalization ability.
|
| 97 |
+
|
| 98 |
+
# 4 EXPERIMENTS
|
| 99 |
+
|
| 100 |
+
In order to verify the effectiveness of the proposed method in improving the generalization ability of networks, we conduct the experiments on a wide range of tasks, including image classification, semantic segmentation, instance retrieval, and robustness towards corruptions, where the training and testing sets have different cases of distribution shifts, such as style shift, synthetic-to-real gap, scenes change, and pixel-level corruption.
|
| 101 |
+
|
| 102 |
+
# 4.1 GENERALIZATION ON MULTI-DOMAIN CLASSIFICATION
|
| 103 |
+
|
| 104 |
+
Setup and Implementation Details: We evaluate the proposed method on PACS (Li et al. (2017)), a widely-used benchmark for domain generalization with four different styles: Art Painting, Cartoon, Photo, and Sketch. The implementation follows the official setup of MixStyle (Zhou et al. (2021b)) with a leave-one-domain-out protocol and ResNet18 (He et al., 2016) is used as the backbone. The random shuffle version of MixStyle is adopted for fair comparisons, which does not use domain labels. In addition to PACS, we also employ Office-Home (Venkateswara et al., 2017) for multidomain generalization experiments in the Appendix.
|
| 105 |
+
|
| 106 |
+
Experiment Results: The experiments results, shown in Table 1, demonstrate our significant improvement over the baseline method, which shows our superiority to the conventional deterministic approach. Especially in Art and Sketch, our method has nearly $10 \%$ improvement in average accuracy. Furthermore, our method also outperforms the competing methods, which indicates our method that models diverse uncertain shifts on feature statistics is effective to improve network generalization ability against different domain shifts. Photo has similiar domain characteristics as ImageNet dataset and the slight drop might be due to the ImageNet pretraining (also discussed in (Xu et al., 2021)). Our DSU augments the features and enlarges the diversity of the training domains. In contrast, the baseline method preserves more pre-trained knowledge from ImageNet thus tends to overfit the Photo style dataset benefiting from pretraining.
|
| 107 |
+
|
| 108 |
+
Table 1: Experiment results of PACS multi-domain classification task. ${ \mathrm { R S C } } ^ { * }$ denotes the reproduced results from pAdaIN (Nuriel et al., 2021).
|
| 109 |
+
|
| 110 |
+
<table><tr><td>Method</td><td>Reference</td><td>Art</td><td>Cartoon</td><td>Photo</td><td>Sketch</td><td>Average (%)</td></tr><tr><td>Baseline</td><td>=</td><td>74.3</td><td>76.7</td><td>96.4</td><td>68.7</td><td>79.0</td></tr><tr><td>Mixup (Zhang et al., 2018)</td><td>ICLR 2018</td><td>76.8</td><td>74.9</td><td>95.8</td><td>66.6</td><td>78.5</td></tr><tr><td>Manifold Mixup (Verma et al.,2019)</td><td>ICML 2019</td><td>75.6</td><td>70.1</td><td>93.5</td><td>65.4</td><td>76.2</td></tr><tr><td>CutMix (Yun et al., 2019)</td><td>ICCV 2019</td><td>74.6</td><td>71.8</td><td>95.6</td><td>65.3</td><td>76.8</td></tr><tr><td>RSC*(Huang et al.,2020)</td><td>ECCV 2020</td><td>78.9</td><td>76.9</td><td>94.1</td><td>76.8</td><td>81.7</td></tr><tr><td>L2A-OT (Zhou et al., 2020a)</td><td>ECCV 2020</td><td>83.3</td><td>78.2</td><td>96.2</td><td>73.6</td><td>82.8</td></tr><tr><td>SagNet (Nam et al., 2021)</td><td>CVPR 2021</td><td>83.6</td><td>77.7</td><td>95.5</td><td>76.3</td><td>83.3</td></tr><tr><td>pAdaIN (Nuriel et al., 2021)</td><td>CVPR 2021</td><td>81.7</td><td>76.6</td><td>96.3</td><td>75.1</td><td>82.5</td></tr><tr><td>MixStyle (Zhou et al., 2021b)</td><td>ICLR 2021</td><td>82.3</td><td>79.0</td><td>96.3</td><td>73.8</td><td>82.8</td></tr><tr><td>DSU</td><td>Ours</td><td>83.6</td><td>79.6</td><td>95.8</td><td>77.6</td><td>84.1</td></tr></table>
|
| 111 |
+
|
| 112 |
+
# 4.2 GENERALIZATION ON SEMANTIC SEGMENTATION
|
| 113 |
+
|
| 114 |
+
Setup and Implementation Details: Semantic segmentation, as a fundamental application for automatic driving, encounters severe performance declines due to scenarios differences (Wang et al., 2020a). GTA5 (Richter et al., 2016) is a synthetic dataset generated from Grand Theft Auto 5 game engine, while Cityscapes (Cordts et al., 2016) is a real-world dataset collected from different cities in primarily Germany. To evaluate the cross-scenario generalization ability of segmentation models, we adopt synthetic GTA5 for training while using real CityScapes for testing. The experiments are conducted on FADA released codes (Wang et al. (2020a)), using DeepLab-v2 (Chen et al., 2018) segmentation network with ResNet101 backbone. Mean Intersection over Union (mIOU) and mean Accuracy (mAcc) of all object categories are used for evaluation.
|
| 115 |
+
|
| 116 |
+
Table 2: Experiment results of semantic segmentation from synthetic GAT5 to real Cityscapes.
|
| 117 |
+
|
| 118 |
+
<table><tr><td>Method</td><td>Reference</td><td>mIOU (%)</td><td>mAcc (%)</td></tr><tr><td>Baseline</td><td>-</td><td>37.0</td><td>51.5</td></tr><tr><td>pAdaIN (Nuriel et al., 2021)</td><td>CVPR 2021</td><td>38.3</td><td>52.1</td></tr><tr><td>Mixstyle (Zhou et al., 2021b)</td><td>ICLR 2021</td><td>40.3</td><td>53.8</td></tr><tr><td>DSU</td><td>Ours</td><td>43.1</td><td>57.0</td></tr></table>
|
| 119 |
+
|
| 120 |
+
Experiment Results: Table 2 shows the experiment results compared to related methods. As for a pixel-level classification task, improper changes of feature statistics might constrain the performances. The variants generated from our method are centered on the original feature statistics with different perturbations. These changes of feature statistics are mild for preserving the detailed information in these dense tasks. Meanwhile, our method can take full use of the diverse driving scenes and generates diverse variants, thus show a significant improvement on mIOU and mAc by $6 . 1 \%$ and $5 . 5 \%$ , respectively. The visualization result is shown in Figure 3.
|
| 121 |
+
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
Figure 3: The visualization on unseen domain Cityscapes with the model trained on synthetic GTA5.
|
| 125 |
+
|
| 126 |
+
# 4.3 GENERALIZATION ON INSTANCE RETRIEVAL
|
| 127 |
+
|
| 128 |
+
Setup and Implementation Details: In this section, person re-identification (ReID), which aims at matching the same person across disjoint camera views, is used to verify the effectiveness of our method on the instance retrieval task. Experiments are conducted on the widely used DukeMTMC (Ristani et al. (2016)) and Market1501 (Zheng et al. (2015)) datasets. The implementation is based on MMT (Ge et al., 2020) released codes and ResNet50 is adopted as the backbone. Meanwhile, mean Average Precision (mAP) and Rank-1 (R1) precision are used as the evaluation criterions.
|
| 129 |
+
|
| 130 |
+
Table 3: Experiment results of instance retrieval on ReID dataset DukeMTMC and Market1501. A $ \mathbf { B }$ denotes models are trained on A while evaluated on B. For fair comparisons, we reproduce the experiments under the same framework.
|
| 131 |
+
|
| 132 |
+
<table><tr><td>Method</td><td>Reference</td><td>Market → Duke mAP (%) R1 (%)</td><td>mAP (%)</td><td>Duke-→Market R1 (%)</td></tr><tr><td>Baseline</td><td>1</td><td>25.8</td><td>42.3 46.1</td><td>26.7</td></tr><tr><td>pAdaIN (Nuriel et al., 2021)</td><td>CVPR 2021</td><td>28.0</td><td>27.9</td><td>54.7 56.1</td></tr><tr><td>MixStyle (Zhou et al.,2021b)</td><td>ICLR 2021</td><td>28.2</td><td>28.1</td><td>56.6</td></tr><tr><td>DSU</td><td>Ours</td><td>32.0</td><td>46.7 52.0 32.4</td><td>63.7</td></tr></table>
|
| 133 |
+
|
| 134 |
+
Experiment Results: ReID is a fine-grained instance retrieval task, where the subtle information of persons is important for retrieving an instance. MixStyle and pAdain rely on a reference sample to generate new feature statistics, which might introduce confounded information from the reference sample. Compared to them, our method does better in maintaining the original information and also has more variant possibilities. The experiment results are demonstrated in Table 3. Our method achieves huge improvement compared to the baseline method and also outperform MixStyle and pAdaIN by a big margin.
|
| 135 |
+
|
| 136 |
+
# 4.4 ROBUSTNESS TOWARDS CORRUPTIONS
|
| 137 |
+
|
| 138 |
+
Setup and Implementation Details: We validate the proposed method for robustness towards corruptions on ImageNet-C (Hendrycks & Dietterich (2019)), which contains 15 different pixel-level corruptions. ResNet50 is trained with 100 epochs for convergence on large-scale ImageNet-1K (Deng et al. (2009)) and the hyperparameter $p$ is set as 0.1 for training in ImageNet. We also add our method on APR (Chen et al. (2021)), a recently state-of-the-art method on ImageNet-C, to verify that our method can be compatible with other image-level augmentation methods. Error is adopted as the evalution metric for clean ImageNet. Mean Corruption Error (mCE) is adopted as evaluation metric for ImageNet-C, which is computed as the average of the 15 different corruption errors and normalized by the corruption error of AlexNet (Krizhevsky et al. (2012)).
|
| 139 |
+
|
| 140 |
+
Experiment Results: Although the corruptions are imposed on the pixel level, they still introduce a shift in the statistics (Benz et al., 2021). So our method shows consistent improvement on ImageNetC. Meanwhile, the instances in the testing set may not always fall into the distribution of the training set, and they still have slight statistic shifts (Gao et al. (2021b)). Thus it can be seen that the withindataset ImageNet accuracy is also increased. When combining APR with our method, mCE can be decreased from $6 5 . 0 \%$ to $6 4 . 1 \%$ , showing that our method can be compatible with state-of-the-art methods on ImageNet-C for further improvement.
|
| 141 |
+
|
| 142 |
+
Table 4: Experiment results of clean image classification on ImageNet, and the robustness toward corruptions on ImageNet-C.
|
| 143 |
+
|
| 144 |
+
<table><tr><td></td><td>Clean(↓）</td><td>Corrupted (↓)</td><td></td><td>Noise</td><td>Implulse</td><td></td><td>Blur</td><td></td><td></td><td></td><td>Weather</td><td></td><td></td><td></td><td>Digital</td><td></td><td></td></tr><tr><td></td><td>Error (%)</td><td>mCE(%)</td><td>Gauss</td><td>Shot</td><td></td><td>Defocus</td><td>Glass</td><td>Motion</td><td>Zoom</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Bright</td><td>Contrast</td><td>Elastic</td><td>Pixel</td><td>JPEG</td></tr><tr><td>Baseline</td><td>23.8</td><td>76.2</td><td></td><td>81</td><td></td><td></td><td>87</td><td>76</td><td>80</td><td>78</td><td>74</td><td>67</td><td></td><td>70</td><td>83</td><td>76</td><td>73</td></tr><tr><td>DSU</td><td>23.4</td><td>73.4</td><td>6</td><td>77</td><td>5</td><td>五</td><td>83</td><td>77</td><td>79</td><td>74</td><td>71</td><td>66</td><td>5</td><td>68</td><td>82</td><td>65</td><td>71</td></tr><tr><td>APR</td><td>24.0</td><td>65.0</td><td>5</td><td></td><td>50</td><td>69</td><td>85</td><td>69</td><td>79</td><td>62</td><td>64</td><td>55</td><td>54</td><td>63</td><td>84</td><td>65</td><td>65</td></tr><tr><td>APR+DSU</td><td>23.7</td><td>64.1</td><td></td><td>56</td><td>49</td><td>69</td><td>84</td><td>67</td><td>78</td><td>61</td><td>63</td><td>51</td><td>53</td><td>56</td><td>83</td><td>66</td><td>65</td></tr></table>
|
| 145 |
+
|
| 146 |
+
# 5 ABLATION STUDY
|
| 147 |
+
|
| 148 |
+
In this section, we perform an extensive ablation study of the proposed method on PACS and segmentation task (GTA5 to Cityscapes) with models trained on ResNet. The effects of different inserted positions and hyper-parameter of the proposed method are analyzed below. Meantime, we also analyze the effects on different choices of uncertainty distribution.
|
| 149 |
+
|
| 150 |
+
Effects of Different Inserted Positions: DSU can be a plugand-play module to be readily inserted at any position. Here we name the positions of ResNet after first Conv, Max Pooling
|
| 151 |
+
|
| 152 |
+
Table 5: Effects of different inserted positions.
|
| 153 |
+
|
| 154 |
+
<table><tr><td>Inserted Positions</td><td>Baseline</td><td>0-3</td><td>1-4</td><td>2-5</td><td>0-5</td></tr><tr><td>PACS</td><td>79.0</td><td>82.2</td><td>83.1</td><td>83.5</td><td>84.1</td></tr><tr><td> GTA5 to Cityscapes</td><td>37.0</td><td>41.1</td><td>40.9</td><td>42.1</td><td>43.1</td></tr></table>
|
| 155 |
+
|
| 156 |
+
layer, 1,2,3,4-th ConvBlock as 0,1,2,3,4,5 respectively. As shown in Table 5, no matter where the modules are inserted, the performances are consistently higher than the baseline method. The results show that inserting the modules at positions 0-5 would have better performances, which also indicates modeling the uncertainty in all training stages will have better effects. Based on the analysis, we plug the module into positions 0-5 in all experiments.
|
| 157 |
+
|
| 158 |
+
Effects of Hyper-parameter: The hyper-parameter of the probability $p$ is to trade off the strength of feature statistics augmentation. As shown in Figure 4, the results are not sensitive to the probability setting and the accuracy reaches the best results when setting $p$ as 0.5, which is also adopted as the default setting in all experiments if not specified.
|
| 159 |
+
|
| 160 |
+
Choices of Uncertainty Distribution: In our method, the Gaussian distribution with uncertainty estimation is adopted as the default setting, we also conduct other distributions for comparisons in Table 6. Specifically, Random denotes directly adding random shifts draw from a fixed Gaussian $\mathcal { N } ( 0 , 1 )$ , and Uniform denotes that the shifts are drawn from $\mathrm { U } ( - \Sigma , \Sigma )$ , where $\Sigma$ is the scope obtained from our uncertainty estimation. As we can see, directly using Gaussian distribution with the improper variant scope will harm the model performances, indicating the variant range of feature statistics should have some instructions. Further analysis about different vanilla Gaussian distributions with pre-defined standard deviation are conducted in the Appendix. Meanwhile, the result of Uniform shows some improvement but is still lower than DSU, which indicates the boundless Gaussian distribution is more helpful to model more diverse variants.
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
Figure 4: The effects on the hyperparameter probability.
|
| 164 |
+
|
| 165 |
+
Table 6: Different choices of distribution for uncertainty.
|
| 166 |
+
|
| 167 |
+
<table><tr><td>Choice</td><td>Baseline</td><td>Random</td><td>Uniform</td><td>DSU</td></tr><tr><td>PACS</td><td>79.0</td><td>76.9</td><td>81.9</td><td>84.1</td></tr><tr><td>GTA5 to Cityscapes</td><td>37.0</td><td>38.2</td><td>41.6</td><td>43.1</td></tr></table>
|
| 168 |
+
|
| 169 |
+
# 6 FURTHER ANALYSIS
|
| 170 |
+
|
| 171 |
+
# 6.1 QUANTITATIVE ANALYSIS ON THE PROPOSED METHOD
|
| 172 |
+
|
| 173 |
+
In this subsection, we will analyze the effects of the proposed method on both intermediate features and feature representations. Quantitative experiments are conducted on PACS, where we choose Art Painting as the unseen testing domain and the rests are used as training domains.
|
| 174 |
+
|
| 175 |
+
To study the phenomena of feature statistic shifts, we capture the intermediate features after the second block in ResNet18 and measure the average feature statistics values of one category in the training and testing domain, respectively. The distributions of feature statistics are shown in Figure 5. As the previous works (Wang et al., 2020b; 2019a) show, the feature statistics extracted from the baseline model show an obvious shift due to different data distribution. It can be seen that the model trained with our method has less shift. Our method can help the model gain robustness towards domain shifts, as it properly models the potential feature statistic shifts.
|
| 176 |
+
|
| 177 |
+
|
| 178 |
+
Figure 5: Quantitative analysis on the shifts of feature statistics (mean and standard deviation) between training source domains and unseen testing domain.
|
| 179 |
+
|
| 180 |
+
# 6.2 VISUALIZATION ON THE SYNTHETIC CHANGES
|
| 181 |
+
|
| 182 |
+
Besides the quantitative experiment results, we also obtain a more intuitional view of the diverse changes provided by our method, through visualizing the reconstruction results using a predefined autoencoder1 (Huang & Belongie (2017)), where the proposed module is inserted into the encoder, and inverse the feature representations into synthetic images after the decoder. As the results shown in Figure 6, the reconstructed images obtained from our probabilistic approach show diverse synthetic changes, such as the environment, object texture, and contrast, etc.
|
| 183 |
+
|
| 184 |
+
|
| 185 |
+
Figure 6: The visualization on diverse synthetic changes obtained from our method.
|
| 186 |
+
|
| 187 |
+
# 7 CONCLUSIONS
|
| 188 |
+
|
| 189 |
+
In this paper, we propose a probabilistic approach to improve the network generalization ability by modeling the uncertainty of domain shifts with synthesized feature statistics during training. Each feature statistic is hypothesized to follow a multi-variate Gaussian distribution for modeling the diverse potential shifts. Due to the generated feature statistics with diverse distribution possibilities, the models can gain better robustness towards diverse domain shifts. Experiment results demonstrate the effectiveness of our method in improving the network generalization ability.
|
| 190 |
+
|
| 191 |
+
# ACKNOWLEDGEMENT
|
| 192 |
+
|
| 193 |
+
This work was supported by the National Natural Science Foundation of China under Grant 62088102, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the $\mathrm { N g }$ Teng Fong Charitable Foundation.
|
| 194 |
+
|
| 195 |
+
# REFERENCES
|
| 196 |
+
|
| 197 |
+
Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151–175, 2010.
|
| 198 |
+
|
| 199 |
+
Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Revisiting batch normalization for improving corruption robustness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 494–503, 2021.
|
| 200 |
+
|
| 201 |
+
Gilles Blanchard, Gyemin Lee, and Clayton Scott. Generalizing from several related classification tasks to a new unlabeled sample. In Advances in Neural Information Processing Systems, 2011.
|
| 202 |
+
|
| 203 |
+
Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5710–5719, 2020.
|
| 204 |
+
|
| 205 |
+
Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In European Conference in Computer Vision (ECCV), 2020.
|
| 206 |
+
|
| 207 |
+
Guangyao Chen, Peixi Peng, Li Ma, Jia Li, Lin Du, and Yonghong Tian. Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021.
|
| 208 |
+
|
| 209 |
+
Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4): 834–848, 2018.
|
| 210 |
+
|
| 211 |
+
Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11580–11590, 2021.
|
| 212 |
+
|
| 213 |
+
Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
|
| 214 |
+
|
| 215 |
+
Yongxing Dai, Xiaotong Li, Jun Liu, Zekun Tong, and Ling-Yu Duan. Generalizable person reidentification with relevance-aware mixture of experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 16145–16154, June 2021.
|
| 216 |
+
|
| 217 |
+
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255, 2009.
|
| 218 |
+
|
| 219 |
+
Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, pp. 1126–1135, 2017.
|
| 220 |
+
|
| 221 |
+
Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of The 33rd International Conference on Machine Learning, volume 48, pp. 1050–1059, 2016.
|
| 222 |
+
|
| 223 |
+
Shang-Hua Gao, Qi Han, Duo Li, Ming-Ming Cheng, and Pai Peng. Representative batch normalization with feature calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8669–8679, 2021a.
|
| 224 |
+
|
| 225 |
+
Shang-Hua Gao, Qi Han, Duo Li, Pai Peng, Ming-Ming Cheng, and Pai Peng. Representative batch normalization with feature calibration. 2021b.
|
| 226 |
+
|
| 227 |
+
Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual mean-teaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020.
|
| 228 |
+
|
| 229 |
+
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
|
| 230 |
+
|
| 231 |
+
Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. International Conference on Learning Representations, 2019.
|
| 232 |
+
|
| 233 |
+
Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple method to improve robustness and uncertainty under data shift. In International Conference on Learning Representations, 2020.
|
| 234 |
+
|
| 235 |
+
X. Huang and S. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 1510–1519, 2017.
|
| 236 |
+
|
| 237 |
+
Zeyi Huang, Haohan Wang, Eric P. Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In The European Conference on Computer Vision (ECCV), pp. 124–140, 2020.
|
| 238 |
+
|
| 239 |
+
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
|
| 240 |
+
|
| 241 |
+
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25:1097–1105, 2012.
|
| 242 |
+
|
| 243 |
+
Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12383–12392, 2021.
|
| 244 |
+
|
| 245 |
+
Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017.
|
| 246 |
+
|
| 247 |
+
Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018a.
|
| 248 |
+
|
| 249 |
+
Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C. Kot. Domain generalization with adversarial feature learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5400–5409, 2018b.
|
| 250 |
+
|
| 251 |
+
Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), 2018c.
|
| 252 |
+
|
| 253 |
+
Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8690–8699, 2021.
|
| 254 |
+
|
| 255 |
+
Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9482–9491, 2021.
|
| 256 |
+
|
| 257 |
+
Xingang Pan, Ping Luo, Jianping Shi, and Xiaoou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), 2018.
|
| 258 |
+
|
| 259 |
+
Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In Proceedings of the 37th International Conference on Machine Learning, volume 119, pp. 7728–7738, 2020.
|
| 260 |
+
|
| 261 |
+
Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European Conference on Computer Vision (ECCV), pp. 102–118, 2016.
|
| 262 |
+
|
| 263 |
+
Ergys Ristani, Francesco Solera, Roger S. Zou, R. Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV Workshops, 2016.
|
| 264 |
+
|
| 265 |
+
Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In The European Conference on Computer Vision (ECCV), pp. 68–83, 2020.
|
| 266 |
+
|
| 267 |
+
Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations, 2018.
|
| 268 |
+
|
| 269 |
+
Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021.
|
| 270 |
+
|
| 271 |
+
Zheyan Shen, Jiashuo Liu, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021.
|
| 272 |
+
|
| 273 |
+
Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019.
|
| 274 |
+
|
| 275 |
+
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929–1958, 2014.
|
| 276 |
+
|
| 277 |
+
Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008.
|
| 278 |
+
|
| 279 |
+
V. Vapnik. Principles of risk minimization for learning theory. In Advances in Neural Information Processing Systems, volume 4, 1992.
|
| 280 |
+
|
| 281 |
+
Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 10(5):988–999, 1999.
|
| 282 |
+
|
| 283 |
+
Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
|
| 284 |
+
|
| 285 |
+
Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David LopezPaz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, volume 97, pp. 6438– 6447, 2019.
|
| 286 |
+
|
| 287 |
+
Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), 2020a.
|
| 288 |
+
|
| 289 |
+
Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Wenjun Zeng, and Tao Qin. Generalizing to unseen domains: A survey on domain generalization. arXiv preprint arXiv:2103.03097, 2021.
|
| 290 |
+
|
| 291 |
+
Ximei Wang, Ying Jin, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Transferable normalization: Towards improving transferability of deep neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019a.
|
| 292 |
+
|
| 293 |
+
Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. In Advances in Neural Information Processing Systems, volume 33, pp. 19212–19223, 2020b.
|
| 294 |
+
|
| 295 |
+
Yulin Wang, Xuran Pan, Shiji Song, Hong Zhang, Gao Huang, and Cheng Wu. Implicit semantic data augmentation for deep networks. In Advances in Neural Information Processing Systems, pp. 12614–12623, 2019b.
|
| 296 |
+
|
| 297 |
+
Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, 2021.
|
| 298 |
+
|
| 299 |
+
Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6022–6031, 2019.
|
| 300 |
+
|
| 301 |
+
Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018.
|
| 302 |
+
|
| 303 |
+
Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1116–1124, 2015.
|
| 304 |
+
|
| 305 |
+
Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In The European Conference on Computer Vision (ECCV), pp. 561–578, 2020a.
|
| 306 |
+
|
| 307 |
+
Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain adaptive ensemble learning. arXiv preprint arXiv:2003.07325, 2020b.
|
| 308 |
+
|
| 309 |
+
Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. arXiv preprint arXiv:2103.02503, 2021a.
|
| 310 |
+
|
| 311 |
+
Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In International Conference on Learning Representations, 2021b.
|
| 312 |
+
|
| 313 |
+
# A APPENDIX
|
| 314 |
+
|
| 315 |
+
# A.1 ALGORITHM
|
| 316 |
+
|
| 317 |
+
The algorithm of the proposed method is illustrated in Algorithm 1.
|
| 318 |
+
|
| 319 |
+
# Algorithm 1: The algorithm of the proposed method (DSU)
|
| 320 |
+
|
| 321 |
+
Input: Intermediate feature $\boldsymbol { x } \in \mathbb { R } ^ { B \times C \times H \times W }$ , probability $p$ to forward this module;
|
| 322 |
+
Output: Intermediate feature $\widehat { x } \in \mathbb { R } ^ { B \times C \times H \times W }$ after considering potential statistics shifts;
|
| 323 |
+
1 Sample $p _ { 0 } \sim U ( 0 , 1 )$ ;
|
| 324 |
+
2 if $p _ { 0 } < p$ and Training then
|
| 325 |
+
3 Compute the channel-wise mean and standard deviation of each instance in a mini-batch;
|
| 326 |
+
4 $\begin{array} { l } { \displaystyle \mu ( x ) = \frac { 1 } { H W } \sum _ { h = 1 } ^ { H } \sum _ { w = 1 } ^ { W } x _ { b , c , h , w } , } \\ { \sigma ^ { 2 } ( x ) = \frac { 1 } { H W } \sum _ { h = 1 } ^ { H } \sum _ { w = 1 } ^ { W } ( x _ { b , c , h , w } - \mu ( x ) ) ^ { 2 } . } \end{array}$
|
| 327 |
+
5
|
| 328 |
+
6 Uncertainty estimation on feature statistics;
|
| 329 |
+
7 $\begin{array} { r l } & { \Sigma _ { \mu } ^ { 2 } ( x ) = \frac { 1 } { B } \displaystyle \sum _ { b = 1 } ^ { B } ( \mu _ { b c } ( x ) - E _ { b } ( \mu _ { b c } ( x ) ) ) ^ { 2 } , } \\ & { \Sigma _ { \sigma } ^ { 2 } ( x ) = \frac { 1 } { B } \displaystyle \sum _ { b = 1 } ^ { B } ( \sigma _ { b c } ( x ) - E _ { b } ( \sigma _ { b c } ( x ) ) ) ^ { 2 } . } \end{array}$
|
| 330 |
+
8
|
| 331 |
+
9 Compute the synthetic feature statistics randomly sampling from the given Guassian distributions;
|
| 332 |
+
10 $\begin{array} { r l } & { \beta ( x ) = \mu ( x ) + \epsilon _ { \mu } \Sigma _ { \sigma } ( x ) , \epsilon _ { \mu } \sim \mathcal { N } ( \mathbf { 0 } , \mathbf { 1 } ) , } \\ & { \gamma ( x ) = \sigma ( x ) + \epsilon _ { \sigma } \Sigma _ { \sigma } ( x ) , \epsilon _ { \sigma } \sim \mathcal { N } ( \mathbf { 0 } , \mathbf { 1 } ) . } \end{array}$ ,
|
| 333 |
+
11
|
| 334 |
+
12 Obtain the feature after considering potential statistics shifts;
|
| 335 |
+
13 $\begin{array} { r } { \widehat { x } = \gamma ( x ) \times \frac { x - \mu ( x ) } { \sigma ( x ) } + \beta ( x ) . } \end{array}$
|
| 336 |
+
14 return the feature $\widehat { x }$ with uncertain feature statistics.
|
| 337 |
+
|
| 338 |
+
# 15 else
|
| 339 |
+
|
| 340 |
+
adopt the original feature $x$ and skip this module.
|
| 341 |
+
|
| 342 |
+
17 end
|
| 343 |
+
|
| 344 |
+
# A.2 MULTI-DOMAIN GENERALIZATION ON OFFICE HOME.
|
| 345 |
+
|
| 346 |
+
In addition to multi-domain classification experiments on PACS, we further evaluate the effectiveness of the proposed method on Office-Home (Venkateswara et al., 2017), which contains 15,500 images of 65 classes for home and office recognition. The experiment results with ResNet18 backbone are shown in Table 7. It can be observed that our method brings obvious improvement over the baseline method and also outperforms the competing methods. By introducing the feature statistics uncertainty, the models trained with our method can learn to alleviate the domain perturbations, such as the style information, and obtain more domain-invariant features. For example, huge improvement can be observed from the results on Clipart, which is a domain with much different style from others.
|
| 347 |
+
|
| 348 |
+
Table 7: Experiment results of Office-Home multi-domain classification task.
|
| 349 |
+
|
| 350 |
+
<table><tr><td>Method</td><td>Reference</td><td>Art</td><td>Clipart</td><td>Product</td><td>Real</td><td>Average (%)</td></tr><tr><td>Baseline</td><td></td><td>58.8</td><td>48.3</td><td>74.2</td><td>76.2</td><td>64.4</td></tr><tr><td>Mixup (Zhang et al.,2018)</td><td>ICLR 2018</td><td>58.2</td><td>49.3</td><td>74.7</td><td>76.1</td><td>64.6</td></tr><tr><td>CrossGrad (Shankar et al., 2018)</td><td>ICLR 2018</td><td>58.4</td><td>49.4</td><td>73.9</td><td>75.8</td><td>64.4</td></tr><tr><td>Manifold Mixup (Verma et al., 2019)</td><td>ICML 2019</td><td>56.2</td><td>46.3</td><td>73.6</td><td>75.2</td><td>62.8</td></tr><tr><td>CutMix (Yun et al., 2019)</td><td>ICCV 2019</td><td>57.9</td><td>48.3</td><td>74.5</td><td>75.6</td><td>64.1</td></tr><tr><td>RSC (Huang et al.,2020)</td><td>ECCV2020</td><td>58.4</td><td>47.9</td><td>71.6</td><td>74.5</td><td>63.1</td></tr><tr><td>L2A-OT (Zhou et al., 2020a)</td><td>ECCV2020</td><td>60.6</td><td>50.1</td><td>74.8</td><td>77.0</td><td>65.6</td></tr><tr><td>MixStyle (Zhou et al.,2021b)</td><td>ICLR 2021</td><td>58.7</td><td>53.4</td><td>74.2</td><td>75.9</td><td>65.5</td></tr><tr><td>DSU</td><td>Ours</td><td>60.2</td><td>54.8</td><td>74.1</td><td>75.1</td><td>66.1</td></tr></table>
|
| 351 |
+
|
| 352 |
+
# A.3 CHOICE OF UNCERTAINTY DISTRIBUTION
|
| 353 |
+
|
| 354 |
+
Table 8: Intensive study about different vanilla Gaussian distributions with pre-defined standard deviation.
|
| 355 |
+
|
| 356 |
+
<table><tr><td>Choice</td><td>Baseline</td><td>Rand(10°)</td><td>Rand(10-1)</td><td>Rand(10-2)</td><td>Rand(10-3)</td><td>DSU</td></tr><tr><td>PACS</td><td>79.0</td><td>76.9</td><td>81.2</td><td>79.3</td><td>79.1</td><td>84.1</td></tr><tr><td>GTA5 to Cityscapes</td><td>37.0</td><td>38.2</td><td>39.8</td><td>40.1</td><td>38.9</td><td>43.1</td></tr></table>
|
| 357 |
+
|
| 358 |
+
Besides the analysis of the uncertainty estimation in the ablation study, we also conduct a more intensive study about the effects of pre-defined uncertainty estimations. Specifically, Rand $( s )$ denotes directly imposing random shifts draw from a fixed Gaussian ${ \mathcal { N } } ( 0 , s ^ { 2 } )$ . The intensive study is shown in Table 8. It can be observed that the results of different fixed distributions are all much lower than the proposed method. Some conclusions could be obtained from the results. (a): Imposing excessive uncertainty might harm the model training and degrade the performance. (b): The best fixed value of uncertainty estimation might vary from different tasks. By contrast, the proposed method can be adaptive to different tasks without any manual adjustment.
|
| 359 |
+
|
| 360 |
+
Table 9: Study about the effects of sharing the same uncertain distribution among different channels.
|
| 361 |
+
|
| 362 |
+
<table><tr><td>Choice</td><td>Baseline</td><td>Channel-share</td><td>DSU</td></tr><tr><td>PACS</td><td>79.0</td><td>80.2</td><td>84.1</td></tr><tr><td>GTA5 to Cityscapes</td><td>37.0</td><td>39.3</td><td>43.1</td></tr></table>
|
| 363 |
+
|
| 364 |
+
We also conduct the experiment to test the effectiveness of treating different channels with different potentials. Channel-share denotes all channels of the sample share the same uncertainty distribution, i.e., using the average uncertainty estimation among channels. As shown in Table 9, the results indicate that sharing the same uncertain distribution among different channels is less effective, which ignores the different potentials of channels and will limit their performances. Meanwhile, the proposed method explicitly considers the different potentials of different channels and brings better performances.
|
| 365 |
+
|
| 366 |
+
# A.4 T-SNE VISUALIZATION
|
| 367 |
+
|
| 368 |
+
To analyze the effects on feature representations, we visualize the feature representation vectors of different categories in unseen domain with t-SNE (Van der Maaten & Hinton, 2008) in Figure 7. The features of the same category become more compact benefiting from the proposed method. Because our method can alleviate the domain perturbations during training and make the model focus on content information, obtaining more invariant features representations.
|
| 369 |
+
|
| 370 |
+
|
| 371 |
+
Figure 7: The t-SNE visualization on unseen PACS domain.
|
| 372 |
+
|
| 373 |
+
# A.5 COMPARISONS TO THE RELATED METHODS
|
| 374 |
+
|
| 375 |
+
Some related methods (Zhou et al., 2021b; Nuriel et al., 2021) also tackle the domain generalization problem by producing synthetic feature statistics. Specifically, we denote the random shuffle copies of the batch feature as $\widehat { \boldsymbol { x } } = \mathrm { s h u f f l e } ( \boldsymbol { x } )$ . pAdaIN (Nuriel et al. (2021)) generate new samples by bswapping feature statistics between the batch samples applied with a random permutation, where $\beta ( x ) { \overset { \vartriangle } { = } } \mu ( { \widehat { x } } )$ and $\gamma ( x ) ~ = ~ \sigma ( { \widehat { x } } )$ . MixStyle (Zhou et al. (2021b)) generates synthesized domain b bsamples by mixing feature statistics information of two instances, where $\beta ( x ) \dot { } = \lambda \mu ( x ) + ( 1 -$ $\lambda ) \mu ( \widehat { x } )$ and $\gamma ( x ) \overset { - } { = } \lambda \sigma ( x ) + ( 1 - \lambda ) \sigma ( \widehat { x } )$ and $\lambda \in ( 0 , 1 )$ is a random interpolation weight.
|
| 376 |
+
|
| 377 |
+
Despite the success, they typically adopt linear manipulation on pairwise samples to generate new feature statistics, which limits the diversity of synthetic changes. Specifically, the direction of the variants is determined by the chosen reference sample and the internal operation also restricts the variant intensity. Our method, not relies on a specific reference sample, is based on the Gaussian distribution that can produce not only linear changes but diverse variants with more possibilities. Due to the boundless range of the Gaussian distribution, our method has the ability to generate feature statistics beyond the scope of training domain, which also breaks the limitation of inner interpolation between training samples. The visualization of the comparisons is shown in Figure 8.
|
| 378 |
+
|
| 379 |
+
|
| 380 |
+
Figure 8: Comparisons with related methods. The variants produced by previous pairwise-based methods are restricted by the combination of chosen sample pair, while our method can generate various feature statistics variants with different combination of directions and intensities.
|
| 381 |
+
|
| 382 |
+
# A.6 WITHIN-DATASET PERFORMANCE
|
| 383 |
+
|
| 384 |
+
In Table 4, we tested the within-dataset performance on the large-scale dataset ImageNet, denoted as ”Clean”. We observed that the top-1 error rate declines from $2 3 . 8 \%$ to $2 3 . 4 \%$ after training with the proposed DSU, indicating that DSU does not sacrifice the in-domain performance to gain the benefits on out-of-distribution domains.The reason might be that instances in the testing set may not always fall into exactly the same distribution of the training set, and they still have slight statistic shifts (Gao et al., 2021a). The proposed DSU can help the trained model improve the robustness to statistics shifts and thus gain better performance in within-dataset ImageNet.
|
| 385 |
+
|
| 386 |
+
Besides the experiments on ImageNet, we also supplement the within-dataset performances on PACS. According to the multi-source training protocol on PACS (Li et al., 2017), the within-domain performance is averaged over multiple training-domain datasets (P,A,C,S denotes Art, Cartoon, Photo and Sketch respectively). As shown in Table 10, it can be observed that the within-dataset performance of our DSU also slightly beats the baseline, verifying the conclusion as on ImageNet.
|
| 387 |
+
|
| 388 |
+
Table 10: Within-dataset performance on PACS. P,A,C,S denote Photo, Art Painting, Cartoon, and Sketch respectively.
|
| 389 |
+
|
| 390 |
+
<table><tr><td>Method</td><td>Reference</td><td>P,C,S</td><td>P,A,S</td><td>A,C,S</td><td>P,A,C</td><td>Average (%)</td></tr><tr><td>Baseline</td><td>1</td><td>95.70</td><td>95.28</td><td>94.22</td><td>96.58</td><td>95.44</td></tr><tr><td>DSU</td><td>Ours</td><td>96.20</td><td>96.32</td><td>95.17</td><td>97.20</td><td>96.21</td></tr></table>
|
| 391 |
+
|
| 392 |
+
# A.7 ABLATION STUDY ON BATCH SIZE
|
| 393 |
+
|
| 394 |
+
In Table 11, we conduct an ablation study on the batch size. As shown in the table, consistent performance gains are observed with various batch sizes on PACS. Note we use the batch size of 64 in our paper, following the original setting in PACS (Li et al., 2017) for fair comparison.
|
| 395 |
+
|
| 396 |
+
Table 11: Ablation study on the effects of batch size.
|
| 397 |
+
|
| 398 |
+
<table><tr><td>batchsize</td><td>Reference</td><td>16</td><td>32</td><td>64</td><td>128</td><td>256</td></tr><tr><td>Baseline</td><td>1</td><td>81.0</td><td>80.2</td><td>79.0</td><td>77.8</td><td>75.6</td></tr><tr><td>DSU</td><td>Ours</td><td>84.9 (+3.9)</td><td>84.5 (+4.3)</td><td>84.1 (+5.1)</td><td>82.1 (+4.3)</td><td>80.4 (+4.8)</td></tr></table>
|
parse/dev/6HN7LHyzGgC/6HN7LHyzGgC_content_list.json
ADDED
|
@@ -0,0 +1,2043 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "UNCERTAINTY MODELING FOR OUT-OF-DISTRIBUTION GENERALIZATION ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
176,
|
| 8 |
+
98,
|
| 9 |
+
740,
|
| 10 |
+
146
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Xiaotong $\\mathbf { L i } ^ { 1 }$ , Yongxing $\\mathbf { D a i } ^ { 1 }$ , Yixiao $\\mathbf { G e ^ { 2 } }$ , $\\mathbf { J u n L i u ^ { 3 } }$ , Ying Shan2, Ling-Yu Duan1,4∗ \n1Peking University, Beijing, China 2ARC Lab, Tencent PCG \n3Singapore University of Technology and Design, Singapore \n4Peng Cheng Laboratory, Shenzhen, China \nlixiaotong@stu.pku.edu.cn, {yongxingdai, lingyu}@pku.edu.cn, \n{yixiaoge, yingsshan}@tencent.com, jun liu@sutd.edu.sg ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
184,
|
| 19 |
+
167,
|
| 20 |
+
756,
|
| 21 |
+
256
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "ABSTRACT ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
454,
|
| 31 |
+
291,
|
| 32 |
+
544,
|
| 33 |
+
306
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "Though remarkable progress has been achieved in various vision tasks, deep neural networks still suffer obvious performance degradation when tested in out-ofdistribution scenarios. We argue that the feature statistics (mean and standard deviation), which carry the domain characteristics of the training data, can be properly manipulated to improve the generalization ability of deep learning models. Common methods often consider the feature statistics as deterministic values measured from the learned features and do not explicitly consider the uncertain statistics discrepancy caused by potential domain shifts during testing. In this paper, we improve the network generalization ability by modeling the uncertainty of domain shifts with synthesized feature statistics during training. Specifically, we hypothesize that the feature statistic, after considering the potential uncertainties, follows a multivariate Gaussian distribution. Hence, each feature statistic is no longer a deterministic value, but a probabilistic point with diverse distribution possibilities. With the uncertain feature statistics, the models can be trained to alleviate the domain perturbations and achieve better robustness against potential domain shifts. Our method can be readily integrated into networks without additional parameters. Extensive experiments demonstrate that our proposed method consistently improves the network generalization ability on multiple vision tasks, including image classification, semantic segmentation, and instance retrieval. The code can be available at https://github.com/lixiaotong97/DSU. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
233,
|
| 42 |
+
323,
|
| 43 |
+
764,
|
| 44 |
+
601
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 INTRODUCTION ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
176,
|
| 54 |
+
627,
|
| 55 |
+
336,
|
| 56 |
+
642
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "Deep neural networks have shown impressive success in computer vision, but with a severe reliance on the assumption that the training and testing domains follow an independent and identical distribution (Ben-David et al., 2010; Vapnik, 1992). This assumption, however, does not hold in many real-world applications. For instance, when employing segmentation models trained on sunny days for rainy and foggy environments (Choi et al., 2021), or recognizing art paintings with models that trained on photographs (Li et al., 2017), inevitable performance drop can often be observed in such out-of-distribution deployment scenarios. Therefore, the problem of domain generalization, aiming to improve the robustness of the network on various unseen testing domains, becomes quite important. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
659,
|
| 66 |
+
825,
|
| 67 |
+
782
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Previous works (Huang & Belongie, 2017; Li et al., 2021) demonstrate that feature statistics (mean and standard deviation), as the moments of the learned features, carry informative domain characteristics of the training data. Domain characteristics primarily refer to the information that is more specific to the individual domains but less relevant to the task objectives, such as the photo style and capturing environment information in object recognition. Consequently, domains with different data distributions generally have inconsistent feature statistics (Wang et al., 2020b; 2019a; Gao et al., 2021a). Most deep learning methods follow Empirical Risk Minimization principle (Vapnik, 1999) to minimize their average error over the training data (Shen et al., 2021). Despite the satisfactory performance on the training domain, these methods do not explicitly consider the uncertain statistics discrepancy caused by potential domain shifts during testing. As a result, the trained models tend to overfit the training domain and show vulnerability to the statistic changes at testing time, substantially limiting the generalization ability of the learned representations. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
790,
|
| 77 |
+
825,
|
| 78 |
+
901
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "image",
|
| 84 |
+
"img_path": "images/3ee5331b3cecd74e6b2fd4294be35facd9a56932b18b4f4ead23f76c4ad075da.jpg",
|
| 85 |
+
"image_caption": [
|
| 86 |
+
"Figure 1: The visualization of reconstructed samples with synthesized feature statistics, using a pre-trained style transfer auto-encoder (Huang & Belongie, 2017). The illustration of the feature statistics shifts, which may vary in both intensity and direction (i.e., different offsets in the vector space of feature statistics). We also show images of “new” domains generated by manipulating feature statistic shifts with different direction and intensity. Note these images are for visualization only, rather than feeding into the network for training. "
|
| 87 |
+
],
|
| 88 |
+
"image_footnote": [],
|
| 89 |
+
"bbox": [
|
| 90 |
+
269,
|
| 91 |
+
99,
|
| 92 |
+
707,
|
| 93 |
+
271
|
| 94 |
+
],
|
| 95 |
+
"page_idx": 1
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"type": "text",
|
| 99 |
+
"text": "",
|
| 100 |
+
"bbox": [
|
| 101 |
+
174,
|
| 102 |
+
377,
|
| 103 |
+
823,
|
| 104 |
+
434
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "Intuitively, the test domains may bring uncertain statistics shifts with different potential directions and intensities compared to the training domain (as shown in Figure 1), implying the uncertain nature of domain shifts. Considering such “uncertainty” of potential domain shifts, synthesizing novel feature statistics variants to model diverse domain shifts can improve the robustness of the trained network to different testing distributions. Towards this end, we introduce a novel probabilistic method to improve the network generalization ability by properly modeling Domain Shifts with Uncertainty (DSU), i.e., characterizing the feature statistics as uncertain distributions. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
440,
|
| 114 |
+
825,
|
| 115 |
+
539
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "In our method, instead of treating each feature statistic as a deterministic point measured from the feature, we hypothesize that the feature statistic, after considering potential uncertainties, follows a multi-variate Gaussian distribution. The distribution “center” is set as each feature’s original statistic value, and the distribution “scope” represents the variant intensity considering underlying domain shifts. Uncertainty estimation is adopted here to depict the distribution “scope” of probabilistic feature statistics. Specifically, we estimate the distribution “scope” based on the variances of the mini-batch statistics in an efficient non-parametric manner. Subsequently, feature statistics variants are randomly sampled from the estimated Gaussian distribution and then used to replace the original deterministic values for modeling diverse domain shifts, as illustrated in Figure 2. Due to the generated feature statistics with diverse distribution possibilities, the models can be trained to properly alleviate the domain perturbations and encode better domain-invariant features. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
545,
|
| 125 |
+
825,
|
| 126 |
+
699
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "Our proposed method is simple yet fairly effective to alleviate performance drop caused by domain shifts, and can be readily integrated into existing networks without bringing additional model parameters or loss constraints. Comprehensive experiments on a wide range of vision tasks demonstrate the superiority of our proposed method, indicating that introducing uncertainty to feature statistics can well improve models’ generalization against domain shifts. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
705,
|
| 136 |
+
825,
|
| 137 |
+
775
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "2 RELATED WORK ",
|
| 144 |
+
"text_level": 1,
|
| 145 |
+
"bbox": [
|
| 146 |
+
176,
|
| 147 |
+
796,
|
| 148 |
+
344,
|
| 149 |
+
811
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 1
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "2.1 DOMAIN GENERALIZATION ",
|
| 156 |
+
"text_level": 1,
|
| 157 |
+
"bbox": [
|
| 158 |
+
176,
|
| 159 |
+
828,
|
| 160 |
+
405,
|
| 161 |
+
842
|
| 162 |
+
],
|
| 163 |
+
"page_idx": 1
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"type": "text",
|
| 167 |
+
"text": "Domain generalization (DG) has been attracting increasing attention in the past few years, which aims to achieve out-of-distribution generalization on unseen target domains using only single or multiple source domain data for training (Blanchard et al. (2011)). Research on addressing this problem has been extensively conducted in the literature (Zhou et al. (2021a); Wang et al. (2021); Shen et al. (2021)). Here some studies that are more related to our work are introduced below. ",
|
| 168 |
+
"bbox": [
|
| 169 |
+
174,
|
| 170 |
+
854,
|
| 171 |
+
823,
|
| 172 |
+
922
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 1
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "Data Augmentation: Data augmentation is an effective manner for improving generalization ability and relieving models from overfitting in training domains. Most augmentation methods adopt various transformations at the image level, such as AugMix (Hendrycks et al. (2020)) and CutMix (Yun et al. (2019)). Besides using handcraft transformations, mixup (Zhang et al. (2018)) trains the model by using pair-wise linearly interpolated samples in both the image and label spaces. Manifold Mixup (Verma et al. (2019)) further adopts this linear interpolation from image level to feature level. Some recent works extend the above transformations to feature statistics for improving model generalization. MixStyle (Zhou et al. (2021b)) adopts linear interpolation on feature statistics of two instances to generate synthesized samples. The pAdaIn (Nuriel et al. (2021)) swaps statistics between the samples applied with a random permutation of the batch. ",
|
| 179 |
+
"bbox": [
|
| 180 |
+
173,
|
| 181 |
+
103,
|
| 182 |
+
825,
|
| 183 |
+
242
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 2
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "Invariant Representation Learning: The main idea of invariant representation learning is to enable models to learn features that are invariant to domain shifts. Domain alignment-based approaches (Li et al. (2018c;b)) learn invariant features by minimizing the distances between different distributions. Instead of enforcing the entire features to be invariant, disentangled feature learning approaches (Chattopadhyay et al. (2020); Piratla et al. (2020)) decouple the features into domain-specific and domain-invariant parts and learn their representations simultaneously. In addition, normalizationbased methods (Pan et al. (2018); Choi et al. (2021)) can also be used to remove the style information to obtain invariant representations. ",
|
| 190 |
+
"bbox": [
|
| 191 |
+
173,
|
| 192 |
+
250,
|
| 193 |
+
825,
|
| 194 |
+
361
|
| 195 |
+
],
|
| 196 |
+
"page_idx": 2
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"type": "text",
|
| 200 |
+
"text": "Learning Strategies: There are also some effective learning strategies that can be leveraged to improve generalization ability. Ensemble learning is an effective technique in boosting model performance. The ensemble predictions using a collection of diverse models (Zhou et al. (2020b)) or modules (Seo et al. (2020)) can be adopted to improve generalization and robustness. Metalearning-based methods (Finn et al. (2017); Li et al. (2018a); Dai et al. (2021)) learn to simulate the domain shifts following an episode training paradigm. Besides, self-challenging methods, such as RSC (Huang et al. (2020)), force the model to learn a general representation by discarding dominant features activated on the training data. ",
|
| 201 |
+
"bbox": [
|
| 202 |
+
174,
|
| 203 |
+
367,
|
| 204 |
+
825,
|
| 205 |
+
479
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "2.2 UNCERTAINTY IN DEEP LEARNING ",
|
| 212 |
+
"text_level": 1,
|
| 213 |
+
"bbox": [
|
| 214 |
+
178,
|
| 215 |
+
494,
|
| 216 |
+
457,
|
| 217 |
+
508
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "Uncertainty capturing the “noise” and “randomness” inherent in the data has received increasing attention in deep representation learning. Variational Auto-encoder (Kingma & Welling (2013)), as an important method for learning generative models, can be regarded as a method to model the data uncertainty in the hidden space. Dropout (Srivastava et al. (2014)), which is widely used in many deep learning models to avoid over-fitting, can be interpreted to represent model uncertainty as a Bayesian approximation (Gal & Ghahramani (2016)). In some works, uncertainty is used to address the issues of low-quality training data. In person re-identification, DistributionNet (Gal & Ghahramani (2016)) adopts uncertainty to model the person images of noise-labels and outliers. In face recognition, DUL (Chang et al. (2020)) and PFE ((Shi & Jain, 2019)) apply data uncertainty to simultaneously learn the feature embedding and its uncertainty, where the uncertainty is learned through a learnable subnetwork to describe the quality of the image. Different from the aforementioned works, our proposed method is used to model the feature statistics uncertainty under potential domain shifts and acts as a feature augmentation method for handling our-of-distribution generalization problem. ",
|
| 224 |
+
"bbox": [
|
| 225 |
+
173,
|
| 226 |
+
520,
|
| 227 |
+
825,
|
| 228 |
+
715
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "3 METHOD ",
|
| 235 |
+
"text_level": 1,
|
| 236 |
+
"bbox": [
|
| 237 |
+
174,
|
| 238 |
+
734,
|
| 239 |
+
282,
|
| 240 |
+
751
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 2
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "text",
|
| 246 |
+
"text": "3.1 PRELIMINARIES ",
|
| 247 |
+
"text_level": 1,
|
| 248 |
+
"bbox": [
|
| 249 |
+
174,
|
| 250 |
+
765,
|
| 251 |
+
326,
|
| 252 |
+
779
|
| 253 |
+
],
|
| 254 |
+
"page_idx": 2
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"type": "text",
|
| 258 |
+
"text": "Given $x \\in \\mathbb { R } ^ { B \\times C \\times H \\times W }$ to be the encoded features in the intermediate layers of the network, we denote $\\mu \\in \\mathbb { R } ^ { B \\times C }$ and $\\sigma \\in \\mathbb { R } ^ { B \\times C }$ as the channel-wise feature mean and standard deviation of each instance in a mini-batch, respectively, which can be formulated as: ",
|
| 259 |
+
"bbox": [
|
| 260 |
+
176,
|
| 261 |
+
790,
|
| 262 |
+
823,
|
| 263 |
+
833
|
| 264 |
+
],
|
| 265 |
+
"page_idx": 2
|
| 266 |
+
},
|
| 267 |
+
{
|
| 268 |
+
"type": "equation",
|
| 269 |
+
"img_path": "images/2ebe4b4fe7ca409823cd841b768f52ea9f8c56c10da50ed4b61e6ccf5f3c8c01.jpg",
|
| 270 |
+
"text": "$$\n\\begin{array} { l } { \\displaystyle \\mu ( x ) = \\frac { 1 } { H W } \\sum _ { h = 1 } ^ { H } \\sum _ { w = 1 } ^ { W } { x _ { b , c , h , w } } , } \\\\ { \\displaystyle \\sigma ^ { 2 } ( x ) = \\frac { 1 } { H W } \\sum _ { h = 1 } ^ { H } \\sum _ { w = 1 } ^ { W } ( { x _ { b , c , h , w } - \\mu ( x ) } ) ^ { 2 } . } \\end{array}\n$$",
|
| 271 |
+
"text_format": "latex",
|
| 272 |
+
"bbox": [
|
| 273 |
+
356,
|
| 274 |
+
833,
|
| 275 |
+
642,
|
| 276 |
+
924
|
| 277 |
+
],
|
| 278 |
+
"page_idx": 2
|
| 279 |
+
},
|
| 280 |
+
{
|
| 281 |
+
"type": "image",
|
| 282 |
+
"img_path": "images/a734dfb70517bdf0df5cb1fcbc19320e05f63417cdf16be43d1aeebcb31fcc84.jpg",
|
| 283 |
+
"image_caption": [
|
| 284 |
+
"Figure 2: Illustration of the proposed method. Feature statistic is assumed to follow a multi-variate Gaussian distribution during training. When passed through this module, the new feature statistics randomly drawn from the corresponding distribution will replace the original ones to model the diverse domain shifts. "
|
| 285 |
+
],
|
| 286 |
+
"image_footnote": [],
|
| 287 |
+
"bbox": [
|
| 288 |
+
269,
|
| 289 |
+
99,
|
| 290 |
+
727,
|
| 291 |
+
280
|
| 292 |
+
],
|
| 293 |
+
"page_idx": 3
|
| 294 |
+
},
|
| 295 |
+
{
|
| 296 |
+
"type": "text",
|
| 297 |
+
"text": "As the abstraction of features, feature statistics can capture informative characteristics of the corresponding domain (such as color, texture, and contrast), according to previous works (Huang & Belongie, 2017; Li et al., 2021). In out-of-distribution scenarios, the feature statistics often show inconsistency with training domain due to different domain characteristics (Wang et al., $2 0 1 9 \\mathrm { a }$ ; Gao et al., 2021a), which is ill-suited to deep learning modules like nonlinearity layers and normalization layer and degenerates the model’s generalization ability (Wang et al., 2020b). However, most of the deep learning methods only treat feature statistics as deterministic values measured from the features while lacking explicit consideration of the potential uncertain statistical discrepancy. Owing to the model’s inherent vulnerability to such discrepancy, the generalization ability of the learned representations is limited. Some recent methods (Nuriel et al., 2021; Zhou et al., 2021b) utilize feature statistics to tackle the domain generalization problem. Despite the success, they typically adopt linear manipulation (i.e., exchange and interpolation) on pairwise samples to generate new feature statistics, which limits the diversity of synthetic changes. Specifically, the direction of their variants is determined by the chosen reference sample and such internal operation restricts their variant intensity. Thus these methods are sub-optimal when handling the diverse and uncertain domain shifts in real world. ",
|
| 298 |
+
"bbox": [
|
| 299 |
+
174,
|
| 300 |
+
397,
|
| 301 |
+
825,
|
| 302 |
+
619
|
| 303 |
+
],
|
| 304 |
+
"page_idx": 3
|
| 305 |
+
},
|
| 306 |
+
{
|
| 307 |
+
"type": "text",
|
| 308 |
+
"text": "3.2 MODELING DOMAIN SHIFTS WITH UNCERTAINTY ",
|
| 309 |
+
"text_level": 1,
|
| 310 |
+
"bbox": [
|
| 311 |
+
173,
|
| 312 |
+
667,
|
| 313 |
+
555,
|
| 314 |
+
681
|
| 315 |
+
],
|
| 316 |
+
"page_idx": 3
|
| 317 |
+
},
|
| 318 |
+
{
|
| 319 |
+
"type": "text",
|
| 320 |
+
"text": "Given the arbitrary testing domains with uncertain feature statistic shifts in both direction and intensity, properly modeling the domain shifts becomes an important task for tackling the challenge of domain generalization problem. ",
|
| 321 |
+
"bbox": [
|
| 322 |
+
176,
|
| 323 |
+
705,
|
| 324 |
+
823,
|
| 325 |
+
747
|
| 326 |
+
],
|
| 327 |
+
"page_idx": 3
|
| 328 |
+
},
|
| 329 |
+
{
|
| 330 |
+
"type": "text",
|
| 331 |
+
"text": "Considering the uncertainty and randomness of domain shifts, it is promising to employ the methods of “uncertainty” to treat the “uncertainty” of domain shifts. In this paper, we propose a novel method by modeling Domain Shifts with Uncertainty (DSU). Instead of treating each feature statistic as a deterministic value measured from the learned feature, we hypothesize that the distribution of each feature statistic, after considering potential uncertainties, follows a multi-variate Gaussian distribution. This means each feature statistic has a probabilistic representation drawn from a certain distribution, i.e., the feature statistics mean and standard deviation follow $\\mathcal { N } ( \\mu , \\Sigma _ { \\mu } ^ { 2 } )$ and $\\textstyle \\mathcal { N } ( \\sigma , \\Sigma _ { \\sigma } ^ { 2 } )$ , respectively. Specifically, the corresponding Gaussian distribution’s center is set as each feature’s original statistics, while the Gaussian distribution’s standard deviation describes the uncertainty scope for different potential shifts. Through randomly sampling diverse synthesized feature statistics with the probabilistic approach, the models can be trained to improve the robustness of the network against statistics shifts. ",
|
| 332 |
+
"bbox": [
|
| 333 |
+
173,
|
| 334 |
+
755,
|
| 335 |
+
825,
|
| 336 |
+
922
|
| 337 |
+
],
|
| 338 |
+
"page_idx": 3
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"type": "text",
|
| 342 |
+
"text": "3.2.1 UNCERTAINTY ESTIMATION ",
|
| 343 |
+
"text_level": 1,
|
| 344 |
+
"bbox": [
|
| 345 |
+
176,
|
| 346 |
+
103,
|
| 347 |
+
423,
|
| 348 |
+
118
|
| 349 |
+
],
|
| 350 |
+
"page_idx": 4
|
| 351 |
+
},
|
| 352 |
+
{
|
| 353 |
+
"type": "text",
|
| 354 |
+
"text": "Taking the uncertainty of domain shifts into consideration, the uncertainty estimation in our method aims to depict the uncertainty scope of each probabilistic feature statistic. However, the testing domain is unknown, which makes it challenging to obtain an appropriate variant range. ",
|
| 355 |
+
"bbox": [
|
| 356 |
+
176,
|
| 357 |
+
127,
|
| 358 |
+
823,
|
| 359 |
+
170
|
| 360 |
+
],
|
| 361 |
+
"page_idx": 4
|
| 362 |
+
},
|
| 363 |
+
{
|
| 364 |
+
"type": "text",
|
| 365 |
+
"text": "Some generative-based studies (Shen & Zhou, 2021; Wang et al., 2019b) show that the variances between features contain implicit semantic meaning and the directions with larger variances can imply potentials of more valuable semantic changes. Inspired by this, we propose a simple yet effective non-parametric method for uncertainty estimation, utilizing the variance of the feature statistics to provide some instructions: ",
|
| 366 |
+
"bbox": [
|
| 367 |
+
173,
|
| 368 |
+
175,
|
| 369 |
+
825,
|
| 370 |
+
246
|
| 371 |
+
],
|
| 372 |
+
"page_idx": 4
|
| 373 |
+
},
|
| 374 |
+
{
|
| 375 |
+
"type": "equation",
|
| 376 |
+
"img_path": "images/6c56111f8f655c257f670f2e365132aee14c63e219834f662047b282e6553ca7.jpg",
|
| 377 |
+
"text": "$$\n\\begin{array} { l } { { \\Sigma _ { \\mu } ^ { 2 } ( x ) = \\displaystyle \\frac { 1 } { B } \\sum _ { b = 1 } ^ { B } ( \\mu ( x ) - \\mathbb { E } _ { b } [ \\mu ( x ) ] ) ^ { 2 } , } } \\\\ { { \\Sigma _ { \\sigma } ^ { 2 } ( x ) = \\displaystyle \\frac { 1 } { B } \\sum _ { b = 1 } ^ { B } ( \\sigma ( x ) - \\mathbb { E } _ { b } [ \\sigma ( x ) ] ) ^ { 2 } . } } \\end{array}\n$$",
|
| 378 |
+
"text_format": "latex",
|
| 379 |
+
"bbox": [
|
| 380 |
+
375,
|
| 381 |
+
253,
|
| 382 |
+
622,
|
| 383 |
+
349
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 4
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "where $\\Sigma _ { \\mu } \\in \\mathbb { R } ^ { C }$ and $\\Sigma _ { \\sigma } \\in \\mathbb { R } ^ { C }$ represent the uncertainty estimation of the feature mean $\\mu$ and feature standard deviation $\\sigma$ , respectively. The magnitudes of uncertainty estimation can reveal the possibility that the corresponding channel may change potentially. Although the underlying distribution of the domain shifts is unpredictable, the uncertainty estimation captured from the minibatch can provide an appropriate and meaningful variation range for each feature channel, which does not harm model training but can simulate diverse potential shifts. ",
|
| 390 |
+
"bbox": [
|
| 391 |
+
173,
|
| 392 |
+
358,
|
| 393 |
+
825,
|
| 394 |
+
444
|
| 395 |
+
],
|
| 396 |
+
"page_idx": 4
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"type": "text",
|
| 400 |
+
"text": "3.2.2 PROBABILISTIC DISTRIBUTION OF FEATURE STATISTICS ",
|
| 401 |
+
"text_level": 1,
|
| 402 |
+
"bbox": [
|
| 403 |
+
174,
|
| 404 |
+
459,
|
| 405 |
+
612,
|
| 406 |
+
474
|
| 407 |
+
],
|
| 408 |
+
"page_idx": 4
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"type": "text",
|
| 412 |
+
"text": "Once the uncertainty estimation of each feature channel is obtained, the Gaussian distribution for probabilistic feature statistics can be established. To use randomness to model the uncertainty, we adopt the random sampling to further exploit the uncertainty in the probabilistic representations. The new feature statistics, mean $\\beta ( x ) \\sim { \\mathrm { \\bar { \\mathcal { N } } } } ( \\mu , \\Sigma _ { \\mu } ^ { 2 } )$ and standard deviation $\\gamma ( x ) \\sim \\bar { \\mathcal { N } } ( \\sigma , \\Sigma _ { \\sigma } ^ { 2 } )$ , can be randomly drawn from the corresponding distributions as: ",
|
| 413 |
+
"bbox": [
|
| 414 |
+
173,
|
| 415 |
+
483,
|
| 416 |
+
825,
|
| 417 |
+
554
|
| 418 |
+
],
|
| 419 |
+
"page_idx": 4
|
| 420 |
+
},
|
| 421 |
+
{
|
| 422 |
+
"type": "equation",
|
| 423 |
+
"img_path": "images/730a594ca451f8525c65d51f912c67889846ccf52676be2f6d15a8a63f94a0bf.jpg",
|
| 424 |
+
"text": "$$\n\\begin{array} { l l } { { \\beta ( x ) = \\mu ( x ) + \\epsilon _ { \\mu } \\Sigma _ { \\mu } ( x ) , } } & { { \\epsilon _ { \\mu } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\mathbf { 1 } ) , } } \\\\ { { } } & { { } } \\\\ { { \\gamma ( x ) = \\sigma ( x ) + \\epsilon _ { \\sigma } \\Sigma _ { \\sigma } ( x ) , } } & { { \\epsilon _ { \\sigma } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\mathbf { 1 } ) . } } \\end{array}\n$$",
|
| 425 |
+
"text_format": "latex",
|
| 426 |
+
"bbox": [
|
| 427 |
+
349,
|
| 428 |
+
571,
|
| 429 |
+
648,
|
| 430 |
+
626
|
| 431 |
+
],
|
| 432 |
+
"page_idx": 4
|
| 433 |
+
},
|
| 434 |
+
{
|
| 435 |
+
"type": "text",
|
| 436 |
+
"text": "Here we use the re-parameterization trick (Kingma & Welling (2013)) to make the sampling operation differentiable, and $\\epsilon _ { \\mu }$ and $\\epsilon _ { \\sigma }$ both follow the standard Gaussian distribution. By exploiting the given Gaussian distribution, random sampling can generate various new feature statistics information with different combinations of directions and intensities. ",
|
| 437 |
+
"bbox": [
|
| 438 |
+
174,
|
| 439 |
+
633,
|
| 440 |
+
823,
|
| 441 |
+
690
|
| 442 |
+
],
|
| 443 |
+
"page_idx": 4
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"type": "text",
|
| 447 |
+
"text": "3.2.3 IMPLEMENTATION ",
|
| 448 |
+
"text_level": 1,
|
| 449 |
+
"bbox": [
|
| 450 |
+
174,
|
| 451 |
+
705,
|
| 452 |
+
354,
|
| 453 |
+
719
|
| 454 |
+
],
|
| 455 |
+
"page_idx": 4
|
| 456 |
+
},
|
| 457 |
+
{
|
| 458 |
+
"type": "text",
|
| 459 |
+
"text": "The implementation of our method is by the means of AdaIN (Huang & Belongie (2017)), and replaces the feature statistics with the randomly drawing ones to achieve the transformation. The final form of the proposed method can be formulated as: ",
|
| 460 |
+
"bbox": [
|
| 461 |
+
174,
|
| 462 |
+
729,
|
| 463 |
+
825,
|
| 464 |
+
772
|
| 465 |
+
],
|
| 466 |
+
"page_idx": 4
|
| 467 |
+
},
|
| 468 |
+
{
|
| 469 |
+
"type": "equation",
|
| 470 |
+
"img_path": "images/c7ada7835c9874b2c19815c9129ba18ecec3d488cec0fd4e01dca52376b7b79e.jpg",
|
| 471 |
+
"text": "$$\n\\begin{array} { r } { \\mathrm { D S U } ( x ) = \\underbrace { \\left( \\sigma ( x ) + \\epsilon _ { \\sigma } \\Sigma _ { \\sigma } ( x ) \\right) } _ { \\gamma ( x ) } \\left( \\frac { x - \\mu ( x ) } { \\sigma ( x ) } \\right) + \\underbrace { \\left( \\mu ( x ) + \\epsilon _ { \\mu } \\Sigma _ { \\mu } ( x ) \\right) } _ { \\beta ( x ) } . } \\end{array}\n$$",
|
| 472 |
+
"text_format": "latex",
|
| 473 |
+
"bbox": [
|
| 474 |
+
276,
|
| 475 |
+
779,
|
| 476 |
+
722,
|
| 477 |
+
827
|
| 478 |
+
],
|
| 479 |
+
"page_idx": 4
|
| 480 |
+
},
|
| 481 |
+
{
|
| 482 |
+
"type": "text",
|
| 483 |
+
"text": "The above operation can be integrated at various positions of the network as a flexible module. Note that the module only works during model training and can be discarded while testing. To trade off the strength of this module, we set a hyperparameter $p$ that denotes the probability to apply it. The algorithm is described in the Appendix. Benefiting from the proposed method, the model trained with uncertain feature statistics will gain better robustness against potential statistics shifts, and thus acquires a better generalization ability. ",
|
| 484 |
+
"bbox": [
|
| 485 |
+
173,
|
| 486 |
+
839,
|
| 487 |
+
826,
|
| 488 |
+
924
|
| 489 |
+
],
|
| 490 |
+
"page_idx": 4
|
| 491 |
+
},
|
| 492 |
+
{
|
| 493 |
+
"type": "text",
|
| 494 |
+
"text": "4 EXPERIMENTS ",
|
| 495 |
+
"text_level": 1,
|
| 496 |
+
"bbox": [
|
| 497 |
+
174,
|
| 498 |
+
102,
|
| 499 |
+
326,
|
| 500 |
+
118
|
| 501 |
+
],
|
| 502 |
+
"page_idx": 5
|
| 503 |
+
},
|
| 504 |
+
{
|
| 505 |
+
"type": "text",
|
| 506 |
+
"text": "In order to verify the effectiveness of the proposed method in improving the generalization ability of networks, we conduct the experiments on a wide range of tasks, including image classification, semantic segmentation, instance retrieval, and robustness towards corruptions, where the training and testing sets have different cases of distribution shifts, such as style shift, synthetic-to-real gap, scenes change, and pixel-level corruption. ",
|
| 507 |
+
"bbox": [
|
| 508 |
+
174,
|
| 509 |
+
132,
|
| 510 |
+
825,
|
| 511 |
+
202
|
| 512 |
+
],
|
| 513 |
+
"page_idx": 5
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"type": "text",
|
| 517 |
+
"text": "4.1 GENERALIZATION ON MULTI-DOMAIN CLASSIFICATION ",
|
| 518 |
+
"text_level": 1,
|
| 519 |
+
"bbox": [
|
| 520 |
+
174,
|
| 521 |
+
219,
|
| 522 |
+
598,
|
| 523 |
+
232
|
| 524 |
+
],
|
| 525 |
+
"page_idx": 5
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"type": "text",
|
| 529 |
+
"text": "Setup and Implementation Details: We evaluate the proposed method on PACS (Li et al. (2017)), a widely-used benchmark for domain generalization with four different styles: Art Painting, Cartoon, Photo, and Sketch. The implementation follows the official setup of MixStyle (Zhou et al. (2021b)) with a leave-one-domain-out protocol and ResNet18 (He et al., 2016) is used as the backbone. The random shuffle version of MixStyle is adopted for fair comparisons, which does not use domain labels. In addition to PACS, we also employ Office-Home (Venkateswara et al., 2017) for multidomain generalization experiments in the Appendix. ",
|
| 530 |
+
"bbox": [
|
| 531 |
+
174,
|
| 532 |
+
243,
|
| 533 |
+
825,
|
| 534 |
+
342
|
| 535 |
+
],
|
| 536 |
+
"page_idx": 5
|
| 537 |
+
},
|
| 538 |
+
{
|
| 539 |
+
"type": "text",
|
| 540 |
+
"text": "Experiment Results: The experiments results, shown in Table 1, demonstrate our significant improvement over the baseline method, which shows our superiority to the conventional deterministic approach. Especially in Art and Sketch, our method has nearly $10 \\%$ improvement in average accuracy. Furthermore, our method also outperforms the competing methods, which indicates our method that models diverse uncertain shifts on feature statistics is effective to improve network generalization ability against different domain shifts. Photo has similiar domain characteristics as ImageNet dataset and the slight drop might be due to the ImageNet pretraining (also discussed in (Xu et al., 2021)). Our DSU augments the features and enlarges the diversity of the training domains. In contrast, the baseline method preserves more pre-trained knowledge from ImageNet thus tends to overfit the Photo style dataset benefiting from pretraining. ",
|
| 541 |
+
"bbox": [
|
| 542 |
+
174,
|
| 543 |
+
348,
|
| 544 |
+
825,
|
| 545 |
+
488
|
| 546 |
+
],
|
| 547 |
+
"page_idx": 5
|
| 548 |
+
},
|
| 549 |
+
{
|
| 550 |
+
"type": "table",
|
| 551 |
+
"img_path": "images/c66be72d5a58ece9bf73158f837b4067edffcb88ad9dc81eda7a1c6ad9a38d26.jpg",
|
| 552 |
+
"table_caption": [
|
| 553 |
+
"Table 1: Experiment results of PACS multi-domain classification task. ${ \\mathrm { R S C } } ^ { * }$ denotes the reproduced results from pAdaIN (Nuriel et al., 2021). "
|
| 554 |
+
],
|
| 555 |
+
"table_footnote": [],
|
| 556 |
+
"table_body": "<table><tr><td>Method</td><td>Reference</td><td>Art</td><td>Cartoon</td><td>Photo</td><td>Sketch</td><td>Average (%)</td></tr><tr><td>Baseline</td><td>=</td><td>74.3</td><td>76.7</td><td>96.4</td><td>68.7</td><td>79.0</td></tr><tr><td>Mixup (Zhang et al., 2018)</td><td>ICLR 2018</td><td>76.8</td><td>74.9</td><td>95.8</td><td>66.6</td><td>78.5</td></tr><tr><td>Manifold Mixup (Verma et al.,2019)</td><td>ICML 2019</td><td>75.6</td><td>70.1</td><td>93.5</td><td>65.4</td><td>76.2</td></tr><tr><td>CutMix (Yun et al., 2019)</td><td>ICCV 2019</td><td>74.6</td><td>71.8</td><td>95.6</td><td>65.3</td><td>76.8</td></tr><tr><td>RSC*(Huang et al.,2020)</td><td>ECCV 2020</td><td>78.9</td><td>76.9</td><td>94.1</td><td>76.8</td><td>81.7</td></tr><tr><td>L2A-OT (Zhou et al., 2020a)</td><td>ECCV 2020</td><td>83.3</td><td>78.2</td><td>96.2</td><td>73.6</td><td>82.8</td></tr><tr><td>SagNet (Nam et al., 2021)</td><td>CVPR 2021</td><td>83.6</td><td>77.7</td><td>95.5</td><td>76.3</td><td>83.3</td></tr><tr><td>pAdaIN (Nuriel et al., 2021)</td><td>CVPR 2021</td><td>81.7</td><td>76.6</td><td>96.3</td><td>75.1</td><td>82.5</td></tr><tr><td>MixStyle (Zhou et al., 2021b)</td><td>ICLR 2021</td><td>82.3</td><td>79.0</td><td>96.3</td><td>73.8</td><td>82.8</td></tr><tr><td>DSU</td><td>Ours</td><td>83.6</td><td>79.6</td><td>95.8</td><td>77.6</td><td>84.1</td></tr></table>",
|
| 557 |
+
"bbox": [
|
| 558 |
+
173,
|
| 559 |
+
529,
|
| 560 |
+
825,
|
| 561 |
+
676
|
| 562 |
+
],
|
| 563 |
+
"page_idx": 5
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"type": "text",
|
| 567 |
+
"text": "4.2 GENERALIZATION ON SEMANTIC SEGMENTATION ",
|
| 568 |
+
"text_level": 1,
|
| 569 |
+
"bbox": [
|
| 570 |
+
173,
|
| 571 |
+
683,
|
| 572 |
+
558,
|
| 573 |
+
696
|
| 574 |
+
],
|
| 575 |
+
"page_idx": 5
|
| 576 |
+
},
|
| 577 |
+
{
|
| 578 |
+
"type": "text",
|
| 579 |
+
"text": "Setup and Implementation Details: Semantic segmentation, as a fundamental application for automatic driving, encounters severe performance declines due to scenarios differences (Wang et al., 2020a). GTA5 (Richter et al., 2016) is a synthetic dataset generated from Grand Theft Auto 5 game engine, while Cityscapes (Cordts et al., 2016) is a real-world dataset collected from different cities in primarily Germany. To evaluate the cross-scenario generalization ability of segmentation models, we adopt synthetic GTA5 for training while using real CityScapes for testing. The experiments are conducted on FADA released codes (Wang et al. (2020a)), using DeepLab-v2 (Chen et al., 2018) segmentation network with ResNet101 backbone. Mean Intersection over Union (mIOU) and mean Accuracy (mAcc) of all object categories are used for evaluation. ",
|
| 580 |
+
"bbox": [
|
| 581 |
+
173,
|
| 582 |
+
708,
|
| 583 |
+
825,
|
| 584 |
+
833
|
| 585 |
+
],
|
| 586 |
+
"page_idx": 5
|
| 587 |
+
},
|
| 588 |
+
{
|
| 589 |
+
"type": "table",
|
| 590 |
+
"img_path": "images/2b216b7ed88c5a59690a2da7d41e7f1f2641beefb8e2d8e23ad9f6bbd5278cd4.jpg",
|
| 591 |
+
"table_caption": [
|
| 592 |
+
"Table 2: Experiment results of semantic segmentation from synthetic GAT5 to real Cityscapes. "
|
| 593 |
+
],
|
| 594 |
+
"table_footnote": [],
|
| 595 |
+
"table_body": "<table><tr><td>Method</td><td>Reference</td><td>mIOU (%)</td><td>mAcc (%)</td></tr><tr><td>Baseline</td><td>-</td><td>37.0</td><td>51.5</td></tr><tr><td>pAdaIN (Nuriel et al., 2021)</td><td>CVPR 2021</td><td>38.3</td><td>52.1</td></tr><tr><td>Mixstyle (Zhou et al., 2021b)</td><td>ICLR 2021</td><td>40.3</td><td>53.8</td></tr><tr><td>DSU</td><td>Ours</td><td>43.1</td><td>57.0</td></tr></table>",
|
| 596 |
+
"bbox": [
|
| 597 |
+
250,
|
| 598 |
+
861,
|
| 599 |
+
743,
|
| 600 |
+
939
|
| 601 |
+
],
|
| 602 |
+
"page_idx": 5
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"type": "text",
|
| 606 |
+
"text": "Experiment Results: Table 2 shows the experiment results compared to related methods. As for a pixel-level classification task, improper changes of feature statistics might constrain the performances. The variants generated from our method are centered on the original feature statistics with different perturbations. These changes of feature statistics are mild for preserving the detailed information in these dense tasks. Meanwhile, our method can take full use of the diverse driving scenes and generates diverse variants, thus show a significant improvement on mIOU and mAc by $6 . 1 \\%$ and $5 . 5 \\%$ , respectively. The visualization result is shown in Figure 3. ",
|
| 607 |
+
"bbox": [
|
| 608 |
+
173,
|
| 609 |
+
103,
|
| 610 |
+
825,
|
| 611 |
+
202
|
| 612 |
+
],
|
| 613 |
+
"page_idx": 6
|
| 614 |
+
},
|
| 615 |
+
{
|
| 616 |
+
"type": "image",
|
| 617 |
+
"img_path": "images/68a21bfca5c82e5cded24c0754d1924eda335e41d01b065155f433ceaf7038ba.jpg",
|
| 618 |
+
"image_caption": [],
|
| 619 |
+
"image_footnote": [],
|
| 620 |
+
"bbox": [
|
| 621 |
+
238,
|
| 622 |
+
208,
|
| 623 |
+
758,
|
| 624 |
+
273
|
| 625 |
+
],
|
| 626 |
+
"page_idx": 6
|
| 627 |
+
},
|
| 628 |
+
{
|
| 629 |
+
"type": "text",
|
| 630 |
+
"text": "Figure 3: The visualization on unseen domain Cityscapes with the model trained on synthetic GTA5. ",
|
| 631 |
+
"bbox": [
|
| 632 |
+
171,
|
| 633 |
+
284,
|
| 634 |
+
821,
|
| 635 |
+
299
|
| 636 |
+
],
|
| 637 |
+
"page_idx": 6
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"type": "text",
|
| 641 |
+
"text": "4.3 GENERALIZATION ON INSTANCE RETRIEVAL",
|
| 642 |
+
"text_level": 1,
|
| 643 |
+
"bbox": [
|
| 644 |
+
174,
|
| 645 |
+
320,
|
| 646 |
+
524,
|
| 647 |
+
334
|
| 648 |
+
],
|
| 649 |
+
"page_idx": 6
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"text": "Setup and Implementation Details: In this section, person re-identification (ReID), which aims at matching the same person across disjoint camera views, is used to verify the effectiveness of our method on the instance retrieval task. Experiments are conducted on the widely used DukeMTMC (Ristani et al. (2016)) and Market1501 (Zheng et al. (2015)) datasets. The implementation is based on MMT (Ge et al., 2020) released codes and ResNet50 is adopted as the backbone. Meanwhile, mean Average Precision (mAP) and Rank-1 (R1) precision are used as the evaluation criterions. ",
|
| 654 |
+
"bbox": [
|
| 655 |
+
173,
|
| 656 |
+
347,
|
| 657 |
+
825,
|
| 658 |
+
431
|
| 659 |
+
],
|
| 660 |
+
"page_idx": 6
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "table",
|
| 664 |
+
"img_path": "images/ae73f3417c12f1967e8893640947266d9533a961f8db2d90fc513a2d7de0e60c.jpg",
|
| 665 |
+
"table_caption": [
|
| 666 |
+
"Table 3: Experiment results of instance retrieval on ReID dataset DukeMTMC and Market1501. A $ \\mathbf { B }$ denotes models are trained on A while evaluated on B. For fair comparisons, we reproduce the experiments under the same framework. "
|
| 667 |
+
],
|
| 668 |
+
"table_footnote": [],
|
| 669 |
+
"table_body": "<table><tr><td>Method</td><td>Reference</td><td>Market → Duke mAP (%) R1 (%)</td><td>mAP (%)</td><td>Duke-→Market R1 (%)</td></tr><tr><td>Baseline</td><td>1</td><td>25.8</td><td>42.3 46.1</td><td>26.7</td></tr><tr><td>pAdaIN (Nuriel et al., 2021)</td><td>CVPR 2021</td><td>28.0</td><td>27.9</td><td>54.7 56.1</td></tr><tr><td>MixStyle (Zhou et al.,2021b)</td><td>ICLR 2021</td><td>28.2</td><td>28.1</td><td>56.6</td></tr><tr><td>DSU</td><td>Ours</td><td>32.0</td><td>46.7 52.0 32.4</td><td>63.7</td></tr></table>",
|
| 670 |
+
"bbox": [
|
| 671 |
+
187,
|
| 672 |
+
497,
|
| 673 |
+
808,
|
| 674 |
+
590
|
| 675 |
+
],
|
| 676 |
+
"page_idx": 6
|
| 677 |
+
},
|
| 678 |
+
{
|
| 679 |
+
"type": "text",
|
| 680 |
+
"text": "Experiment Results: ReID is a fine-grained instance retrieval task, where the subtle information of persons is important for retrieving an instance. MixStyle and pAdain rely on a reference sample to generate new feature statistics, which might introduce confounded information from the reference sample. Compared to them, our method does better in maintaining the original information and also has more variant possibilities. The experiment results are demonstrated in Table 3. Our method achieves huge improvement compared to the baseline method and also outperform MixStyle and pAdaIN by a big margin. ",
|
| 681 |
+
"bbox": [
|
| 682 |
+
173,
|
| 683 |
+
604,
|
| 684 |
+
825,
|
| 685 |
+
704
|
| 686 |
+
],
|
| 687 |
+
"page_idx": 6
|
| 688 |
+
},
|
| 689 |
+
{
|
| 690 |
+
"type": "text",
|
| 691 |
+
"text": "4.4 ROBUSTNESS TOWARDS CORRUPTIONS ",
|
| 692 |
+
"text_level": 1,
|
| 693 |
+
"bbox": [
|
| 694 |
+
176,
|
| 695 |
+
723,
|
| 696 |
+
486,
|
| 697 |
+
737
|
| 698 |
+
],
|
| 699 |
+
"page_idx": 6
|
| 700 |
+
},
|
| 701 |
+
{
|
| 702 |
+
"type": "text",
|
| 703 |
+
"text": "Setup and Implementation Details: We validate the proposed method for robustness towards corruptions on ImageNet-C (Hendrycks & Dietterich (2019)), which contains 15 different pixel-level corruptions. ResNet50 is trained with 100 epochs for convergence on large-scale ImageNet-1K (Deng et al. (2009)) and the hyperparameter $p$ is set as 0.1 for training in ImageNet. We also add our method on APR (Chen et al. (2021)), a recently state-of-the-art method on ImageNet-C, to verify that our method can be compatible with other image-level augmentation methods. Error is adopted as the evalution metric for clean ImageNet. Mean Corruption Error (mCE) is adopted as evaluation metric for ImageNet-C, which is computed as the average of the 15 different corruption errors and normalized by the corruption error of AlexNet (Krizhevsky et al. (2012)). ",
|
| 704 |
+
"bbox": [
|
| 705 |
+
173,
|
| 706 |
+
750,
|
| 707 |
+
825,
|
| 708 |
+
875
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 6
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "text",
|
| 714 |
+
"text": "Experiment Results: Although the corruptions are imposed on the pixel level, they still introduce a shift in the statistics (Benz et al., 2021). So our method shows consistent improvement on ImageNetC. Meanwhile, the instances in the testing set may not always fall into the distribution of the training set, and they still have slight statistic shifts (Gao et al. (2021b)). Thus it can be seen that the withindataset ImageNet accuracy is also increased. When combining APR with our method, mCE can be decreased from $6 5 . 0 \\%$ to $6 4 . 1 \\%$ , showing that our method can be compatible with state-of-the-art methods on ImageNet-C for further improvement. ",
|
| 715 |
+
"bbox": [
|
| 716 |
+
176,
|
| 717 |
+
882,
|
| 718 |
+
823,
|
| 719 |
+
924
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 6
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "table",
|
| 725 |
+
"img_path": "images/68305dcad86b3a12f6aff52d6f759852ee32d1d7c20a15bb1b62d418890d68b1.jpg",
|
| 726 |
+
"table_caption": [
|
| 727 |
+
"Table 4: Experiment results of clean image classification on ImageNet, and the robustness toward corruptions on ImageNet-C. "
|
| 728 |
+
],
|
| 729 |
+
"table_footnote": [],
|
| 730 |
+
"table_body": "<table><tr><td></td><td>Clean(↓）</td><td>Corrupted (↓)</td><td></td><td>Noise</td><td>Implulse</td><td></td><td>Blur</td><td></td><td></td><td></td><td>Weather</td><td></td><td></td><td></td><td>Digital</td><td></td><td></td></tr><tr><td></td><td>Error (%)</td><td>mCE(%)</td><td>Gauss</td><td>Shot</td><td></td><td>Defocus</td><td>Glass</td><td>Motion</td><td>Zoom</td><td>Snow</td><td>Frost</td><td>Fog</td><td>Bright</td><td>Contrast</td><td>Elastic</td><td>Pixel</td><td>JPEG</td></tr><tr><td>Baseline</td><td>23.8</td><td>76.2</td><td></td><td>81</td><td></td><td></td><td>87</td><td>76</td><td>80</td><td>78</td><td>74</td><td>67</td><td></td><td>70</td><td>83</td><td>76</td><td>73</td></tr><tr><td>DSU</td><td>23.4</td><td>73.4</td><td>6</td><td>77</td><td>5</td><td>五</td><td>83</td><td>77</td><td>79</td><td>74</td><td>71</td><td>66</td><td>5</td><td>68</td><td>82</td><td>65</td><td>71</td></tr><tr><td>APR</td><td>24.0</td><td>65.0</td><td>5</td><td></td><td>50</td><td>69</td><td>85</td><td>69</td><td>79</td><td>62</td><td>64</td><td>55</td><td>54</td><td>63</td><td>84</td><td>65</td><td>65</td></tr><tr><td>APR+DSU</td><td>23.7</td><td>64.1</td><td></td><td>56</td><td>49</td><td>69</td><td>84</td><td>67</td><td>78</td><td>61</td><td>63</td><td>51</td><td>53</td><td>56</td><td>83</td><td>66</td><td>65</td></tr></table>",
|
| 731 |
+
"bbox": [
|
| 732 |
+
174,
|
| 733 |
+
127,
|
| 734 |
+
823,
|
| 735 |
+
178
|
| 736 |
+
],
|
| 737 |
+
"page_idx": 7
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"type": "text",
|
| 741 |
+
"text": "",
|
| 742 |
+
"bbox": [
|
| 743 |
+
174,
|
| 744 |
+
205,
|
| 745 |
+
825,
|
| 746 |
+
262
|
| 747 |
+
],
|
| 748 |
+
"page_idx": 7
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"type": "text",
|
| 752 |
+
"text": "5 ABLATION STUDY ",
|
| 753 |
+
"text_level": 1,
|
| 754 |
+
"bbox": [
|
| 755 |
+
176,
|
| 756 |
+
296,
|
| 757 |
+
356,
|
| 758 |
+
313
|
| 759 |
+
],
|
| 760 |
+
"page_idx": 7
|
| 761 |
+
},
|
| 762 |
+
{
|
| 763 |
+
"type": "text",
|
| 764 |
+
"text": "In this section, we perform an extensive ablation study of the proposed method on PACS and segmentation task (GTA5 to Cityscapes) with models trained on ResNet. The effects of different inserted positions and hyper-parameter of the proposed method are analyzed below. Meantime, we also analyze the effects on different choices of uncertainty distribution. ",
|
| 765 |
+
"bbox": [
|
| 766 |
+
174,
|
| 767 |
+
337,
|
| 768 |
+
825,
|
| 769 |
+
392
|
| 770 |
+
],
|
| 771 |
+
"page_idx": 7
|
| 772 |
+
},
|
| 773 |
+
{
|
| 774 |
+
"type": "text",
|
| 775 |
+
"text": "Effects of Different Inserted Positions: DSU can be a plugand-play module to be readily inserted at any position. Here we name the positions of ResNet after first Conv, Max Pooling ",
|
| 776 |
+
"bbox": [
|
| 777 |
+
173,
|
| 778 |
+
400,
|
| 779 |
+
387,
|
| 780 |
+
483
|
| 781 |
+
],
|
| 782 |
+
"page_idx": 7
|
| 783 |
+
},
|
| 784 |
+
{
|
| 785 |
+
"type": "table",
|
| 786 |
+
"img_path": "images/2f77cce4c4b23e2789674b4582285799c7d54da979581b5a0a7025ca8d23217a.jpg",
|
| 787 |
+
"table_caption": [
|
| 788 |
+
"Table 5: Effects of different inserted positions. "
|
| 789 |
+
],
|
| 790 |
+
"table_footnote": [],
|
| 791 |
+
"table_body": "<table><tr><td>Inserted Positions</td><td>Baseline</td><td>0-3</td><td>1-4</td><td>2-5</td><td>0-5</td></tr><tr><td>PACS</td><td>79.0</td><td>82.2</td><td>83.1</td><td>83.5</td><td>84.1</td></tr><tr><td> GTA5 to Cityscapes</td><td>37.0</td><td>41.1</td><td>40.9</td><td>42.1</td><td>43.1</td></tr></table>",
|
| 792 |
+
"bbox": [
|
| 793 |
+
398,
|
| 794 |
+
428,
|
| 795 |
+
821,
|
| 796 |
+
478
|
| 797 |
+
],
|
| 798 |
+
"page_idx": 7
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"type": "text",
|
| 802 |
+
"text": "layer, 1,2,3,4-th ConvBlock as 0,1,2,3,4,5 respectively. As shown in Table 5, no matter where the modules are inserted, the performances are consistently higher than the baseline method. The results show that inserting the modules at positions 0-5 would have better performances, which also indicates modeling the uncertainty in all training stages will have better effects. Based on the analysis, we plug the module into positions 0-5 in all experiments. ",
|
| 803 |
+
"bbox": [
|
| 804 |
+
173,
|
| 805 |
+
483,
|
| 806 |
+
820,
|
| 807 |
+
553
|
| 808 |
+
],
|
| 809 |
+
"page_idx": 7
|
| 810 |
+
},
|
| 811 |
+
{
|
| 812 |
+
"type": "text",
|
| 813 |
+
"text": "Effects of Hyper-parameter: The hyper-parameter of the probability $p$ is to trade off the strength of feature statistics augmentation. As shown in Figure 4, the results are not sensitive to the probability setting and the accuracy reaches the best results when setting $p$ as 0.5, which is also adopted as the default setting in all experiments if not specified. ",
|
| 814 |
+
"bbox": [
|
| 815 |
+
174,
|
| 816 |
+
560,
|
| 817 |
+
549,
|
| 818 |
+
656
|
| 819 |
+
],
|
| 820 |
+
"page_idx": 7
|
| 821 |
+
},
|
| 822 |
+
{
|
| 823 |
+
"type": "text",
|
| 824 |
+
"text": "Choices of Uncertainty Distribution: In our method, the Gaussian distribution with uncertainty estimation is adopted as the default setting, we also conduct other distributions for comparisons in Table 6. Specifically, Random denotes directly adding random shifts draw from a fixed Gaussian $\\mathcal { N } ( 0 , 1 )$ , and Uniform denotes that the shifts are drawn from $\\mathrm { U } ( - \\Sigma , \\Sigma )$ , where $\\Sigma$ is the scope obtained from our uncertainty estimation. As we can see, directly using Gaussian distribution with the improper variant scope will harm the model performances, indicating the variant range of feature statistics should have some instructions. Further analysis about different vanilla Gaussian distributions with pre-defined standard deviation are conducted in the Appendix. Meanwhile, the result of Uniform shows some improvement but is still lower than DSU, which indicates the boundless Gaussian distribution is more helpful to model more diverse variants. ",
|
| 825 |
+
"bbox": [
|
| 826 |
+
174,
|
| 827 |
+
665,
|
| 828 |
+
549,
|
| 829 |
+
691
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 7
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "text",
|
| 835 |
+
"text": "",
|
| 836 |
+
"bbox": [
|
| 837 |
+
173,
|
| 838 |
+
693,
|
| 839 |
+
825,
|
| 840 |
+
816
|
| 841 |
+
],
|
| 842 |
+
"page_idx": 7
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"type": "image",
|
| 846 |
+
"img_path": "images/eb2c6212021035f11adb3701509d09bba3bc1d9ed7ce60b307b7f16a134421e9.jpg",
|
| 847 |
+
"image_caption": [
|
| 848 |
+
"Figure 4: The effects on the hyperparameter probability. "
|
| 849 |
+
],
|
| 850 |
+
"image_footnote": [],
|
| 851 |
+
"bbox": [
|
| 852 |
+
575,
|
| 853 |
+
558,
|
| 854 |
+
820,
|
| 855 |
+
643
|
| 856 |
+
],
|
| 857 |
+
"page_idx": 7
|
| 858 |
+
},
|
| 859 |
+
{
|
| 860 |
+
"type": "table",
|
| 861 |
+
"img_path": "images/8c31cd6d52deea3b343eccbbffe08cfee395a926d56d40c11e6f02ecc3cb5469.jpg",
|
| 862 |
+
"table_caption": [
|
| 863 |
+
"Table 6: Different choices of distribution for uncertainty. "
|
| 864 |
+
],
|
| 865 |
+
"table_footnote": [],
|
| 866 |
+
"table_body": "<table><tr><td>Choice</td><td>Baseline</td><td>Random</td><td>Uniform</td><td>DSU</td></tr><tr><td>PACS</td><td>79.0</td><td>76.9</td><td>81.9</td><td>84.1</td></tr><tr><td>GTA5 to Cityscapes</td><td>37.0</td><td>38.2</td><td>41.6</td><td>43.1</td></tr></table>",
|
| 867 |
+
"bbox": [
|
| 868 |
+
279,
|
| 869 |
+
862,
|
| 870 |
+
714,
|
| 871 |
+
912
|
| 872 |
+
],
|
| 873 |
+
"page_idx": 7
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "text",
|
| 877 |
+
"text": "6 FURTHER ANALYSIS ",
|
| 878 |
+
"text_level": 1,
|
| 879 |
+
"bbox": [
|
| 880 |
+
176,
|
| 881 |
+
102,
|
| 882 |
+
377,
|
| 883 |
+
118
|
| 884 |
+
],
|
| 885 |
+
"page_idx": 8
|
| 886 |
+
},
|
| 887 |
+
{
|
| 888 |
+
"type": "text",
|
| 889 |
+
"text": "6.1 QUANTITATIVE ANALYSIS ON THE PROPOSED METHOD ",
|
| 890 |
+
"text_level": 1,
|
| 891 |
+
"bbox": [
|
| 892 |
+
174,
|
| 893 |
+
133,
|
| 894 |
+
593,
|
| 895 |
+
147
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 8
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "text",
|
| 901 |
+
"text": "In this subsection, we will analyze the effects of the proposed method on both intermediate features and feature representations. Quantitative experiments are conducted on PACS, where we choose Art Painting as the unseen testing domain and the rests are used as training domains. ",
|
| 902 |
+
"bbox": [
|
| 903 |
+
174,
|
| 904 |
+
160,
|
| 905 |
+
825,
|
| 906 |
+
202
|
| 907 |
+
],
|
| 908 |
+
"page_idx": 8
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"type": "text",
|
| 912 |
+
"text": "To study the phenomena of feature statistic shifts, we capture the intermediate features after the second block in ResNet18 and measure the average feature statistics values of one category in the training and testing domain, respectively. The distributions of feature statistics are shown in Figure 5. As the previous works (Wang et al., 2020b; 2019a) show, the feature statistics extracted from the baseline model show an obvious shift due to different data distribution. It can be seen that the model trained with our method has less shift. Our method can help the model gain robustness towards domain shifts, as it properly models the potential feature statistic shifts. ",
|
| 913 |
+
"bbox": [
|
| 914 |
+
173,
|
| 915 |
+
208,
|
| 916 |
+
826,
|
| 917 |
+
305
|
| 918 |
+
],
|
| 919 |
+
"page_idx": 8
|
| 920 |
+
},
|
| 921 |
+
{
|
| 922 |
+
"type": "image",
|
| 923 |
+
"img_path": "images/17b4651e1b926d357fa953211aec3e884b58569d5a56fc2937a2918a86050585.jpg",
|
| 924 |
+
"image_caption": [
|
| 925 |
+
"Figure 5: Quantitative analysis on the shifts of feature statistics (mean and standard deviation) between training source domains and unseen testing domain. "
|
| 926 |
+
],
|
| 927 |
+
"image_footnote": [],
|
| 928 |
+
"bbox": [
|
| 929 |
+
271,
|
| 930 |
+
309,
|
| 931 |
+
727,
|
| 932 |
+
391
|
| 933 |
+
],
|
| 934 |
+
"page_idx": 8
|
| 935 |
+
},
|
| 936 |
+
{
|
| 937 |
+
"type": "text",
|
| 938 |
+
"text": "6.2 VISUALIZATION ON THE SYNTHETIC CHANGES ",
|
| 939 |
+
"text_level": 1,
|
| 940 |
+
"bbox": [
|
| 941 |
+
174,
|
| 942 |
+
446,
|
| 943 |
+
539,
|
| 944 |
+
460
|
| 945 |
+
],
|
| 946 |
+
"page_idx": 8
|
| 947 |
+
},
|
| 948 |
+
{
|
| 949 |
+
"type": "text",
|
| 950 |
+
"text": "Besides the quantitative experiment results, we also obtain a more intuitional view of the diverse changes provided by our method, through visualizing the reconstruction results using a predefined autoencoder1 (Huang & Belongie (2017)), where the proposed module is inserted into the encoder, and inverse the feature representations into synthetic images after the decoder. As the results shown in Figure 6, the reconstructed images obtained from our probabilistic approach show diverse synthetic changes, such as the environment, object texture, and contrast, etc. ",
|
| 951 |
+
"bbox": [
|
| 952 |
+
173,
|
| 953 |
+
472,
|
| 954 |
+
826,
|
| 955 |
+
556
|
| 956 |
+
],
|
| 957 |
+
"page_idx": 8
|
| 958 |
+
},
|
| 959 |
+
{
|
| 960 |
+
"type": "image",
|
| 961 |
+
"img_path": "images/8961e5d8f47d0a06768da6910ee5df27805c86ece4db6111eb881f43f90c0807.jpg",
|
| 962 |
+
"image_caption": [
|
| 963 |
+
"Figure 6: The visualization on diverse synthetic changes obtained from our method. "
|
| 964 |
+
],
|
| 965 |
+
"image_footnote": [],
|
| 966 |
+
"bbox": [
|
| 967 |
+
222,
|
| 968 |
+
569,
|
| 969 |
+
776,
|
| 970 |
+
637
|
| 971 |
+
],
|
| 972 |
+
"page_idx": 8
|
| 973 |
+
},
|
| 974 |
+
{
|
| 975 |
+
"type": "text",
|
| 976 |
+
"text": "7 CONCLUSIONS ",
|
| 977 |
+
"text_level": 1,
|
| 978 |
+
"bbox": [
|
| 979 |
+
176,
|
| 980 |
+
693,
|
| 981 |
+
328,
|
| 982 |
+
708
|
| 983 |
+
],
|
| 984 |
+
"page_idx": 8
|
| 985 |
+
},
|
| 986 |
+
{
|
| 987 |
+
"type": "text",
|
| 988 |
+
"text": "In this paper, we propose a probabilistic approach to improve the network generalization ability by modeling the uncertainty of domain shifts with synthesized feature statistics during training. Each feature statistic is hypothesized to follow a multi-variate Gaussian distribution for modeling the diverse potential shifts. Due to the generated feature statistics with diverse distribution possibilities, the models can gain better robustness towards diverse domain shifts. Experiment results demonstrate the effectiveness of our method in improving the network generalization ability. ",
|
| 989 |
+
"bbox": [
|
| 990 |
+
173,
|
| 991 |
+
723,
|
| 992 |
+
825,
|
| 993 |
+
808
|
| 994 |
+
],
|
| 995 |
+
"page_idx": 8
|
| 996 |
+
},
|
| 997 |
+
{
|
| 998 |
+
"type": "text",
|
| 999 |
+
"text": "ACKNOWLEDGEMENT ",
|
| 1000 |
+
"text_level": 1,
|
| 1001 |
+
"bbox": [
|
| 1002 |
+
176,
|
| 1003 |
+
825,
|
| 1004 |
+
326,
|
| 1005 |
+
838
|
| 1006 |
+
],
|
| 1007 |
+
"page_idx": 8
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "text",
|
| 1011 |
+
"text": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the $\\mathrm { N g }$ Teng Fong Charitable Foundation. ",
|
| 1012 |
+
"bbox": [
|
| 1013 |
+
176,
|
| 1014 |
+
851,
|
| 1015 |
+
825,
|
| 1016 |
+
892
|
| 1017 |
+
],
|
| 1018 |
+
"page_idx": 8
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"type": "text",
|
| 1022 |
+
"text": "REFERENCES ",
|
| 1023 |
+
"text_level": 1,
|
| 1024 |
+
"bbox": [
|
| 1025 |
+
176,
|
| 1026 |
+
102,
|
| 1027 |
+
287,
|
| 1028 |
+
117
|
| 1029 |
+
],
|
| 1030 |
+
"page_idx": 9
|
| 1031 |
+
},
|
| 1032 |
+
{
|
| 1033 |
+
"type": "text",
|
| 1034 |
+
"text": "Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151–175, 2010. ",
|
| 1035 |
+
"bbox": [
|
| 1036 |
+
174,
|
| 1037 |
+
126,
|
| 1038 |
+
825,
|
| 1039 |
+
167
|
| 1040 |
+
],
|
| 1041 |
+
"page_idx": 9
|
| 1042 |
+
},
|
| 1043 |
+
{
|
| 1044 |
+
"type": "text",
|
| 1045 |
+
"text": "Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Revisiting batch normalization for improving corruption robustness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 494–503, 2021. ",
|
| 1046 |
+
"bbox": [
|
| 1047 |
+
176,
|
| 1048 |
+
179,
|
| 1049 |
+
823,
|
| 1050 |
+
222
|
| 1051 |
+
],
|
| 1052 |
+
"page_idx": 9
|
| 1053 |
+
},
|
| 1054 |
+
{
|
| 1055 |
+
"type": "text",
|
| 1056 |
+
"text": "Gilles Blanchard, Gyemin Lee, and Clayton Scott. Generalizing from several related classification tasks to a new unlabeled sample. In Advances in Neural Information Processing Systems, 2011. ",
|
| 1057 |
+
"bbox": [
|
| 1058 |
+
176,
|
| 1059 |
+
231,
|
| 1060 |
+
821,
|
| 1061 |
+
261
|
| 1062 |
+
],
|
| 1063 |
+
"page_idx": 9
|
| 1064 |
+
},
|
| 1065 |
+
{
|
| 1066 |
+
"type": "text",
|
| 1067 |
+
"text": "Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5710–5719, 2020. ",
|
| 1068 |
+
"bbox": [
|
| 1069 |
+
173,
|
| 1070 |
+
270,
|
| 1071 |
+
826,
|
| 1072 |
+
314
|
| 1073 |
+
],
|
| 1074 |
+
"page_idx": 9
|
| 1075 |
+
},
|
| 1076 |
+
{
|
| 1077 |
+
"type": "text",
|
| 1078 |
+
"text": "Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In European Conference in Computer Vision (ECCV), 2020. ",
|
| 1079 |
+
"bbox": [
|
| 1080 |
+
173,
|
| 1081 |
+
324,
|
| 1082 |
+
826,
|
| 1083 |
+
367
|
| 1084 |
+
],
|
| 1085 |
+
"page_idx": 9
|
| 1086 |
+
},
|
| 1087 |
+
{
|
| 1088 |
+
"type": "text",
|
| 1089 |
+
"text": "Guangyao Chen, Peixi Peng, Li Ma, Jia Li, Lin Du, and Yonghong Tian. Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. ",
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
176,
|
| 1092 |
+
376,
|
| 1093 |
+
821,
|
| 1094 |
+
420
|
| 1095 |
+
],
|
| 1096 |
+
"page_idx": 9
|
| 1097 |
+
},
|
| 1098 |
+
{
|
| 1099 |
+
"type": "text",
|
| 1100 |
+
"text": "Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4): 834–848, 2018. ",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
173,
|
| 1103 |
+
429,
|
| 1104 |
+
825,
|
| 1105 |
+
486
|
| 1106 |
+
],
|
| 1107 |
+
"page_idx": 9
|
| 1108 |
+
},
|
| 1109 |
+
{
|
| 1110 |
+
"type": "text",
|
| 1111 |
+
"text": "Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T. Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11580–11590, 2021. ",
|
| 1112 |
+
"bbox": [
|
| 1113 |
+
173,
|
| 1114 |
+
496,
|
| 1115 |
+
825,
|
| 1116 |
+
554
|
| 1117 |
+
],
|
| 1118 |
+
"page_idx": 9
|
| 1119 |
+
},
|
| 1120 |
+
{
|
| 1121 |
+
"type": "text",
|
| 1122 |
+
"text": "Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. ",
|
| 1123 |
+
"bbox": [
|
| 1124 |
+
174,
|
| 1125 |
+
564,
|
| 1126 |
+
825,
|
| 1127 |
+
619
|
| 1128 |
+
],
|
| 1129 |
+
"page_idx": 9
|
| 1130 |
+
},
|
| 1131 |
+
{
|
| 1132 |
+
"type": "text",
|
| 1133 |
+
"text": "Yongxing Dai, Xiaotong Li, Jun Liu, Zekun Tong, and Ling-Yu Duan. Generalizable person reidentification with relevance-aware mixture of experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 16145–16154, June 2021. ",
|
| 1134 |
+
"bbox": [
|
| 1135 |
+
174,
|
| 1136 |
+
630,
|
| 1137 |
+
823,
|
| 1138 |
+
672
|
| 1139 |
+
],
|
| 1140 |
+
"page_idx": 9
|
| 1141 |
+
},
|
| 1142 |
+
{
|
| 1143 |
+
"type": "text",
|
| 1144 |
+
"text": "Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255, 2009. ",
|
| 1145 |
+
"bbox": [
|
| 1146 |
+
173,
|
| 1147 |
+
683,
|
| 1148 |
+
825,
|
| 1149 |
+
726
|
| 1150 |
+
],
|
| 1151 |
+
"page_idx": 9
|
| 1152 |
+
},
|
| 1153 |
+
{
|
| 1154 |
+
"type": "text",
|
| 1155 |
+
"text": "Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, pp. 1126–1135, 2017. ",
|
| 1156 |
+
"bbox": [
|
| 1157 |
+
173,
|
| 1158 |
+
736,
|
| 1159 |
+
825,
|
| 1160 |
+
779
|
| 1161 |
+
],
|
| 1162 |
+
"page_idx": 9
|
| 1163 |
+
},
|
| 1164 |
+
{
|
| 1165 |
+
"type": "text",
|
| 1166 |
+
"text": "Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of The 33rd International Conference on Machine Learning, volume 48, pp. 1050–1059, 2016. ",
|
| 1167 |
+
"bbox": [
|
| 1168 |
+
173,
|
| 1169 |
+
789,
|
| 1170 |
+
823,
|
| 1171 |
+
832
|
| 1172 |
+
],
|
| 1173 |
+
"page_idx": 9
|
| 1174 |
+
},
|
| 1175 |
+
{
|
| 1176 |
+
"type": "text",
|
| 1177 |
+
"text": "Shang-Hua Gao, Qi Han, Duo Li, Ming-Ming Cheng, and Pai Peng. Representative batch normalization with feature calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8669–8679, 2021a. ",
|
| 1178 |
+
"bbox": [
|
| 1179 |
+
173,
|
| 1180 |
+
842,
|
| 1181 |
+
823,
|
| 1182 |
+
885
|
| 1183 |
+
],
|
| 1184 |
+
"page_idx": 9
|
| 1185 |
+
},
|
| 1186 |
+
{
|
| 1187 |
+
"type": "text",
|
| 1188 |
+
"text": "Shang-Hua Gao, Qi Han, Duo Li, Pai Peng, Ming-Ming Cheng, and Pai Peng. Representative batch normalization with feature calibration. 2021b. ",
|
| 1189 |
+
"bbox": [
|
| 1190 |
+
173,
|
| 1191 |
+
895,
|
| 1192 |
+
821,
|
| 1193 |
+
924
|
| 1194 |
+
],
|
| 1195 |
+
"page_idx": 9
|
| 1196 |
+
},
|
| 1197 |
+
{
|
| 1198 |
+
"type": "text",
|
| 1199 |
+
"text": "Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual mean-teaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. ",
|
| 1200 |
+
"bbox": [
|
| 1201 |
+
174,
|
| 1202 |
+
103,
|
| 1203 |
+
821,
|
| 1204 |
+
146
|
| 1205 |
+
],
|
| 1206 |
+
"page_idx": 10
|
| 1207 |
+
},
|
| 1208 |
+
{
|
| 1209 |
+
"type": "text",
|
| 1210 |
+
"text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. ",
|
| 1211 |
+
"bbox": [
|
| 1212 |
+
174,
|
| 1213 |
+
155,
|
| 1214 |
+
825,
|
| 1215 |
+
198
|
| 1216 |
+
],
|
| 1217 |
+
"page_idx": 10
|
| 1218 |
+
},
|
| 1219 |
+
{
|
| 1220 |
+
"type": "text",
|
| 1221 |
+
"text": "Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. International Conference on Learning Representations, 2019. ",
|
| 1222 |
+
"bbox": [
|
| 1223 |
+
174,
|
| 1224 |
+
207,
|
| 1225 |
+
820,
|
| 1226 |
+
237
|
| 1227 |
+
],
|
| 1228 |
+
"page_idx": 10
|
| 1229 |
+
},
|
| 1230 |
+
{
|
| 1231 |
+
"type": "text",
|
| 1232 |
+
"text": "Dan Hendrycks, Norman Mu, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple method to improve robustness and uncertainty under data shift. In International Conference on Learning Representations, 2020. ",
|
| 1233 |
+
"bbox": [
|
| 1234 |
+
174,
|
| 1235 |
+
244,
|
| 1236 |
+
825,
|
| 1237 |
+
289
|
| 1238 |
+
],
|
| 1239 |
+
"page_idx": 10
|
| 1240 |
+
},
|
| 1241 |
+
{
|
| 1242 |
+
"type": "text",
|
| 1243 |
+
"text": "X. Huang and S. Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In 2017 IEEE International Conference on Computer Vision (ICCV), pp. 1510–1519, 2017. ",
|
| 1244 |
+
"bbox": [
|
| 1245 |
+
174,
|
| 1246 |
+
297,
|
| 1247 |
+
821,
|
| 1248 |
+
327
|
| 1249 |
+
],
|
| 1250 |
+
"page_idx": 10
|
| 1251 |
+
},
|
| 1252 |
+
{
|
| 1253 |
+
"type": "text",
|
| 1254 |
+
"text": "Zeyi Huang, Haohan Wang, Eric P. Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In The European Conference on Computer Vision (ECCV), pp. 124–140, 2020. ",
|
| 1255 |
+
"bbox": [
|
| 1256 |
+
174,
|
| 1257 |
+
335,
|
| 1258 |
+
823,
|
| 1259 |
+
366
|
| 1260 |
+
],
|
| 1261 |
+
"page_idx": 10
|
| 1262 |
+
},
|
| 1263 |
+
{
|
| 1264 |
+
"type": "text",
|
| 1265 |
+
"text": "Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. ",
|
| 1266 |
+
"bbox": [
|
| 1267 |
+
173,
|
| 1268 |
+
373,
|
| 1269 |
+
823,
|
| 1270 |
+
404
|
| 1271 |
+
],
|
| 1272 |
+
"page_idx": 10
|
| 1273 |
+
},
|
| 1274 |
+
{
|
| 1275 |
+
"type": "text",
|
| 1276 |
+
"text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25:1097–1105, 2012. ",
|
| 1277 |
+
"bbox": [
|
| 1278 |
+
174,
|
| 1279 |
+
411,
|
| 1280 |
+
825,
|
| 1281 |
+
455
|
| 1282 |
+
],
|
| 1283 |
+
"page_idx": 10
|
| 1284 |
+
},
|
| 1285 |
+
{
|
| 1286 |
+
"type": "text",
|
| 1287 |
+
"text": "Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12383–12392, 2021. ",
|
| 1288 |
+
"bbox": [
|
| 1289 |
+
174,
|
| 1290 |
+
464,
|
| 1291 |
+
825,
|
| 1292 |
+
508
|
| 1293 |
+
],
|
| 1294 |
+
"page_idx": 10
|
| 1295 |
+
},
|
| 1296 |
+
{
|
| 1297 |
+
"type": "text",
|
| 1298 |
+
"text": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. ",
|
| 1299 |
+
"bbox": [
|
| 1300 |
+
173,
|
| 1301 |
+
516,
|
| 1302 |
+
823,
|
| 1303 |
+
559
|
| 1304 |
+
],
|
| 1305 |
+
"page_idx": 10
|
| 1306 |
+
},
|
| 1307 |
+
{
|
| 1308 |
+
"type": "text",
|
| 1309 |
+
"text": "Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018a. ",
|
| 1310 |
+
"bbox": [
|
| 1311 |
+
174,
|
| 1312 |
+
568,
|
| 1313 |
+
826,
|
| 1314 |
+
611
|
| 1315 |
+
],
|
| 1316 |
+
"page_idx": 10
|
| 1317 |
+
},
|
| 1318 |
+
{
|
| 1319 |
+
"type": "text",
|
| 1320 |
+
"text": "Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C. Kot. Domain generalization with adversarial feature learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 5400–5409, 2018b. ",
|
| 1321 |
+
"bbox": [
|
| 1322 |
+
174,
|
| 1323 |
+
621,
|
| 1324 |
+
825,
|
| 1325 |
+
664
|
| 1326 |
+
],
|
| 1327 |
+
"page_idx": 10
|
| 1328 |
+
},
|
| 1329 |
+
{
|
| 1330 |
+
"type": "text",
|
| 1331 |
+
"text": "Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), 2018c. ",
|
| 1332 |
+
"bbox": [
|
| 1333 |
+
174,
|
| 1334 |
+
672,
|
| 1335 |
+
825,
|
| 1336 |
+
715
|
| 1337 |
+
],
|
| 1338 |
+
"page_idx": 10
|
| 1339 |
+
},
|
| 1340 |
+
{
|
| 1341 |
+
"type": "text",
|
| 1342 |
+
"text": "Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8690–8699, 2021. ",
|
| 1343 |
+
"bbox": [
|
| 1344 |
+
173,
|
| 1345 |
+
724,
|
| 1346 |
+
826,
|
| 1347 |
+
768
|
| 1348 |
+
],
|
| 1349 |
+
"page_idx": 10
|
| 1350 |
+
},
|
| 1351 |
+
{
|
| 1352 |
+
"type": "text",
|
| 1353 |
+
"text": "Oren Nuriel, Sagie Benaim, and Lior Wolf. Permuted adain: Reducing the bias towards global statistics in image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 9482–9491, 2021. ",
|
| 1354 |
+
"bbox": [
|
| 1355 |
+
174,
|
| 1356 |
+
776,
|
| 1357 |
+
823,
|
| 1358 |
+
820
|
| 1359 |
+
],
|
| 1360 |
+
"page_idx": 10
|
| 1361 |
+
},
|
| 1362 |
+
{
|
| 1363 |
+
"type": "text",
|
| 1364 |
+
"text": "Xingang Pan, Ping Luo, Jianping Shi, and Xiaoou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. ",
|
| 1365 |
+
"bbox": [
|
| 1366 |
+
173,
|
| 1367 |
+
829,
|
| 1368 |
+
823,
|
| 1369 |
+
872
|
| 1370 |
+
],
|
| 1371 |
+
"page_idx": 10
|
| 1372 |
+
},
|
| 1373 |
+
{
|
| 1374 |
+
"type": "text",
|
| 1375 |
+
"text": "Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In Proceedings of the 37th International Conference on Machine Learning, volume 119, pp. 7728–7738, 2020. ",
|
| 1376 |
+
"bbox": [
|
| 1377 |
+
176,
|
| 1378 |
+
881,
|
| 1379 |
+
825,
|
| 1380 |
+
924
|
| 1381 |
+
],
|
| 1382 |
+
"page_idx": 10
|
| 1383 |
+
},
|
| 1384 |
+
{
|
| 1385 |
+
"type": "text",
|
| 1386 |
+
"text": "Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European Conference on Computer Vision (ECCV), pp. 102–118, 2016. ",
|
| 1387 |
+
"bbox": [
|
| 1388 |
+
171,
|
| 1389 |
+
103,
|
| 1390 |
+
825,
|
| 1391 |
+
133
|
| 1392 |
+
],
|
| 1393 |
+
"page_idx": 11
|
| 1394 |
+
},
|
| 1395 |
+
{
|
| 1396 |
+
"type": "text",
|
| 1397 |
+
"text": "Ergys Ristani, Francesco Solera, Roger S. Zou, R. Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In ECCV Workshops, 2016. ",
|
| 1398 |
+
"bbox": [
|
| 1399 |
+
173,
|
| 1400 |
+
141,
|
| 1401 |
+
823,
|
| 1402 |
+
171
|
| 1403 |
+
],
|
| 1404 |
+
"page_idx": 11
|
| 1405 |
+
},
|
| 1406 |
+
{
|
| 1407 |
+
"type": "text",
|
| 1408 |
+
"text": "Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In The European Conference on Computer Vision (ECCV), pp. 68–83, 2020. ",
|
| 1409 |
+
"bbox": [
|
| 1410 |
+
176,
|
| 1411 |
+
179,
|
| 1412 |
+
821,
|
| 1413 |
+
223
|
| 1414 |
+
],
|
| 1415 |
+
"page_idx": 11
|
| 1416 |
+
},
|
| 1417 |
+
{
|
| 1418 |
+
"type": "text",
|
| 1419 |
+
"text": "Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. Generalizing across domains via cross-gradient training. In International Conference on Learning Representations, 2018. ",
|
| 1420 |
+
"bbox": [
|
| 1421 |
+
173,
|
| 1422 |
+
232,
|
| 1423 |
+
825,
|
| 1424 |
+
275
|
| 1425 |
+
],
|
| 1426 |
+
"page_idx": 11
|
| 1427 |
+
},
|
| 1428 |
+
{
|
| 1429 |
+
"type": "text",
|
| 1430 |
+
"text": "Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. ",
|
| 1431 |
+
"bbox": [
|
| 1432 |
+
174,
|
| 1433 |
+
284,
|
| 1434 |
+
823,
|
| 1435 |
+
314
|
| 1436 |
+
],
|
| 1437 |
+
"page_idx": 11
|
| 1438 |
+
},
|
| 1439 |
+
{
|
| 1440 |
+
"type": "text",
|
| 1441 |
+
"text": "Zheyan Shen, Jiashuo Liu, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. Towards out-of-distribution generalization: A survey. arXiv preprint arXiv:2108.13624, 2021. ",
|
| 1442 |
+
"bbox": [
|
| 1443 |
+
173,
|
| 1444 |
+
323,
|
| 1445 |
+
825,
|
| 1446 |
+
352
|
| 1447 |
+
],
|
| 1448 |
+
"page_idx": 11
|
| 1449 |
+
},
|
| 1450 |
+
{
|
| 1451 |
+
"type": "text",
|
| 1452 |
+
"text": "Yichun Shi and Anil K. Jain. Probabilistic face embeddings. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. ",
|
| 1453 |
+
"bbox": [
|
| 1454 |
+
173,
|
| 1455 |
+
361,
|
| 1456 |
+
825,
|
| 1457 |
+
391
|
| 1458 |
+
],
|
| 1459 |
+
"page_idx": 11
|
| 1460 |
+
},
|
| 1461 |
+
{
|
| 1462 |
+
"type": "text",
|
| 1463 |
+
"text": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929–1958, 2014. ",
|
| 1464 |
+
"bbox": [
|
| 1465 |
+
174,
|
| 1466 |
+
400,
|
| 1467 |
+
825,
|
| 1468 |
+
443
|
| 1469 |
+
],
|
| 1470 |
+
"page_idx": 11
|
| 1471 |
+
},
|
| 1472 |
+
{
|
| 1473 |
+
"type": "text",
|
| 1474 |
+
"text": "Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. ",
|
| 1475 |
+
"bbox": [
|
| 1476 |
+
168,
|
| 1477 |
+
452,
|
| 1478 |
+
825,
|
| 1479 |
+
481
|
| 1480 |
+
],
|
| 1481 |
+
"page_idx": 11
|
| 1482 |
+
},
|
| 1483 |
+
{
|
| 1484 |
+
"type": "text",
|
| 1485 |
+
"text": "V. Vapnik. Principles of risk minimization for learning theory. In Advances in Neural Information Processing Systems, volume 4, 1992. ",
|
| 1486 |
+
"bbox": [
|
| 1487 |
+
171,
|
| 1488 |
+
489,
|
| 1489 |
+
823,
|
| 1490 |
+
520
|
| 1491 |
+
],
|
| 1492 |
+
"page_idx": 11
|
| 1493 |
+
},
|
| 1494 |
+
{
|
| 1495 |
+
"type": "text",
|
| 1496 |
+
"text": "Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 10(5):988–999, 1999. ",
|
| 1497 |
+
"bbox": [
|
| 1498 |
+
169,
|
| 1499 |
+
529,
|
| 1500 |
+
823,
|
| 1501 |
+
558
|
| 1502 |
+
],
|
| 1503 |
+
"page_idx": 11
|
| 1504 |
+
},
|
| 1505 |
+
{
|
| 1506 |
+
"type": "text",
|
| 1507 |
+
"text": "Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. ",
|
| 1508 |
+
"bbox": [
|
| 1509 |
+
173,
|
| 1510 |
+
568,
|
| 1511 |
+
823,
|
| 1512 |
+
611
|
| 1513 |
+
],
|
| 1514 |
+
"page_idx": 11
|
| 1515 |
+
},
|
| 1516 |
+
{
|
| 1517 |
+
"type": "text",
|
| 1518 |
+
"text": "Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David LopezPaz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In Proceedings of the 36th International Conference on Machine Learning, volume 97, pp. 6438– 6447, 2019. ",
|
| 1519 |
+
"bbox": [
|
| 1520 |
+
174,
|
| 1521 |
+
619,
|
| 1522 |
+
825,
|
| 1523 |
+
676
|
| 1524 |
+
],
|
| 1525 |
+
"page_idx": 11
|
| 1526 |
+
},
|
| 1527 |
+
{
|
| 1528 |
+
"type": "text",
|
| 1529 |
+
"text": "Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), 2020a. ",
|
| 1530 |
+
"bbox": [
|
| 1531 |
+
173,
|
| 1532 |
+
685,
|
| 1533 |
+
825,
|
| 1534 |
+
729
|
| 1535 |
+
],
|
| 1536 |
+
"page_idx": 11
|
| 1537 |
+
},
|
| 1538 |
+
{
|
| 1539 |
+
"type": "text",
|
| 1540 |
+
"text": "Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Wenjun Zeng, and Tao Qin. Generalizing to unseen domains: A survey on domain generalization. arXiv preprint arXiv:2103.03097, 2021. ",
|
| 1541 |
+
"bbox": [
|
| 1542 |
+
171,
|
| 1543 |
+
738,
|
| 1544 |
+
825,
|
| 1545 |
+
767
|
| 1546 |
+
],
|
| 1547 |
+
"page_idx": 11
|
| 1548 |
+
},
|
| 1549 |
+
{
|
| 1550 |
+
"type": "text",
|
| 1551 |
+
"text": "Ximei Wang, Ying Jin, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Transferable normalization: Towards improving transferability of deep neural networks. In Advances in Neural Information Processing Systems, volume 32, 2019a. ",
|
| 1552 |
+
"bbox": [
|
| 1553 |
+
174,
|
| 1554 |
+
776,
|
| 1555 |
+
823,
|
| 1556 |
+
819
|
| 1557 |
+
],
|
| 1558 |
+
"page_idx": 11
|
| 1559 |
+
},
|
| 1560 |
+
{
|
| 1561 |
+
"type": "text",
|
| 1562 |
+
"text": "Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. In Advances in Neural Information Processing Systems, volume 33, pp. 19212–19223, 2020b. ",
|
| 1563 |
+
"bbox": [
|
| 1564 |
+
174,
|
| 1565 |
+
829,
|
| 1566 |
+
823,
|
| 1567 |
+
872
|
| 1568 |
+
],
|
| 1569 |
+
"page_idx": 11
|
| 1570 |
+
},
|
| 1571 |
+
{
|
| 1572 |
+
"type": "text",
|
| 1573 |
+
"text": "Yulin Wang, Xuran Pan, Shiji Song, Hong Zhang, Gao Huang, and Cheng Wu. Implicit semantic data augmentation for deep networks. In Advances in Neural Information Processing Systems, pp. 12614–12623, 2019b. ",
|
| 1574 |
+
"bbox": [
|
| 1575 |
+
176,
|
| 1576 |
+
881,
|
| 1577 |
+
825,
|
| 1578 |
+
924
|
| 1579 |
+
],
|
| 1580 |
+
"page_idx": 11
|
| 1581 |
+
},
|
| 1582 |
+
{
|
| 1583 |
+
"type": "text",
|
| 1584 |
+
"text": "Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual representation learning via random convolutions. In International Conference on Learning Representations, 2021. ",
|
| 1585 |
+
"bbox": [
|
| 1586 |
+
174,
|
| 1587 |
+
103,
|
| 1588 |
+
823,
|
| 1589 |
+
146
|
| 1590 |
+
],
|
| 1591 |
+
"page_idx": 12
|
| 1592 |
+
},
|
| 1593 |
+
{
|
| 1594 |
+
"type": "text",
|
| 1595 |
+
"text": "Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 6022–6031, 2019. ",
|
| 1596 |
+
"bbox": [
|
| 1597 |
+
178,
|
| 1598 |
+
155,
|
| 1599 |
+
823,
|
| 1600 |
+
198
|
| 1601 |
+
],
|
| 1602 |
+
"page_idx": 12
|
| 1603 |
+
},
|
| 1604 |
+
{
|
| 1605 |
+
"type": "text",
|
| 1606 |
+
"text": "Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. ",
|
| 1607 |
+
"bbox": [
|
| 1608 |
+
173,
|
| 1609 |
+
205,
|
| 1610 |
+
820,
|
| 1611 |
+
236
|
| 1612 |
+
],
|
| 1613 |
+
"page_idx": 12
|
| 1614 |
+
},
|
| 1615 |
+
{
|
| 1616 |
+
"type": "text",
|
| 1617 |
+
"text": "Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1116–1124, 2015. ",
|
| 1618 |
+
"bbox": [
|
| 1619 |
+
176,
|
| 1620 |
+
243,
|
| 1621 |
+
823,
|
| 1622 |
+
286
|
| 1623 |
+
],
|
| 1624 |
+
"page_idx": 12
|
| 1625 |
+
},
|
| 1626 |
+
{
|
| 1627 |
+
"type": "text",
|
| 1628 |
+
"text": "Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Learning to generate novel domains for domain generalization. In The European Conference on Computer Vision (ECCV), pp. 561–578, 2020a. ",
|
| 1629 |
+
"bbox": [
|
| 1630 |
+
171,
|
| 1631 |
+
295,
|
| 1632 |
+
825,
|
| 1633 |
+
338
|
| 1634 |
+
],
|
| 1635 |
+
"page_idx": 12
|
| 1636 |
+
},
|
| 1637 |
+
{
|
| 1638 |
+
"type": "text",
|
| 1639 |
+
"text": "Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain adaptive ensemble learning. arXiv preprint arXiv:2003.07325, 2020b. ",
|
| 1640 |
+
"bbox": [
|
| 1641 |
+
173,
|
| 1642 |
+
347,
|
| 1643 |
+
823,
|
| 1644 |
+
376
|
| 1645 |
+
],
|
| 1646 |
+
"page_idx": 12
|
| 1647 |
+
},
|
| 1648 |
+
{
|
| 1649 |
+
"type": "text",
|
| 1650 |
+
"text": "Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. arXiv preprint arXiv:2103.02503, 2021a. ",
|
| 1651 |
+
"bbox": [
|
| 1652 |
+
173,
|
| 1653 |
+
385,
|
| 1654 |
+
821,
|
| 1655 |
+
414
|
| 1656 |
+
],
|
| 1657 |
+
"page_idx": 12
|
| 1658 |
+
},
|
| 1659 |
+
{
|
| 1660 |
+
"type": "text",
|
| 1661 |
+
"text": "Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. In International Conference on Learning Representations, 2021b. ",
|
| 1662 |
+
"bbox": [
|
| 1663 |
+
173,
|
| 1664 |
+
422,
|
| 1665 |
+
825,
|
| 1666 |
+
452
|
| 1667 |
+
],
|
| 1668 |
+
"page_idx": 12
|
| 1669 |
+
},
|
| 1670 |
+
{
|
| 1671 |
+
"type": "text",
|
| 1672 |
+
"text": "A APPENDIX ",
|
| 1673 |
+
"text_level": 1,
|
| 1674 |
+
"bbox": [
|
| 1675 |
+
176,
|
| 1676 |
+
102,
|
| 1677 |
+
299,
|
| 1678 |
+
118
|
| 1679 |
+
],
|
| 1680 |
+
"page_idx": 13
|
| 1681 |
+
},
|
| 1682 |
+
{
|
| 1683 |
+
"type": "text",
|
| 1684 |
+
"text": "A.1 ALGORITHM ",
|
| 1685 |
+
"text_level": 1,
|
| 1686 |
+
"bbox": [
|
| 1687 |
+
174,
|
| 1688 |
+
136,
|
| 1689 |
+
307,
|
| 1690 |
+
151
|
| 1691 |
+
],
|
| 1692 |
+
"page_idx": 13
|
| 1693 |
+
},
|
| 1694 |
+
{
|
| 1695 |
+
"type": "text",
|
| 1696 |
+
"text": "The algorithm of the proposed method is illustrated in Algorithm 1. ",
|
| 1697 |
+
"bbox": [
|
| 1698 |
+
173,
|
| 1699 |
+
165,
|
| 1700 |
+
616,
|
| 1701 |
+
180
|
| 1702 |
+
],
|
| 1703 |
+
"page_idx": 13
|
| 1704 |
+
},
|
| 1705 |
+
{
|
| 1706 |
+
"type": "text",
|
| 1707 |
+
"text": "Algorithm 1: The algorithm of the proposed method (DSU) ",
|
| 1708 |
+
"text_level": 1,
|
| 1709 |
+
"bbox": [
|
| 1710 |
+
173,
|
| 1711 |
+
203,
|
| 1712 |
+
570,
|
| 1713 |
+
218
|
| 1714 |
+
],
|
| 1715 |
+
"page_idx": 13
|
| 1716 |
+
},
|
| 1717 |
+
{
|
| 1718 |
+
"type": "text",
|
| 1719 |
+
"text": "Input: Intermediate feature $\\boldsymbol { x } \\in \\mathbb { R } ^ { B \\times C \\times H \\times W }$ , probability $p$ to forward this module; \nOutput: Intermediate feature $\\widehat { x } \\in \\mathbb { R } ^ { B \\times C \\times H \\times W }$ after considering potential statistics shifts; \n1 Sample $p _ { 0 } \\sim U ( 0 , 1 )$ ; \n2 if $p _ { 0 } < p$ and Training then \n3 Compute the channel-wise mean and standard deviation of each instance in a mini-batch; \n4 $\\begin{array} { l } { \\displaystyle \\mu ( x ) = \\frac { 1 } { H W } \\sum _ { h = 1 } ^ { H } \\sum _ { w = 1 } ^ { W } x _ { b , c , h , w } , } \\\\ { \\sigma ^ { 2 } ( x ) = \\frac { 1 } { H W } \\sum _ { h = 1 } ^ { H } \\sum _ { w = 1 } ^ { W } ( x _ { b , c , h , w } - \\mu ( x ) ) ^ { 2 } . } \\end{array}$ \n5 \n6 Uncertainty estimation on feature statistics; \n7 $\\begin{array} { r l } & { \\Sigma _ { \\mu } ^ { 2 } ( x ) = \\frac { 1 } { B } \\displaystyle \\sum _ { b = 1 } ^ { B } ( \\mu _ { b c } ( x ) - E _ { b } ( \\mu _ { b c } ( x ) ) ) ^ { 2 } , } \\\\ & { \\Sigma _ { \\sigma } ^ { 2 } ( x ) = \\frac { 1 } { B } \\displaystyle \\sum _ { b = 1 } ^ { B } ( \\sigma _ { b c } ( x ) - E _ { b } ( \\sigma _ { b c } ( x ) ) ) ^ { 2 } . } \\end{array}$ \n8 \n9 Compute the synthetic feature statistics randomly sampling from the given Guassian distributions; \n10 $\\begin{array} { r l } & { \\beta ( x ) = \\mu ( x ) + \\epsilon _ { \\mu } \\Sigma _ { \\sigma } ( x ) , \\epsilon _ { \\mu } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\mathbf { 1 } ) , } \\\\ & { \\gamma ( x ) = \\sigma ( x ) + \\epsilon _ { \\sigma } \\Sigma _ { \\sigma } ( x ) , \\epsilon _ { \\sigma } \\sim \\mathcal { N } ( \\mathbf { 0 } , \\mathbf { 1 } ) . } \\end{array}$ , \n11 \n12 Obtain the feature after considering potential statistics shifts; \n13 $\\begin{array} { r } { \\widehat { x } = \\gamma ( x ) \\times \\frac { x - \\mu ( x ) } { \\sigma ( x ) } + \\beta ( x ) . } \\end{array}$ \n14 return the feature $\\widehat { x }$ with uncertain feature statistics. ",
|
| 1720 |
+
"bbox": [
|
| 1721 |
+
158,
|
| 1722 |
+
219,
|
| 1723 |
+
797,
|
| 1724 |
+
530
|
| 1725 |
+
],
|
| 1726 |
+
"page_idx": 13
|
| 1727 |
+
},
|
| 1728 |
+
{
|
| 1729 |
+
"type": "text",
|
| 1730 |
+
"text": "15 else ",
|
| 1731 |
+
"text_level": 1,
|
| 1732 |
+
"bbox": [
|
| 1733 |
+
156,
|
| 1734 |
+
532,
|
| 1735 |
+
200,
|
| 1736 |
+
544
|
| 1737 |
+
],
|
| 1738 |
+
"page_idx": 13
|
| 1739 |
+
},
|
| 1740 |
+
{
|
| 1741 |
+
"type": "text",
|
| 1742 |
+
"text": "adopt the original feature $x$ and skip this module. ",
|
| 1743 |
+
"bbox": [
|
| 1744 |
+
179,
|
| 1745 |
+
546,
|
| 1746 |
+
488,
|
| 1747 |
+
560
|
| 1748 |
+
],
|
| 1749 |
+
"page_idx": 13
|
| 1750 |
+
},
|
| 1751 |
+
{
|
| 1752 |
+
"type": "text",
|
| 1753 |
+
"text": "17 end ",
|
| 1754 |
+
"bbox": [
|
| 1755 |
+
156,
|
| 1756 |
+
563,
|
| 1757 |
+
199,
|
| 1758 |
+
574
|
| 1759 |
+
],
|
| 1760 |
+
"page_idx": 13
|
| 1761 |
+
},
|
| 1762 |
+
{
|
| 1763 |
+
"type": "text",
|
| 1764 |
+
"text": "A.2 MULTI-DOMAIN GENERALIZATION ON OFFICE HOME. ",
|
| 1765 |
+
"text_level": 1,
|
| 1766 |
+
"bbox": [
|
| 1767 |
+
173,
|
| 1768 |
+
618,
|
| 1769 |
+
588,
|
| 1770 |
+
633
|
| 1771 |
+
],
|
| 1772 |
+
"page_idx": 13
|
| 1773 |
+
},
|
| 1774 |
+
{
|
| 1775 |
+
"type": "text",
|
| 1776 |
+
"text": "In addition to multi-domain classification experiments on PACS, we further evaluate the effectiveness of the proposed method on Office-Home (Venkateswara et al., 2017), which contains 15,500 images of 65 classes for home and office recognition. The experiment results with ResNet18 backbone are shown in Table 7. It can be observed that our method brings obvious improvement over the baseline method and also outperforms the competing methods. By introducing the feature statistics uncertainty, the models trained with our method can learn to alleviate the domain perturbations, such as the style information, and obtain more domain-invariant features. For example, huge improvement can be observed from the results on Clipart, which is a domain with much different style from others. ",
|
| 1777 |
+
"bbox": [
|
| 1778 |
+
173,
|
| 1779 |
+
646,
|
| 1780 |
+
825,
|
| 1781 |
+
772
|
| 1782 |
+
],
|
| 1783 |
+
"page_idx": 13
|
| 1784 |
+
},
|
| 1785 |
+
{
|
| 1786 |
+
"type": "table",
|
| 1787 |
+
"img_path": "images/7a3dd849f039730bd5c03528dcaba2f444a31b6e31f2880b9b0e75a203b6c315.jpg",
|
| 1788 |
+
"table_caption": [
|
| 1789 |
+
"Table 7: Experiment results of Office-Home multi-domain classification task. "
|
| 1790 |
+
],
|
| 1791 |
+
"table_footnote": [],
|
| 1792 |
+
"table_body": "<table><tr><td>Method</td><td>Reference</td><td>Art</td><td>Clipart</td><td>Product</td><td>Real</td><td>Average (%)</td></tr><tr><td>Baseline</td><td></td><td>58.8</td><td>48.3</td><td>74.2</td><td>76.2</td><td>64.4</td></tr><tr><td>Mixup (Zhang et al.,2018)</td><td>ICLR 2018</td><td>58.2</td><td>49.3</td><td>74.7</td><td>76.1</td><td>64.6</td></tr><tr><td>CrossGrad (Shankar et al., 2018)</td><td>ICLR 2018</td><td>58.4</td><td>49.4</td><td>73.9</td><td>75.8</td><td>64.4</td></tr><tr><td>Manifold Mixup (Verma et al., 2019)</td><td>ICML 2019</td><td>56.2</td><td>46.3</td><td>73.6</td><td>75.2</td><td>62.8</td></tr><tr><td>CutMix (Yun et al., 2019)</td><td>ICCV 2019</td><td>57.9</td><td>48.3</td><td>74.5</td><td>75.6</td><td>64.1</td></tr><tr><td>RSC (Huang et al.,2020)</td><td>ECCV2020</td><td>58.4</td><td>47.9</td><td>71.6</td><td>74.5</td><td>63.1</td></tr><tr><td>L2A-OT (Zhou et al., 2020a)</td><td>ECCV2020</td><td>60.6</td><td>50.1</td><td>74.8</td><td>77.0</td><td>65.6</td></tr><tr><td>MixStyle (Zhou et al.,2021b)</td><td>ICLR 2021</td><td>58.7</td><td>53.4</td><td>74.2</td><td>75.9</td><td>65.5</td></tr><tr><td>DSU</td><td>Ours</td><td>60.2</td><td>54.8</td><td>74.1</td><td>75.1</td><td>66.1</td></tr></table>",
|
| 1793 |
+
"bbox": [
|
| 1794 |
+
174,
|
| 1795 |
+
804,
|
| 1796 |
+
818,
|
| 1797 |
+
939
|
| 1798 |
+
],
|
| 1799 |
+
"page_idx": 13
|
| 1800 |
+
},
|
| 1801 |
+
{
|
| 1802 |
+
"type": "text",
|
| 1803 |
+
"text": "A.3 CHOICE OF UNCERTAINTY DISTRIBUTION ",
|
| 1804 |
+
"text_level": 1,
|
| 1805 |
+
"bbox": [
|
| 1806 |
+
176,
|
| 1807 |
+
103,
|
| 1808 |
+
503,
|
| 1809 |
+
117
|
| 1810 |
+
],
|
| 1811 |
+
"page_idx": 14
|
| 1812 |
+
},
|
| 1813 |
+
{
|
| 1814 |
+
"type": "table",
|
| 1815 |
+
"img_path": "images/0084e73431eff7fbed41d0d6aae726b4f3c20933281f86186cf37efeea98cd24.jpg",
|
| 1816 |
+
"table_caption": [
|
| 1817 |
+
"Table 8: Intensive study about different vanilla Gaussian distributions with pre-defined standard deviation. "
|
| 1818 |
+
],
|
| 1819 |
+
"table_footnote": [],
|
| 1820 |
+
"table_body": "<table><tr><td>Choice</td><td>Baseline</td><td>Rand(10°)</td><td>Rand(10-1)</td><td>Rand(10-2)</td><td>Rand(10-3)</td><td>DSU</td></tr><tr><td>PACS</td><td>79.0</td><td>76.9</td><td>81.2</td><td>79.3</td><td>79.1</td><td>84.1</td></tr><tr><td>GTA5 to Cityscapes</td><td>37.0</td><td>38.2</td><td>39.8</td><td>40.1</td><td>38.9</td><td>43.1</td></tr></table>",
|
| 1821 |
+
"bbox": [
|
| 1822 |
+
176,
|
| 1823 |
+
165,
|
| 1824 |
+
823,
|
| 1825 |
+
213
|
| 1826 |
+
],
|
| 1827 |
+
"page_idx": 14
|
| 1828 |
+
},
|
| 1829 |
+
{
|
| 1830 |
+
"type": "text",
|
| 1831 |
+
"text": "Besides the analysis of the uncertainty estimation in the ablation study, we also conduct a more intensive study about the effects of pre-defined uncertainty estimations. Specifically, Rand $( s )$ denotes directly imposing random shifts draw from a fixed Gaussian ${ \\mathcal { N } } ( 0 , s ^ { 2 } )$ . The intensive study is shown in Table 8. It can be observed that the results of different fixed distributions are all much lower than the proposed method. Some conclusions could be obtained from the results. (a): Imposing excessive uncertainty might harm the model training and degrade the performance. (b): The best fixed value of uncertainty estimation might vary from different tasks. By contrast, the proposed method can be adaptive to different tasks without any manual adjustment. ",
|
| 1832 |
+
"bbox": [
|
| 1833 |
+
173,
|
| 1834 |
+
215,
|
| 1835 |
+
825,
|
| 1836 |
+
327
|
| 1837 |
+
],
|
| 1838 |
+
"page_idx": 14
|
| 1839 |
+
},
|
| 1840 |
+
{
|
| 1841 |
+
"type": "table",
|
| 1842 |
+
"img_path": "images/1766c5920c267c8019a411a212bce1c1a33acd29754ab5e43bce6ec2cd9866e8.jpg",
|
| 1843 |
+
"table_caption": [
|
| 1844 |
+
"Table 9: Study about the effects of sharing the same uncertain distribution among different channels. "
|
| 1845 |
+
],
|
| 1846 |
+
"table_footnote": [],
|
| 1847 |
+
"table_body": "<table><tr><td>Choice</td><td>Baseline</td><td>Channel-share</td><td>DSU</td></tr><tr><td>PACS</td><td>79.0</td><td>80.2</td><td>84.1</td></tr><tr><td>GTA5 to Cityscapes</td><td>37.0</td><td>39.3</td><td>43.1</td></tr></table>",
|
| 1848 |
+
"bbox": [
|
| 1849 |
+
299,
|
| 1850 |
+
352,
|
| 1851 |
+
696,
|
| 1852 |
+
400
|
| 1853 |
+
],
|
| 1854 |
+
"page_idx": 14
|
| 1855 |
+
},
|
| 1856 |
+
{
|
| 1857 |
+
"type": "text",
|
| 1858 |
+
"text": "We also conduct the experiment to test the effectiveness of treating different channels with different potentials. Channel-share denotes all channels of the sample share the same uncertainty distribution, i.e., using the average uncertainty estimation among channels. As shown in Table 9, the results indicate that sharing the same uncertain distribution among different channels is less effective, which ignores the different potentials of channels and will limit their performances. Meanwhile, the proposed method explicitly considers the different potentials of different channels and brings better performances. ",
|
| 1859 |
+
"bbox": [
|
| 1860 |
+
173,
|
| 1861 |
+
401,
|
| 1862 |
+
825,
|
| 1863 |
+
497
|
| 1864 |
+
],
|
| 1865 |
+
"page_idx": 14
|
| 1866 |
+
},
|
| 1867 |
+
{
|
| 1868 |
+
"type": "text",
|
| 1869 |
+
"text": "A.4 T-SNE VISUALIZATION ",
|
| 1870 |
+
"text_level": 1,
|
| 1871 |
+
"bbox": [
|
| 1872 |
+
176,
|
| 1873 |
+
515,
|
| 1874 |
+
380,
|
| 1875 |
+
529
|
| 1876 |
+
],
|
| 1877 |
+
"page_idx": 14
|
| 1878 |
+
},
|
| 1879 |
+
{
|
| 1880 |
+
"type": "text",
|
| 1881 |
+
"text": "To analyze the effects on feature representations, we visualize the feature representation vectors of different categories in unseen domain with t-SNE (Van der Maaten & Hinton, 2008) in Figure 7. The features of the same category become more compact benefiting from the proposed method. Because our method can alleviate the domain perturbations during training and make the model focus on content information, obtaining more invariant features representations. ",
|
| 1882 |
+
"bbox": [
|
| 1883 |
+
174,
|
| 1884 |
+
540,
|
| 1885 |
+
547,
|
| 1886 |
+
665
|
| 1887 |
+
],
|
| 1888 |
+
"page_idx": 14
|
| 1889 |
+
},
|
| 1890 |
+
{
|
| 1891 |
+
"type": "image",
|
| 1892 |
+
"img_path": "images/a2062cc013e717f391a1cbd3b216c94f5d01678c5b8c0e440372bf8dbc599b39.jpg",
|
| 1893 |
+
"image_caption": [
|
| 1894 |
+
"Figure 7: The t-SNE visualization on unseen PACS domain. "
|
| 1895 |
+
],
|
| 1896 |
+
"image_footnote": [],
|
| 1897 |
+
"bbox": [
|
| 1898 |
+
562,
|
| 1899 |
+
535,
|
| 1900 |
+
823,
|
| 1901 |
+
618
|
| 1902 |
+
],
|
| 1903 |
+
"page_idx": 14
|
| 1904 |
+
},
|
| 1905 |
+
{
|
| 1906 |
+
"type": "text",
|
| 1907 |
+
"text": "A.5 COMPARISONS TO THE RELATED METHODS ",
|
| 1908 |
+
"text_level": 1,
|
| 1909 |
+
"bbox": [
|
| 1910 |
+
174,
|
| 1911 |
+
683,
|
| 1912 |
+
516,
|
| 1913 |
+
695
|
| 1914 |
+
],
|
| 1915 |
+
"page_idx": 14
|
| 1916 |
+
},
|
| 1917 |
+
{
|
| 1918 |
+
"type": "text",
|
| 1919 |
+
"text": "Some related methods (Zhou et al., 2021b; Nuriel et al., 2021) also tackle the domain generalization problem by producing synthetic feature statistics. Specifically, we denote the random shuffle copies of the batch feature as $\\widehat { \\boldsymbol { x } } = \\mathrm { s h u f f l e } ( \\boldsymbol { x } )$ . pAdaIN (Nuriel et al. (2021)) generate new samples by bswapping feature statistics between the batch samples applied with a random permutation, where $\\beta ( x ) { \\overset { \\vartriangle } { = } } \\mu ( { \\widehat { x } } )$ and $\\gamma ( x ) ~ = ~ \\sigma ( { \\widehat { x } } )$ . MixStyle (Zhou et al. (2021b)) generates synthesized domain b bsamples by mixing feature statistics information of two instances, where $\\beta ( x ) \\dot { } = \\lambda \\mu ( x ) + ( 1 -$ $\\lambda ) \\mu ( \\widehat { x } )$ and $\\gamma ( x ) \\overset { - } { = } \\lambda \\sigma ( x ) + ( 1 - \\lambda ) \\sigma ( \\widehat { x } )$ and $\\lambda \\in ( 0 , 1 )$ is a random interpolation weight. ",
|
| 1920 |
+
"bbox": [
|
| 1921 |
+
173,
|
| 1922 |
+
708,
|
| 1923 |
+
825,
|
| 1924 |
+
806
|
| 1925 |
+
],
|
| 1926 |
+
"page_idx": 14
|
| 1927 |
+
},
|
| 1928 |
+
{
|
| 1929 |
+
"type": "text",
|
| 1930 |
+
"text": "Despite the success, they typically adopt linear manipulation on pairwise samples to generate new feature statistics, which limits the diversity of synthetic changes. Specifically, the direction of the variants is determined by the chosen reference sample and the internal operation also restricts the variant intensity. Our method, not relies on a specific reference sample, is based on the Gaussian distribution that can produce not only linear changes but diverse variants with more possibilities. Due to the boundless range of the Gaussian distribution, our method has the ability to generate feature statistics beyond the scope of training domain, which also breaks the limitation of inner interpolation between training samples. The visualization of the comparisons is shown in Figure 8. ",
|
| 1931 |
+
"bbox": [
|
| 1932 |
+
174,
|
| 1933 |
+
811,
|
| 1934 |
+
825,
|
| 1935 |
+
924
|
| 1936 |
+
],
|
| 1937 |
+
"page_idx": 14
|
| 1938 |
+
},
|
| 1939 |
+
{
|
| 1940 |
+
"type": "image",
|
| 1941 |
+
"img_path": "images/6e20dcee25376ad8cc877cac1674a188ed4b4186141d52f56fc7c69b8922d085.jpg",
|
| 1942 |
+
"image_caption": [
|
| 1943 |
+
"Figure 8: Comparisons with related methods. The variants produced by previous pairwise-based methods are restricted by the combination of chosen sample pair, while our method can generate various feature statistics variants with different combination of directions and intensities. "
|
| 1944 |
+
],
|
| 1945 |
+
"image_footnote": [],
|
| 1946 |
+
"bbox": [
|
| 1947 |
+
173,
|
| 1948 |
+
102,
|
| 1949 |
+
815,
|
| 1950 |
+
281
|
| 1951 |
+
],
|
| 1952 |
+
"page_idx": 15
|
| 1953 |
+
},
|
| 1954 |
+
{
|
| 1955 |
+
"type": "text",
|
| 1956 |
+
"text": "A.6 WITHIN-DATASET PERFORMANCE ",
|
| 1957 |
+
"text_level": 1,
|
| 1958 |
+
"bbox": [
|
| 1959 |
+
176,
|
| 1960 |
+
369,
|
| 1961 |
+
450,
|
| 1962 |
+
383
|
| 1963 |
+
],
|
| 1964 |
+
"page_idx": 15
|
| 1965 |
+
},
|
| 1966 |
+
{
|
| 1967 |
+
"type": "text",
|
| 1968 |
+
"text": "In Table 4, we tested the within-dataset performance on the large-scale dataset ImageNet, denoted as ”Clean”. We observed that the top-1 error rate declines from $2 3 . 8 \\%$ to $2 3 . 4 \\%$ after training with the proposed DSU, indicating that DSU does not sacrifice the in-domain performance to gain the benefits on out-of-distribution domains.The reason might be that instances in the testing set may not always fall into exactly the same distribution of the training set, and they still have slight statistic shifts (Gao et al., 2021a). The proposed DSU can help the trained model improve the robustness to statistics shifts and thus gain better performance in within-dataset ImageNet. ",
|
| 1969 |
+
"bbox": [
|
| 1970 |
+
173,
|
| 1971 |
+
395,
|
| 1972 |
+
825,
|
| 1973 |
+
494
|
| 1974 |
+
],
|
| 1975 |
+
"page_idx": 15
|
| 1976 |
+
},
|
| 1977 |
+
{
|
| 1978 |
+
"type": "text",
|
| 1979 |
+
"text": "Besides the experiments on ImageNet, we also supplement the within-dataset performances on PACS. According to the multi-source training protocol on PACS (Li et al., 2017), the within-domain performance is averaged over multiple training-domain datasets (P,A,C,S denotes Art, Cartoon, Photo and Sketch respectively). As shown in Table 10, it can be observed that the within-dataset performance of our DSU also slightly beats the baseline, verifying the conclusion as on ImageNet. ",
|
| 1980 |
+
"bbox": [
|
| 1981 |
+
173,
|
| 1982 |
+
501,
|
| 1983 |
+
825,
|
| 1984 |
+
570
|
| 1985 |
+
],
|
| 1986 |
+
"page_idx": 15
|
| 1987 |
+
},
|
| 1988 |
+
{
|
| 1989 |
+
"type": "table",
|
| 1990 |
+
"img_path": "images/135c095680bc6837df2842654d522b781753cb19b1d312f48ab916b803b8e317.jpg",
|
| 1991 |
+
"table_caption": [
|
| 1992 |
+
"Table 10: Within-dataset performance on PACS. P,A,C,S denote Photo, Art Painting, Cartoon, and Sketch respectively. "
|
| 1993 |
+
],
|
| 1994 |
+
"table_footnote": [],
|
| 1995 |
+
"table_body": "<table><tr><td>Method</td><td>Reference</td><td>P,C,S</td><td>P,A,S</td><td>A,C,S</td><td>P,A,C</td><td>Average (%)</td></tr><tr><td>Baseline</td><td>1</td><td>95.70</td><td>95.28</td><td>94.22</td><td>96.58</td><td>95.44</td></tr><tr><td>DSU</td><td>Ours</td><td>96.20</td><td>96.32</td><td>95.17</td><td>97.20</td><td>96.21</td></tr></table>",
|
| 1996 |
+
"bbox": [
|
| 1997 |
+
261,
|
| 1998 |
+
625,
|
| 1999 |
+
725,
|
| 2000 |
+
671
|
| 2001 |
+
],
|
| 2002 |
+
"page_idx": 15
|
| 2003 |
+
},
|
| 2004 |
+
{
|
| 2005 |
+
"type": "text",
|
| 2006 |
+
"text": "A.7 ABLATION STUDY ON BATCH SIZE ",
|
| 2007 |
+
"text_level": 1,
|
| 2008 |
+
"bbox": [
|
| 2009 |
+
176,
|
| 2010 |
+
679,
|
| 2011 |
+
455,
|
| 2012 |
+
693
|
| 2013 |
+
],
|
| 2014 |
+
"page_idx": 15
|
| 2015 |
+
},
|
| 2016 |
+
{
|
| 2017 |
+
"type": "text",
|
| 2018 |
+
"text": "In Table 11, we conduct an ablation study on the batch size. As shown in the table, consistent performance gains are observed with various batch sizes on PACS. Note we use the batch size of 64 in our paper, following the original setting in PACS (Li et al., 2017) for fair comparison. ",
|
| 2019 |
+
"bbox": [
|
| 2020 |
+
174,
|
| 2021 |
+
704,
|
| 2022 |
+
825,
|
| 2023 |
+
747
|
| 2024 |
+
],
|
| 2025 |
+
"page_idx": 15
|
| 2026 |
+
},
|
| 2027 |
+
{
|
| 2028 |
+
"type": "table",
|
| 2029 |
+
"img_path": "images/dcf0adec0a0b8408bc2349e5363aef0364a61cd8f01c1157acbc8ad5e5b6f66a.jpg",
|
| 2030 |
+
"table_caption": [
|
| 2031 |
+
"Table 11: Ablation study on the effects of batch size. "
|
| 2032 |
+
],
|
| 2033 |
+
"table_footnote": [],
|
| 2034 |
+
"table_body": "<table><tr><td>batchsize</td><td>Reference</td><td>16</td><td>32</td><td>64</td><td>128</td><td>256</td></tr><tr><td>Baseline</td><td>1</td><td>81.0</td><td>80.2</td><td>79.0</td><td>77.8</td><td>75.6</td></tr><tr><td>DSU</td><td>Ours</td><td>84.9 (+3.9)</td><td>84.5 (+4.3)</td><td>84.1 (+5.1)</td><td>82.1 (+4.3)</td><td>80.4 (+4.8)</td></tr></table>",
|
| 2035 |
+
"bbox": [
|
| 2036 |
+
202,
|
| 2037 |
+
786,
|
| 2038 |
+
789,
|
| 2039 |
+
834
|
| 2040 |
+
],
|
| 2041 |
+
"page_idx": 15
|
| 2042 |
+
}
|
| 2043 |
+
]
|
parse/dev/6HN7LHyzGgC/6HN7LHyzGgC_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/6HN7LHyzGgC/6HN7LHyzGgC_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/6dlC7E1H_9/6dlC7E1H_9_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/6dlC7E1H_9/6dlC7E1H_9_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/89ia77nZ8u/89ia77nZ8u.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/89ia77nZ8u/89ia77nZ8u_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/89ia77nZ8u/89ia77nZ8u_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/89ia77nZ8u/89ia77nZ8u_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/9XFSbDPmdW/9XFSbDPmdW.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/9XFSbDPmdW/9XFSbDPmdW_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/9XFSbDPmdW/9XFSbDPmdW_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/A6X9y8n4sT/A6X9y8n4sT.md
ADDED
|
@@ -0,0 +1,308 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization
|
| 2 |
+
|
| 3 |
+
Chao Xu2∗ Haian Jin3,4∗ Linghao Chen1,4∗ Mukund Varma T1 Zexiang Xu6 Hao Su1
|
| 4 |
+
|
| 5 |
+
1 UC San Diego 2 UCLA 3 Cornell University 4 Zhejiang University 5 Adobe Research
|
| 6 |
+
|
| 7 |
+
Project Website: http://one-2-3-45.com
|
| 8 |
+
|
| 9 |
+
# Abstract
|
| 10 |
+
|
| 11 |
+
Single image 3D reconstruction is an important but challenging task that requires extensive knowledge of our natural world. Many existing methods solve this problem by optimizing a neural radiance field under the guidance of 2D diffusion models but suffer from lengthy optimization time, 3D inconsistency results, and poor geometry. In this work, we propose a novel method that takes a single image of any object as input and generates a full 360-degree 3D textured mesh in a single feed-forward pass. Given a single image, we first use a view-conditioned 2D diffusion model, Zero123, to generate multi-view images for the input view, and then aim to lift them up to 3D space. Since traditional reconstruction methods struggle with inconsistent multi-view predictions, we build our 3D reconstruction module upon an SDF-based generalizable neural surface reconstruction method and propose several critical training strategies to enable the reconstruction of 360- degree meshes. Without costly optimizations, our method reconstructs 3D shapes in significantly less time than existing methods. Moreover, our method favors better geometry, generates more 3D consistent results, and adheres more closely to the input image. We evaluate our approach on both synthetic data and in-the-wild images and demonstrate its superiority in terms of both mesh quality and runtime. In addition, our approach can seamlessly support the text-to-3D task by integrating with off-the-shelf text-to-image diffusion models.
|
| 12 |
+
|
| 13 |
+
# 1 Introduction
|
| 14 |
+
|
| 15 |
+
Single image 3D reconstruction, the task of reconstructing a 3D model of an object from a single 2D image, is a long-standing problem in the computer vision community and is crucial for a wide range of applications, such as robotic object manipulation and navigation, 3D content creation, as well as AR/VR [47; 9; 92]. The problem is challenging as it requires not only the reconstruction of visible parts but also the hallucination of invisible regions. Consequently, this problem is often ill-posed and corresponds to multiple plausible solutions because of insufficient evidence from a single image. On the other hand, humans can adeptly infer unseen 3D content based on our extensive knowledge of the 3D world. To endow intelligent agents with this ability, many existing methods [31; 19; 25; 11; 87; 91; 16; 83; 39; 10; 37] exploit class-specific priors by training 3D generative networks on 3D shape datasets [4]. However, these methods often fail to generalize to unseen categories, and their reconstruction quality is constrained by the limited size of public 3D datasets.
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
Figure 1: One-2-3-45 reconstructs a full $3 6 0 ^ { \circ }$ mesh of any object in 45 seconds given a single image of it. In each example, we showcase the input image in the left column, alongside the generated textured and textureless meshes from three different views.
|
| 19 |
+
|
| 20 |
+
In this work, we pursue a generic solution to turn an image of any object, regardless of its category, into a high-quality 3D textured mesh. To achieve this, we propose a novel approach that can effectively utilize the strong priors learned by 2D diffusion models for 3D reconstruction. Compared to 3D data, 2D images are more readily available and scalable. Recent 2D generative models (e.g., DALLE [64; 63], Imagen [69], and Stable Diffusion [68]) and visual-language models (e.g., CLIP [61]) have made significant strides by pre-training on Internet-scale image datasets. Since they learn a wide range of visual concepts and possess strong priors about our 3D world, it is natural to marry 3D tasks with them. Consequently, an emerging body of research [26; 23; 52; 60; 36], as exemplified by DreamField [26], DreamFusion [60], and Magic3D [36], employs 2D diffusion models or vision language models to assist 3D generative tasks. The common paradigm of them is to perform per-shape optimization with differentiable rendering and the guidance of the CLIP model or 2D diffusion models. While many other 3D representations have been explored, neural fields are the most commonly used representation during optimization.
|
| 21 |
+
|
| 22 |
+
Although these optimization-based methods have achieved impressive results on both text-to-3D [60; 26; 36] and image-to-3D tasks [48; 72], they face some common dilemmas: (a) time-consuming. Pershape optimization typically involves tens of thousands of iterations of full-image volume rendering and prior model inferences, resulting in typically tens of minutes per shape. (b) memory intensive. Since the full image is required for the 2D prior model, the volume rendering can be memory-intensive when the image resolution goes up. (c) 3D inconsistent. Since the 2D prior model only sees a single view at each iteration and tries to make every view look like the input, they often generate 3D inconsistent shapes (e.g., with two faces, or the Janus problem [48; 60]). (d) poor geometry. Many methods utilize the density field as the representation in volume rendering. It is common that they produce good RGB renderings but extracting high-quality mesh tends to be difficult.
|
| 23 |
+
|
| 24 |
+
In this paper, instead of following the common optimization-based paradigm, we propose a novel approach to utilize 2D prior models for 3D modeling. At the heart of our approach is the combination of a 2D diffusion model with a cost-volume-based 3D reconstruction technique, enabling the reconstruction of a high-quality $3 6 0 ^ { \circ }$ textured mesh from a single image in a feed-forward pass without per-scene optimization. Specifically, we leverage a recent 2D diffusion model, Zero123 [41], which is fine-tuned on Stable Diffusion [68] to predict novel views of the input image given the camera transformation. We utilize it to generate multi-view predictions of the input single image so that we can leverage multi-view 3D reconstruction techniques to obtain a 3D mesh. There are two challenges associated with reconstruction from synthesized multi-view predictions: (a) the inherent lack of perfect consistency within the multi-view predictions, which can lead to severe failures in optimization-based methods such as NeRF methods [53; 5]. (b) the camera pose of the input image is required but unknown. To tackle them, we build our reconstruction module upon a cost volume-based neural surface reconstruction approach, SparseNeuS [45], which is a variant of MVSNeRF [6]. Additionally, we introduce a series of essential training strategies that enable the reconstruction of 360-degree meshes from inherently inconsistent multi-view predictions. We also propose an elevation estimation module that estimates the elevation of the input shape in Zero123’s canonical coordinate system, which is used to compute the camera poses required by the reconstruction module.
|
| 25 |
+
|
| 26 |
+
By integrating the three modules of multi-view synthesis, elevation estimation, and 3D reconstruction, our method can reconstruct 3D meshes of any object from a single image in a feed-forward manner. Without costly optimizations, our method reconstructs 3D shapes in significantly less time, e.g., in just 45 seconds. Our method favors better geometry due to the use of SDF representations, and generates more consistent 3D meshes, thanks to the camera-conditioned multi-view predictions. Moreover, our reconstruction adheres more closely to the input image compared to existing methods. See Figure 1 for some of our example results. We evaluate our method on both synthetic data and real images and demonstrate that our method outperforms existing methods in terms of both quality and efficiency.
|
| 27 |
+
|
| 28 |
+
# 2 Related Work
|
| 29 |
+
|
| 30 |
+
# 2.1 3D Generation Guided by 2D Prior Models
|
| 31 |
+
|
| 32 |
+
Recently, 2D generative models (e.g., DALL-E [64; 63], Imagen [69], and Stable Diffusion [68]) and vision-language models (e.g., CLIP [61]) have learned a wide range of visual concepts by pre-training on Internet-scale image datasets. They possess powerful priors about our 3D world and have inspired a growing body of research to employ 2D prior models for assisting 3D understanding [38; 40] and generative tasks. Exemplified by DreamField [26], DreamFusion [60], and Magic3D [36], a line of works follows the paradigm of per-shape optimization. They typically optimize a 3D representation (i.e., NeRF, mesh, SMPL human model) and utilize differentiable rendering to generate 2D images from various views. The images are then fed to the CLIP model [23; 26; 52; 35; 3; 32; 2; 28; 89; 43] or 2D diffusion model [60; 36; 72; 48; 13; 78; 88; 51; 99; 62; 75] for calculating the loss functions, which are used to guide the 3D shape optimization. In addition to optimization-based 3D shape generation, some works train a 3D generative model but leverage the embedding space of CLIP [8; 44; 71], and some works focus on generating textures or materials for input meshes using 2D models’ prior [52; 82; 7; 51; 67].
|
| 33 |
+
|
| 34 |
+
# 2.2 Single Image to 3D
|
| 35 |
+
|
| 36 |
+
Before the emergence of CLIP and large-scale 2D diffusion models, people often learn 3D priors from 3D synthetic data [4] or real scans [65]. Unlike 2D images, 3D data can be represented in various formats and numerous representation-specific 3D generative models have been proposed. By combining 2D image encoder and 3D generators, they generate 3D data in various representations, including 3D voxels [19; 85; 11; 87; 86; 91], point clouds [16; 94; 20; 1; 49; 96], polygon meshes [31; 79; 83; 56], and parametric models [59; 100; 101]. Recently, there has been an increasing number of work on learning to generate a 3D implicit field from a single image [90; 50; 70; 25; 58; 18; 21; 27; 55; 84; 54].
|
| 37 |
+
|
| 38 |
+
As previously mentioned, several recent works leverage 2D diffusion models to perform per-shape optimization, allowing for the text-to-3D task [60; 36; 26] given that diffusion models are typically conditioned on text. To enable the generation of 3D models from a single image, some works [48; 13; 51] utilize textual inversion [17], to find the best-matching text embedding for the input image, which is then fed into a diffusion model. NeuralLift-360 [24] adds a CLIP loss to enforce similarity between the rendered image and the input image. 3DFuse [72] finetunes the Stable Diffusion model with LoRA layers [24] and a sparse depth injector to ensure greater 3D consistency. A recent work
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
Figure 2: Our method consists of three primary components: (a) Multi-view synthesis: we use a view-conditioned 2D diffusion model, Zero123 [41], to generate multi-view images in a two-stage manner. The input of Zero123 includes a single image and a relative camera transformation, which is parameterized by the relative spherical coordinates $( \Delta \theta , \Delta \phi , \Delta r )$ . (b) Pose estimation: we estimate the elevation angle $\theta$ of the input image based on four nearby views generated by Zero123. We then obtain the poses of the multi-view images by combining the specified relative poses with the estimated pose of the input view. (c) 3D reconstruction: We feed the multi-view posed images to an SDF-based generalizable neural surface reconstruction module for $3 6 0 ^ { \circ }$ mesh reconstruction.
|
| 42 |
+
|
| 43 |
+
Zero123 [41; 73] finetunes the Stable Diffusion model [69] to generate a novel view of the input image based on relative camera pose. In addition to these methods, OpenAI trains a 3D native diffusion model Point-E [57], which uses several million internal 3D models to generate point clouds. Very recently, they published another model Shap-E [30] which is trained to generate parameters of implicit functions that can be used for producing textured meshes or neural radiance fields.
|
| 44 |
+
|
| 45 |
+
# 2.3 Generalizable Neural Reconstruction
|
| 46 |
+
|
| 47 |
+
Traditional NeRF-like methods [53; 80] use a neural network to represent a single scene and require per-scene optimization. However, some approaches aim to learn priors across scenes and generalize to novel scenes. These methods typically take a few source views as input and leverage 2D networks for extracting 2D features. The pixel features are then unprojected into 3D space, and a NeRF-based rendering pipeline is applied on top of them. In this way, they can generate a 3D implicit field given a few source views in a single feed-forward pass. Among the methods, some [81; 65; 22; 95; 93; 42; 34; 76; 77] directly aggregate 2D features with MLPs or transformers, while others explicitly construct the 3D feature/cost volume [6; 29; 98; 45], and utilize the voxel feature for decoding density and color. In addition to the density field representation, some methods such as SparseNeuS [45] and VolRecon [66] utilize SDF representations for geometry reconstruction.
|
| 48 |
+
|
| 49 |
+
# 3 Method
|
| 50 |
+
|
| 51 |
+
Our overall pipeline is illustrated in Figure 2. In Section 3.1, we introduce a view-conditioned 2D diffusion model, Zero123 [41], which is used to generate multi-view images. In Section 3.2, we show that traditional NeRF-based and SDF-based methods fail to reconstruct high-quality meshes from inconsistent multi-view predictions even given ground truth camera poses. Therefore, in Section 3.3, we propose a cost volume-based neural surface reconstruction module that can be trained to handle inconsistent multi-view predictions and reconstruct a 3D mesh in a single feed-forward pass. Specifically, we build upon the SparseNeuS [45] and introduce several critical training strategies to support $3 6 0 ^ { \circ }$ mesh reconstruction. Additionally, in Section 3.4, we demonstrate the necessity of estimating the pose of the input view in Zero123’s canonical space for 3D reconstruction. While the azimuth and radius can be arbitrarily specified, we propose a novel module that utilizes four nearby views generated by Zero123 to estimate the elevation of the input view.
|
| 52 |
+
|
| 53 |
+
# 3.1 Zero123: View-Conditioned 2D Diffusion
|
| 54 |
+
|
| 55 |
+
Recent 2D diffusion models [64; 69; 68] have demonstrated the ability to learn a wide range of visual concepts and strong priors by training on internet-scale data. While the original diffusion models mainly focused on the task of text-to-image, recent work [97; 24] has shown that fine-tuning pretrained models allows us to add various conditional controls to the diffusion models and generate images based on specific conditions. Several conditions, such as canny edges, user scribbles, depth, and normal maps, have already proven effective [97].
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
Figure 3: NeRF-based method [53] and SDF-based method [80] fail to reconstruct high-quality meshes given multi-view images predicted by Zero123. See Figure 1 for our reconstruction results.
|
| 59 |
+
|
| 60 |
+
The recent work Zero123 [41] shares a similar spirit and aims to add viewpoint condition control for the Stable Diffusion model [68]. Specifically, given a single RGB image of an object and a relative camera transformation, Zero123 aims to control the diffusion model to synthesize a new image under this transformed camera view. To achieve this, Zero123 fine-tunes the Stable Diffusion on paired images with their relative camera transformations, synthesized from a large-scale 3D dataset [12]. During the creation of the fine-tuning dataset, Zero123 assumes that the object is centered at the origin of the coordinate system and uses a spherical camera, i.e., the camera is placed on the sphere’s surface and always looks at the origin. For two camera poses $( \theta _ { 1 } , \phi _ { 1 } , r _ { 1 } )$ and $( \theta _ { 2 } , \phi _ { 2 } , r _ { 2 } )$ , where $\theta _ { i }$ , $\phi _ { i }$ , and $r _ { i }$ denote the polar angle, azimuth angle, and radius, their relative camera transformation is parameterized as $( \bar { \theta _ { 2 } } - \theta _ { 1 } , \bar { \phi } _ { 2 } - \phi _ { 1 } , r _ { 2 } - r _ { 1 } )$ . They aim to learn a model $f$ , such that $f ( x _ { 1 } , \theta _ { 2 } - \theta _ { 1 } , \phi _ { 2 } - \phi _ { 1 } , r _ { 2 } - r _ { 1 } )$ is perceptually similar to $x _ { 2 }$ , where $x _ { 1 }$ and $x _ { 2 }$ are two images of an object captured from different views. Zero123 finds that such fine-tuning enables the Stable Diffusion model to learn a generic mechanism for controlling the camera viewpoints, which extrapolates outside of the objects seen in the fine-tuning dataset.
|
| 61 |
+
|
| 62 |
+
# 3.2 Can NeRF Optimization Lift Multi-View Predictions to 3D?
|
| 63 |
+
|
| 64 |
+
Given a single image of an object, we can utilize Zero123 [41] to generate multi-view images, but can we use traditional NeRF-based or SDF-based methods [5; 80] to reconstruct high-quality 3D meshes from these predictions? We conduct a small experiment to test this hypothesis. Given a single image, we first generate 32 multi-view images using Zero123, with camera poses uniformly sampled from the sphere surface. We then feed the predictions to a NeRF-based method (TensoRF [53]) and an SDF-based method (NeuS [80]), which optimize density and SDF fields, respectively. However, as shown in Figure 3, both methods fail to produce satisfactory results, generating numerous distortions and floaters. This is primarily due to the inconsistency of Zero123’s predictions. In Figure 4, we compare Zero123’s predictions with ground-truth renderings. We can see that the overall PSNR is not very high, particularly when the input relative pose is large or the target pose is at unusual locations (e.g., from the bottom or the top). However, the mask IoU (most regions are greater than 0.95) and CLIP similarity are relatively good. This suggests that Zero123 tends to generate predictions that are perceptually similar to the ground truth and have similar contours or boundaries, but the pixel-level appearance may not be exactly the same. Nevertheless, such inconsistencies between the source views are already fatal to traditional optimization-based methods. Although the original Zero123 paper proposes another method for lifting its multi-view predictions, we will demonstrate in experiments that it also fails to yield perfect results and entails time-consuming optimization.
|
| 65 |
+
|
| 66 |
+
# 3.3 Neural Surface Reconstruction from Imperfect Multi-View Predictions
|
| 67 |
+
|
| 68 |
+
Instead of using optimization-based approaches, we base our reconstruction module on a generalizable SDF reconstruction method SparseNeuS [45], which is essentially a variant of the MVSNeRF [6] pipeline that combines multi-view stereo, neural scene representation, and volume rendering. As illustrated in Figure 2, our reconstruction module takes multiple source images with corresponding camera poses as input and generates a textured mesh in a single feed-forward pass. In this section, we will first briefly describe the network pipeline of the module and then explain how we train the module, select the source images, and generate textured meshes. Additionally, in Section 3.4, we will discuss how we generate the camera poses for the source images.
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
Figure 4: We analyze the prediction quality of Zero123 by comparing its predictions to ground truth renderings across various view transformations. For each view transformation, we report the average PSNR, mask IoU, and CLIP similarity of 100 shapes from the Objaverse [12] dataset. The prediction mask is calculated by considering foreground objects (i.e., non-white regions). Zero123 provides more accurate predictions when the view transformation is small.
|
| 72 |
+
|
| 73 |
+
As shown in Figure 2, our reconstruction module takes $m$ posed source images as input. The module begins by extracting $m$ 2D feature maps using a 2D feature network. Next, the module builds a 3D cost volume whose contents are computed by first projecting each 3D voxel to $m$ 2D feature planes and then fetching the variance of the features across the $m$ projected 2D locations. The cost volume is then processed using a sparse 3D CNN to obtain a geometry volume that encodes the underlying geometry of the input shape. To predict the SDF at an arbitrary 3D point, an MLP network takes the 3D coordinate and its corresponding interpolated features from the geometry encoding volume as input. To predict the color of a 3D point, another MLP network takes as input the 2D features at the projected locations, interpolated features from the geometry volume, and the viewing direction of the query ray relative to the viewing direction of the source images. The network predicts the blending weights for each source view, and the color of the 3D point is predicted as the weighted sum of its projected colors. Finally, an SDF-based rendering technique is applied on top of the two MLP networks for RGB and mask rendering [80]. In each iteration, we randomly choose one view to build the cost volume and another view for rendering supervision.
|
| 74 |
+
|
| 75 |
+
2-Stage Source View Selection and Groundtruth-Prediction Mixed Training. Although the original SparseNeuS [45] paper only demonstrated frontal view reconstruction, we have extended it to reconstruct 360-degree meshes in a single feed-forward pass by selecting source views in a particular way. Specifically, our reconstruction model is trained on a 3D object dataset while freezing Zero123. We follow Zero123 to normalize the training shapes and use a spherical camera model. For each shape, we first render $n$ ground-truth RGB images from $n$ camera poses uniformly placed on the sphere. For each of the $n$ views, we use Zero123 to predict four nearby views. During training, we feed all $4 \times n$ predictions with ground-truth poses into the reconstruction module and randomly choose one of the $n$ ground-truth RGB images views as the target view. We call this view selection strategy as 2-stage source view selection. We supervise the training with both the ground-truth RGB and mask values. In this way, the module can learn to handle the inconsistent predictions from Zero123 and reconstruct a consistent $3 6 0 ^ { \circ }$ mesh. We argue that our two-stage source view selection strategy is critical since uniformly choosing $n \times 4$ source views from the sphere surface would result in larger distances between the camera poses. However, cost volume-based methods [45; 29; 6] typically rely on very close source views to find local correspondences. Furthermore, as shown in Figure 4, when the relative pose is small (e.g., 10 degrees apart), Zero123 can provide very accurate and consistent predictions and thus can be used to find local correspondences and infer the geometry.
|
| 76 |
+
|
| 77 |
+
During training, we utilize $n$ ground-truth renderings in the initial stage. We find that employing $n$ predicted images at this stage would suffer from notable inconsistencies across different views, complicating the network’s ability to learn sharp details (see examples in ablation study). However, during inference, we can replace the $n$ ground-truth renderings with Zero123 predictions, as shown in Figure 2, the network can automatically generalize to some extent. We will show in the experiments that this groundtruth-prediction mixed training strategy is also important. To export the textured mesh, we use marching cubes [46] to extract the mesh from the predicted SDF field and query the color of the mesh vertices as described in [80]. Although our reconstruction module is trained on a 3D dataset, we find that it mainly relies on local correspondences and can generalize to unseen shapes very well.
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
Figure 5: Qualitative examples of One-2-3-45 for both synthetic and real images. Each triplet showcases an input image, a textured mesh, and a textureless mesh.
|
| 81 |
+
|
| 82 |
+
# 3.4 Camera Pose Estimation
|
| 83 |
+
|
| 84 |
+
Our reconstruction module requires camera poses for the $4 \times n$ source view images. Note that we adopt Zero123 for image synthesis, which parameterizes cameras in a canonical spherical coordinate frame, $( \theta , \phi , r )$ , where $\theta$ , $\phi$ and $r$ represent the elevation, azimuth, and radius. While we can arbitrarily adjust the azimuth angle $\phi$ and the radius $r$ of all source view images simultaneously, resulting in the rotation and scaling of the reconstructed object accordingly, this parameterization requires knowing the absolute elevation angle $\theta$ of one camera to determine the relative poses of all cameras in a standard XYZ frame. More specifically, the relative poses between camera $( \theta _ { 0 } , \phi _ { 0 } , r _ { 0 } )$ and camera $( \theta _ { 0 } + \Delta \theta , \phi _ { 0 } + \Delta \phi , r _ { 0 } )$ vary for different $\theta _ { 0 }$ even when $\Delta \theta$ and $\Delta \phi$ are the same. Because of this, changing the elevation angles of all source images together (e.g., by 30 degrees up or 30 degrees down) will lead to the distortion of the reconstructed shape (see Figure 10 for examples).
|
| 85 |
+
|
| 86 |
+
Therefore, we propose an elevation estimation module to infer the elevation angle of the input image. First, we use Zero123 to predict four nearby views of the input image. Then we enumerate all possible elevation angles in a coarse-to-fine manner. For each elevation candidate angle, we compute the corresponding camera poses for the four images and calculate a reprojection error for this set of camera poses to measure the consistency between the images and the camera poses. The elevation angle with the smallest reprojection error is used to generate the camera poses for all $4 \times n$ source views by combining the pose of the input view and the relative poses. Please refer to the appendix for details on how we calculate the reprojection error for a set of posed images.
|
| 87 |
+
|
| 88 |
+
# 4 Experiments
|
| 89 |
+
|
| 90 |
+
# 4.1 Implementation Details
|
| 91 |
+
|
| 92 |
+
For each input image, we generate $n = 8$ images by choosing camera poses uniformly placed on the sphere surface and then generate 4 local images $1 0 ^ { \circ }$ apart) for each of the 8 views, resulting in 32 source-view images for reconstruction. During training, we freeze the Zero123 [41] model and train our reconstruction module on the Objaverse-LVIS [12] dataset, which contains 46K 3D models in 1,156 categories. We use BlenderProc [14] to render ground-truth RGB images. For images with background, we utilize an off-the-shelf segmentation network SAM [33] with bounding-box prompts for background removal. Please refer to the appendix for more details.
|
| 93 |
+
|
| 94 |
+
# 4.2 Single Image to 3D Mesh
|
| 95 |
+
|
| 96 |
+
We present qualitative examples of our method in Figures 1 and 5, illustrating its effectiveness in handling both synthetic images and real images. We also compare One-2-3-45 with existing zero-shot single image 3D reconstruction approaches, including Point-E [57], Shap-E [30], Zero123 (Stable
|
| 97 |
+
|
| 98 |
+
|
| 99 |
+
Figure 6: We compare One-2-3-45 with Point-E [57], Shap-E [30], Zero123 (Stable Dreamfusion version) [41], 3DFuse [72], and RealFusion [48]. In each example, we present both the textured and textureless meshes. As 3DFuse [72] and RealFusion [48] do not natively support the export of textured meshes, we showcase the results of volume rendering instead.
|
| 100 |
+
|
| 101 |
+
Table 1: Quantitative Comparison on GSO [15] and Objaverse [12] datasets.
|
| 102 |
+
|
| 103 |
+
<table><tr><td></td><td>Prior Source</td><td>F-Score GSO Obj.</td><td>avg.</td><td>CLIP Similarity GSO Obj. avg.</td><td></td><td>Time</td></tr><tr><td>Point-E [57] Shap-E [30]</td><td>internal 3D data</td><td>81.0 81.0 83.4</td><td>81.0</td><td>74.3</td><td>78.5 76.4</td><td>78s 27s</td></tr><tr><td>Zero123+SD [41]</td><td></td><td></td><td>81.2 82.3</td><td>79.6</td><td>82.1 80.9</td><td></td></tr><tr><td></td><td>2D</td><td>75.1</td><td>69.9 72.5</td><td>71.0</td><td>72.7 71.9</td><td>~15min</td></tr><tr><td>RealFusion [48]</td><td>diffusion</td><td>66.7</td><td>59.3 63.0</td><td>69.3</td><td>69.5 69.4</td><td>~90min</td></tr><tr><td>3DFuse [72]</td><td></td><td>60.7 60.2</td><td>60.4</td><td>71.4</td><td>74.072.7</td><td>~30min</td></tr><tr><td></td><td>models</td><td>84.0</td><td></td><td></td><td>79.778.1</td><td>45s</td></tr><tr><td>Ours</td><td></td><td>83.1</td><td>83.5</td><td>76.4</td><td></td><td></td></tr></table>
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
Figure 7: Error distribution of predicted elevations. The median and average are 5.4 and 9.7 degrees.
|
| 107 |
+
|
| 108 |
+
Dreamfusion version) [41], 3DFuse [72], and RealFusion [48]. Among them, Point-E and Shap-E are two 3D native diffusion models released by OpenAI, which are trained on several million internal 3D data, while others are optimization-based approaches leveraging priors from Stable Diffusion [68].
|
| 109 |
+
|
| 110 |
+
Figure 6 presents the qualitative comparison. While most methods can generate plausible 3D meshes from a single image, notable differences exist among them in terms of geometry quality, adherence to the input, and overall 3D consistency. In terms of geometry quality, approaches like RealFusion [48] and 3DFuse [72], which optimize a neural radiance field, face challenges in extracting high-quality meshes. Likewise, Point-E [57] produces a sparse point cloud as its output, resulting in numerous holes on the reconstructed meshes. In contrast, our approach utilizes an SDF presentation and favors better geometry. Regarding adherence to the input, we observe that most baseline methods struggle to preserve the similarity to the input image. Although Shap-E performs slightly better, it still produces lots of failure cases (see the backpack without shoulder straps, distorted shoe, and stool with three legs). In contrast, our approach leverages a powerful 2D diffusion model to directly produce high-quality multi-view images, rather than relying on 3D space hallucination. This strategy provides better adherence to the input views, alleviates the burden of the 3D reconstruction module, and yields results that are more finely attuned to the input. Furthermore, many approaches encounter challenges in achieving consistent 3D results (also known as the Janus problem [48; 60]), as highlighted in the right figure (two-handle mug, multi-face Mario, and two-face backpack). One of the contributing factors to this issue is that several methods optimize each view independently, striving to make each view resemble the input. In contrast, our method capitalizes on the view-conditioned 2D diffusion model, inherently enhancing 3D consistency.
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
Figure 8: Ablations on training strategies of the reconstruction module and the number of views.
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
We also quantitatively compare the approaches on Objaverse [12] and GoogleScannedObjects (GSO) [15] datasets. For each dataset, we randomly choose 20 shapes and render a single image per shape for evaluation. To align the predictions with the ground-truth mesh, we linearly search the scaling factor and the rotation angle, apply Iterative Closest Point (ICP) for sampled point clouds, and select the one with the most number of inliers. We follow RealFusion [48] to report F-score (with a threshold of 0.05) and CLIP similarity, and the runtime on an A100 GPU. As shown in Table 1, our method outperforms all baseline approaches in terms of F-Score. As for CLIP similarity, we surpass all methods except a concurrent work Shap-E [30]. We find that CLIP similarity is very sensitive to the color distribution and less discriminative in local geometry variations (i.e., the number of legs of a stool, the number of handles of a mug). Regarding running time, our method demonstrates a notable advantage over optimization-based approaches and performs on par with 3D native diffusion models, such as Point-E [57] and Shap-E [30]. Specifically, our 3D reconstruction module reconstructs a 3D mesh in approximately 5 seconds, with the remaining time primarily spent on Zero123 predictions, which takes roughly 1 second per image on an A100 GPU.
|
| 118 |
+
|
| 119 |
+
# 4.3 Ablation Study
|
| 120 |
+
|
| 121 |
+
Training strategies. We ablate our training strategies in Figure 8. We found that without our 2-stage source view selection strategy, a network trained to consume 31 uniformly posed Zero123 predictions (fourth column) suffers from severe inconsistency among source views, causing the reconstruction module to fail completely. If we feed only 7 source views (sixth column) without the four nearby views, the reconstruction fails to capture local correspondence and cannot reconstruct fine-grained geometry. During training, we first render $n$ ground-truth renderings and then use Zero123 to predict four nearby views for each of them. If we train directly on $8 \times 4$ ground-truth renderings without Zero123 prediction during training (second column), it fails to generalize well to Zero123 predictions during inference, with many missing regions. Instead, if we replace the $n$ ground-truth renderings with $n$ Zero123 predictions during training (first column), the network fail to generate sharp details (see the strips of the backpack).
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
Figure 9: $3 6 0 ^ { \circ }$ reconstruction vs. multiview fusion. Meshes from different views are in different colors.
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
Figure 10: Incorrect elevations lead to distorted reconstruction. Our elevation estimation module can predict an accurate elevation of the input view.
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
Figure 11: Text to 3D. First row: “a bear in cowboy suit.” Second row: “a kungfu cat.” We utilize DALL-E 2 [63] to generate an image conditioned on the text and then lift it to 3D. We compare our method with Stable Dreamfusion [60] and 3DFuse [72]. For baselines, volume renderings are shown.
|
| 131 |
+
|
| 132 |
+
Elevation estimation. Our reconstruction module relies on accurate elevation angles of the input view. In Figure 10, we demonstrate the impact of providing incorrect elevation angles (e.g., altering the elevation angles of source views by $\pm 3 0 ^ { \circ }$ ), which results in distorted reconstruction results. Instead, utilizing our predicted elevation angles can perfectly match results with ground truth elevations. We also quantitatively test our elevation estimation module by rendering 1,700 images from random camera poses. As shown in Figure 7, our elevation estimation module predicts accurate elevations.
|
| 133 |
+
|
| 134 |
+
Number of source views. In Figure 8, we also investigate the impact of varying the number of source views on 3D reconstruction. We observe that our method is not very sensitive to the number of views as long as the reconstruction module is retrained with the corresponding setting.
|
| 135 |
+
|
| 136 |
+
$3 6 0 ^ { \circ }$ reconstruction vs. multi-view fusion. While our method reconstructs a $3 6 0 ^ { \circ }$ mesh in a single pass, most existing generalizable neural reconstruction approaches [45; 29; 6] primarily focus on frontal view reconstruction. An alternative approach is to independently infer the geometry for each view and subsequently fuse them together. However, we have observed that this strategy often struggles with multi-view fusion due to inconsistent Zero123 predictions, as illustrated in Figure 9.
|
| 137 |
+
|
| 138 |
+
# 4.4 Text to 3D Mesh
|
| 139 |
+
|
| 140 |
+
As shown in Figure 11, by integrating with off-the-shelf text-to-image 2D diffusion models [68; 63], our method can be naturally extended to support text-to-image-3D tasks and generate high-quality textured meshes in a short time. See supplementary for more examples.
|
| 141 |
+
|
| 142 |
+
# 5 Conclusion
|
| 143 |
+
|
| 144 |
+
In this paper, we present a novel method for reconstructing a high-quality $3 6 0 ^ { \circ }$ mesh of any object from a single image of it. In comparison to existing zero-shot approaches, our results exhibit superior geometry, enhanced 3D consistency, and a remarkable adherence to the input image. Notably, our approach reconstructs meshes in a single forward pass without the need for time-consuming optimization, resulting in significantly reduced processing time. Furthermore, our method can be effortlessly extended to support the text-to-3D task.
|
| 145 |
+
|
| 146 |
+
# Acknowledgments
|
| 147 |
+
|
| 148 |
+
This work is supported in part by gifts from Qualcomm. We would like to thank Ruoxi Shi, Xinyue Wei, Hansheng Chen, Jiayuan Gu, Fanbo Xiang, Xiaoshuai Zhang, and Yulin Liu for their helpful discussions and manuscript proofreading.
|
| 149 |
+
|
| 150 |
+
We would like to thank the following sketchfab users for the models used for the demo images in this paper: dimaponomar2019 (backpack), danielpeng (bag), pmlzbt233 (wooden barrel), felixyadomi (cactus), avianinda (burger), shedmon (robocat), ie-niels (stool), phucn (armchair), techCIR (mug), sabriny (fox). All models are CC-By licensed.
|
| 151 |
+
|
| 152 |
+
# References
|
| 153 |
+
|
| 154 |
+
[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40–49. PMLR, 2018.
|
| 155 |
+
[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022.
|
| 156 |
+
[3] Zehranaz Canfes, M Furkan Atasoy, Alara Dirik, and Pinar Yanardag. Text and image guided 3d avatar generation and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4421–4431, 2023.
|
| 157 |
+
[4] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015.
|
| 158 |
+
[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, pages 333–350. Springer, 2022.
|
| 159 |
+
[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124– 14133, 2021.
|
| 160 |
+
[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023.
|
| 161 |
+
[8] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. arXiv preprint arXiv:2212.04493, 2022.
|
| 162 |
+
[9] Han-Pang Chiu, Leslie Pack Kaelbling, and Tomas Lozano-P ´ erez. Automatic class-specific 3d ´ reconstruction from a single image. CSAIL, pages 1–9, 2009.
|
| 163 |
+
[10] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. 2023.
|
| 164 |
+
[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628–644. Springer, 2016.
|
| 165 |
+
[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022.
|
| 166 |
+
[13] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv preprint arXiv:2212.03267, 2022.
|
| 167 |
+
[14] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023.
|
| 168 |
+
[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A highquality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553–2560. IEEE, 2022.
|
| 169 |
+
[16] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017.
|
| 170 |
+
[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.
|
| 171 |
+
[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841–31854, 2022.
|
| 172 |
+
[19] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 484–499. Springer, 2016.
|
| 173 |
+
[20] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mach ˆ e approach to learning 3d surface generation. In ´ Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216–224, 2018.
|
| 174 |
+
[21] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oguz. 3dgen: Triplane latent ˘ diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023.
|
| 175 |
+
[22] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700–4709, 2021.
|
| 176 |
+
[23] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022.
|
| 177 |
+
[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.
|
| 178 |
+
[25] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Planes vs. chairs: Category-guided 3d shape learning without any 3d cues. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part I, pages 727–744. Springer, 2022.
|
| 179 |
+
[26] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867–876, 2022.
|
| 180 |
+
[27] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949–12958, 2021.
|
| 181 |
+
[28] Nikolay Jetchev. Clipmatrix: Text-controlled creation of 3d textured meshes. arXiv preprint arXiv:2109.12922, 2021.
|
| 182 |
+
[29] Mohammad Mahdi Johari, Yann Lepoittevin, and Franc¸ois Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365–18375, 2022.
|
| 183 |
+
[30] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023.
|
| 184 |
+
[31] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning categoryspecific mesh reconstruction from image collectionsgirdhar2016learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371–386, 2018.
|
| 185 |
+
[32] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Text to mesh without 3d supervision using limit subdivision. arXiv preprint arXiv:2203.13333, 2022.
|
| 186 |
+
[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023.
|
| 187 |
+
[34] Jona´s Kulh ˇ anek, Erik Derner, Torsten Sattler, and Robert Babu ´ ska. Viewformer: Nerf-free ˇ neural rendering from few images using transformers. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pages 198–216. Springer, 2022.
|
| 188 |
+
[35] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022.
|
| 189 |
+
[36] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022.
|
| 190 |
+
[37] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11596–11603, 2020.
|
| 191 |
+
[38] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. arXiv preprint arXiv:2305.10764, 2023.
|
| 192 |
+
[39] Minghua Liu, Minhyuk Sung, Radomir Mech, and Hao Su. Deepmetahandles: Learning deformation meta-handles of 3d meshes with biharmonic coordinates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12–21, 2021.
|
| 193 |
+
[40] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21736–21746, 2023.
|
| 194 |
+
[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.
|
| 195 |
+
[42] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824–7833, 2022.
|
| 196 |
+
[43] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss: Image as stetting stone for text-guided 3d shape generation. arXiv preprint arXiv:2209.04145, 2022.
|
| 197 |
+
[44] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss++: Image as stepping stone for text-guided 3d shape generation. arXiv preprint arXiv:2303.15181, 2023.
|
| 198 |
+
[45] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, pages 210–227. Springer, 2022.
|
| 199 |
+
[46] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163–169, 1987.
|
| 200 |
+
[47] Oier Mees, Maxim Tatarchenko, Thomas Brox, and Wolfram Burgard. Self-supervised 3d shape and viewpoint estimation from single images for robotics. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6083–6089. IEEE, 2019.
|
| 201 |
+
[48] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: 360 $\{ \backslash \deg \}$ reconstruction of any object from a single image. arXiv preprint arXiv:2302.10663, 2023.
|
| 202 |
+
[49] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. pc2: Projection—conditioned point cloud diffusion for single-image 3d reconstruction. arXiv preprint arXiv:2302.10668, 2023.
|
| 203 |
+
[50] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460–4470, 2019.
|
| 204 |
+
[51] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022.
|
| 205 |
+
[52] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492–13502, 2022.
|
| 206 |
+
[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021.
|
| 207 |
+
[54] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306–315, 2022.
|
| 208 |
+
[55] Norman Muller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bul ¨ o, Matthias Nießner, and \` Peter Kontschieder. Autorf: Learning 3d object radiance fields from single view observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971–3980, 2022.
|
| 209 |
+
[56] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. Polygen: An autoregressive generative model of 3d meshes. In International conference on machine learning, pages 7220– 7229. PMLR, 2020.
|
| 210 |
+
[57] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022.
|
| 211 |
+
[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019.
|
| 212 |
+
[59] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975–10985, 2019.
|
| 213 |
+
[60] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022.
|
| 214 |
+
[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021.
|
| 215 |
+
[62] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023.
|
| 216 |
+
[63] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.
|
| 217 |
+
[64] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821–8831. PMLR, 2021.
|
| 218 |
+
[65] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021.
|
| 219 |
+
[66] Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine Susstrunk. Volrecon: ¨ Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. arXiv preprint arXiv:2212.08067, 2022.
|
| 220 |
+
[67] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023.
|
| 221 |
+
[68] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. ¨ High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022.
|
| 222 |
+
[69] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.
|
| 223 |
+
[70] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304–2314, 2019.
|
| 224 |
+
[71] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603–18613, 2022.
|
| 225 |
+
[72] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, JinHwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023.
|
| 226 |
+
[73] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero $^ { 1 2 3 + + }$ : a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023.
|
| 227 |
+
[74] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021.
|
| 228 |
+
[75] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182–15192, 2021.
|
| 229 |
+
[77] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nerf needs? In The Eleventh International Conference on Learning Representations, 2022.
|
| 230 |
+
[78] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022.
|
| 231 |
+
[79] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52–67, 2018.
|
| 232 |
+
[80] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021.
|
| 233 |
+
[81] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021.
|
| 234 |
+
[82] Jiacheng Wei, Hao Wang, Jiashi Feng, Guosheng Lin, and Kim-Hui Yap. Taps3d: Text-guided 3d textured shape generation from pseudo supervision, 2023.
|
| 235 |
+
[83] Chao Wen, Yinda Zhang, Zhuwen Li, and Yanwei Fu. Pixel2mesh $^ { + + }$ : Multi-view 3d mesh generation via deformation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1042–1051, 2019.
|
| 236 |
+
[84] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. arXiv preprint arXiv:2301.08247, 2023.
|
| 237 |
+
[85] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017.
|
| 238 |
+
[86] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690–2698, 2019.
|
| 239 |
+
[87] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox $^ { + + }$ : Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12):2919–2935, 2020.
|
| 240 |
+
[88] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift360: Lifting an in-the-wild 2d photo to a 3d object with $\mathrm { 3 6 0 ~ \{ \backslash { d e g } \} }$ views. arXiv preprint arXiv:2211.16431, 2022.
|
| 241 |
+
[89] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. arXiv preprint arXiv:2212.14704, 2022.
|
| 242 |
+
[90] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019.
|
| 243 |
+
[91] Farid Yagubbayli, Yida Wang, Alessio Tonioni, and Federico Tombari. Legoformer: Transformers for block-by-block multi-view 3d reconstruction. arXiv preprint arXiv:2106.12102, 2021.
|
| 244 |
+
[92] Daniel Yang, Tarik Tosun, Benjamin Eisner, Volkan Isler, and Daniel Lee. Robotic grasping through combined image-based grasp proposal and 3d reconstruction. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6350–6356. IEEE, 2021.
|
| 245 |
+
[93] Hao Yang, Lanqing Hong, Aoxue Li, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Liwei Wang. Contranerf: Generalizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. arXiv preprint arXiv:2303.11052, 2023.
|
| 246 |
+
[94] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018.
|
| 247 |
+
[95] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578–4587, 2021.
|
| 248 |
+
[96] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. [97] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. [98] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449–5458, 2022. [99] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023.
|
| 249 |
+
[100] Silvia Zuffi, Angjoo Kanazawa, and Michael J Black. Lions and tigers and bears: Capturing non-rigid, 3d, articulated shape from images. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 3955–3963, 2018.
|
| 250 |
+
[101] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365–6373, 2017.
|
| 251 |
+
|
| 252 |
+
# Appendix
|
| 253 |
+
|
| 254 |
+
We first show more qualitative comparison in Section A, which is followed by a demonstration of additional examples on real-world images and the text-to-3D task in Sections B and C respectively. Furthermore, we present the details of our elevation estimation module in Section D, training and evaluation details in Section E. We finally show the failure cases and discuss the limitations in Section F.
|
| 255 |
+
|
| 256 |
+
# A More Qualitative Comparison
|
| 257 |
+
|
| 258 |
+
|
| 259 |
+
Figure 12: We compare One-2-3-45 with Point-E [57], Shap-E [30], Zero123 (Stable Dreamfusion version) [41], 3DFuse [72], and RealFusion [48]. In each example, we present both the textured and textureless meshes. As 3DFuse [72] and RealFusion [48] do not natively support the export of textured meshes, we showcase the results of volume rendering instead.
|
| 260 |
+
|
| 261 |
+
In Figure 12, we demonstrate more qualitative comparison on Objaverse [12] and GoogleScannedObjects (GSO) [15] datasets. Note that all test shapes are not seen during the training of our 3D reconstruction module.
|
| 262 |
+
|
| 263 |
+
# B More Examples on Real-World Images
|
| 264 |
+
|
| 265 |
+
In Figure 13, we showcase more examples on real-world images and compare our method with the concurrent method Shap-E [30]. The input images are from unsplash.com or captured by ourselves. Note that our results exhibit a closer adherence to the input image.
|
| 266 |
+
|
| 267 |
+
# C More Examples on Text-to-3D
|
| 268 |
+
|
| 269 |
+
In Figure 14, we present additional examples for the text-to-3D task. It is evident that existing approaches struggle to capture fine-grained details, such as a tree hollow, or achieve compositionality, as seen in examples like an orange stool with green legs, a pineapple-shaped Havana hat, or a rocking horse chair. In contrast, our method produces superior results that adhere more closely to the input text. We hypothesize that controlling such fine-grained attributes in the 3D space using existing optimization strategies is inherently challenging. However, by leveraging established 2D text-to-image diffusion models, our method becomes more effective in lifting a single 2D image to a corresponding 3D textured mesh.
|
| 270 |
+
|
| 271 |
+
|
| 272 |
+
Figure 13: We compare One-2-3-45 with Shap-E [30] on real-world images. In each example, we present the input image, generated textured and textureless meshes.
|
| 273 |
+
|
| 274 |
+
# D Details of Elevation Estimation
|
| 275 |
+
|
| 276 |
+
To estimate the elevation angle $\theta$ of the input image, we first utilize Zero123 [41] to predict four nearby views (10 degrees apart) of the input view. With these predicted views, we proceed to enumerate all possible elevation angles and compute the re-projection error for each candidate angle.
|
| 277 |
+
|
| 278 |
+
|
| 279 |
+
Figure 14: Text-to-3D: We compare our method against two native text-to-3D approaches Stable DreamFusion [60] and 3DFuse [72]. To enable text-to-3D, our method first uses a pretrained text-toimage model DALL-E 2 [63] to generate an image from input text (prompted with “3d model, long shot”), and then uplifts the image to a 3D textured mesh.
|
| 280 |
+
|
| 281 |
+
The re-projection error assesses the consistency between camera poses and image observations, akin to the bundle adjustment module employed in the Structure-from-Motion (SfM) pipeline.
|
| 282 |
+
|
| 283 |
+
Specifically, we enumerate all candidate elevation angles in a coarse-to-fine manner. In the coarse stage, we enumerate elevation angles with a 10-degree interval. Once we have determined the elevation angle $e ^ { * }$ associated with the smallest re-projection error, we proceed to the fine stage. In this stage, we enumerate elevation angle candidates ranging from $e ^ { * } - 1 0 ^ { \circ }$ to $e ^ { * } + 1 0 ^ { \circ }$ with a 1-degree interval. This coarse-to-fine design facilitates rapid estimation, completing the elevation estimation module in under 1 second for each shape.
|
| 284 |
+
|
| 285 |
+
Given a set of four predicted nearby views, we perform feature matching to identify corresponding keypoints across each pair of images (a total of six pairs) using an off-the-shelf module LoFTR [74]. For each elevation angle candidate, we calculate the camera pose for the input image by employing the spherical coordinate system with a radius of 1.2 and an azimuth angle of 0. Note that the azimuth angle $\phi$ and the radius $r$ can be arbitrarily adjusted, resulting in the rotation and scaling of the reconstructed object accordingly. Subsequently, we obtain the camera poses for the four predicted views by incorporating the specified delta poses.
|
| 286 |
+
|
| 287 |
+
Once we have the four posed images, we compute the re-projection error by enumerating triplet images. For each triplet of images $( a , b , c )$ sharing a set of keypoints $P$ , we consider each point $p \in P$ . Utilizing images $a$ and $b$ , we perform triangulation to determine the 3D location of $p$ . We then project the 3D point onto the third image $c$ and calculate the reprojection error, which is defined as the $l 1$ distance between the reprojected 2D pixel and the estimated keypoint in image $c$ . By enumerating all image triplets and their corresponding shared keypoints, we obtain the mean projection error for each elevation angle candidate.
|
| 288 |
+
|
| 289 |
+
# E Details of Training and Evaluation
|
| 290 |
+
|
| 291 |
+
Training We train the reconstruction module using the following loss function:
|
| 292 |
+
|
| 293 |
+
$$
|
| 294 |
+
\mathcal { L } = \mathcal { L } _ { r g b } + \lambda _ { 1 } \mathcal { L } _ { e i k o n a l } + \lambda _ { 2 } \mathcal { L } _ { s p a r s i t y }
|
| 295 |
+
$$
|
| 296 |
+
|
| 297 |
+
where $\mathcal { L } _ { r g b }$ represents the $l 1$ loss between the rendered and ground truth color, weighted by the sum of accumulated weights; $\mathcal { L } _ { e i k o n a l }$ and $\mathcal { L } _ { s p a r s i t y }$ are the Eikonal and sparsity terms, respectively, following SparseNeuS [45]. We empirically set the weights as $\lambda _ { 0 } = 1$ , $\lambda _ { 1 } = 0 . 1$ , and $\lambda _ { 2 } = 0 . 0 2$ . For $\lambda _ { 2 }$ , we adopt a linear warm-up strategy following SparseNeuS [45]. To train our reconstruction module, we utilize the LVIS subset of the Objaverse [12] dataset, which consists of 46k 3D models across 1,156 categories. The reconstruction module is trained for $3 0 0 \mathrm { k }$ iterations using two A10 GPUs, with the training process lasting approximately 6 days. It is important to note that our reconstruction module does not heavily rely on large-scale training data, as it primarily leverages local correspondence to infer the geometry, which is relatively easier to learn and generalize.
|
| 298 |
+
|
| 299 |
+
Evaluation We evaluate all baseline approaches using their official codebase. Since the approaches take only a single image as input, the predicted mesh may not have the same scale and transformation as the ground-truth mesh. To ensure a fair comparison, we employ the following process to align the predicted mesh with the ground-truth mesh. First, we align the up direction for the results generated by each approach. Next, for each generated mesh, we perform a linear search over scales and rotation angles along the up direction. After applying each pair of scale and $\mathbf { Z }$ -rotation, we utilize the Iterative Closest Point (ICP) algorithm to align the transformed mesh to the ground-truth mesh. Finally, we select the mesh with the largest number of inliers as the final alignment. This alignment process helps us establish a consistent reference frame for evaluating the predicted meshes across different approaches. To calculate CLIP similarity, we render both ground-truth and generated meshes, capturing 24 views around the 3D shape from fixed viewpoints - 12 views at $3 0 ^ { \circ }$ elevation and 12 views at $0 ^ { \circ }$ elevation.
|
| 300 |
+
|
| 301 |
+
# F Failure Cases and Limitations
|
| 302 |
+
|
| 303 |
+
Our method relies on Zero123 for generating multi-view images, which introduces challenges due to its occasional production of inconsistent results. In Figure 15, we present two typical cases that exemplify such inconsistencies. The first case involves an input view that lacks sufficient information, such as the back view of a fox. In this scenario, Zero123 struggles to generate consistent predictions for the invisible regions, such as the face of the fox. As a consequence, our method may encounter difficulties in accurately inferring the geometry for those regions. The second case involves an input view with ambiguous or complex structures, such as the pulp and peel of a banana. In such situations, Zero123’s ability to accurately infer the underlying geometry becomes limited. As a result, our method may be affected by the inconsistent predictions generated by Zero123. It is important to acknowledge that these limitations arise from the occasional scenarios, and they can impact the performance of our method in certain cases. Addressing these challenges and refining the reliability of Zero123’s predictions remain areas for further investigation and improvement.
|
| 304 |
+
|
| 305 |
+
|
| 306 |
+
Figure 15: Failure cases. Our method relies on Zero123 to generate multi-view images, and we encounter challenges when Zero123 generates inconsistent results. (a) The input view lacks sufficient information. (b) The input view contains ambiguous or complicated structures.
|
| 307 |
+
|
| 308 |
+
We have also noticed slight artifacts on the back side of our generated results. As one of the first works in combining view-conditioned 2D diffusion models with generalizable multi-view reconstruction, we believe that there is still ample room for exploring more advanced reconstruction techniques and incorporating additional regularizations. By doing so, we expect to significantly mitigate the minor artifacts and further enhance results in the future.
|
parse/dev/A6X9y8n4sT/A6X9y8n4sT_content_list.json
ADDED
|
@@ -0,0 +1,1297 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
210,
|
| 8 |
+
122,
|
| 9 |
+
789,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Chao Xu2∗ Haian Jin3,4∗ Linghao Chen1,4∗ Mukund Varma T1 Zexiang Xu6 Hao Su1 ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
346,
|
| 19 |
+
224,
|
| 20 |
+
759,
|
| 21 |
+
271
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "1 UC San Diego 2 UCLA 3 Cornell University 4 Zhejiang University 5 Adobe Research ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
199,
|
| 30 |
+
284,
|
| 31 |
+
792,
|
| 32 |
+
300
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Project Website: http://one-2-3-45.com ",
|
| 39 |
+
"bbox": [
|
| 40 |
+
351,
|
| 41 |
+
313,
|
| 42 |
+
647,
|
| 43 |
+
328
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Abstract ",
|
| 50 |
+
"text_level": 1,
|
| 51 |
+
"bbox": [
|
| 52 |
+
462,
|
| 53 |
+
338,
|
| 54 |
+
535,
|
| 55 |
+
354
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "Single image 3D reconstruction is an important but challenging task that requires extensive knowledge of our natural world. Many existing methods solve this problem by optimizing a neural radiance field under the guidance of 2D diffusion models but suffer from lengthy optimization time, 3D inconsistency results, and poor geometry. In this work, we propose a novel method that takes a single image of any object as input and generates a full 360-degree 3D textured mesh in a single feed-forward pass. Given a single image, we first use a view-conditioned 2D diffusion model, Zero123, to generate multi-view images for the input view, and then aim to lift them up to 3D space. Since traditional reconstruction methods struggle with inconsistent multi-view predictions, we build our 3D reconstruction module upon an SDF-based generalizable neural surface reconstruction method and propose several critical training strategies to enable the reconstruction of 360- degree meshes. Without costly optimizations, our method reconstructs 3D shapes in significantly less time than existing methods. Moreover, our method favors better geometry, generates more 3D consistent results, and adheres more closely to the input image. We evaluate our approach on both synthetic data and in-the-wild images and demonstrate its superiority in terms of both mesh quality and runtime. In addition, our approach can seamlessly support the text-to-3D task by integrating with off-the-shelf text-to-image diffusion models. ",
|
| 62 |
+
"bbox": [
|
| 63 |
+
233,
|
| 64 |
+
373,
|
| 65 |
+
766,
|
| 66 |
+
635
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "1 Introduction ",
|
| 73 |
+
"text_level": 1,
|
| 74 |
+
"bbox": [
|
| 75 |
+
176,
|
| 76 |
+
667,
|
| 77 |
+
310,
|
| 78 |
+
684
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Single image 3D reconstruction, the task of reconstructing a 3D model of an object from a single 2D image, is a long-standing problem in the computer vision community and is crucial for a wide range of applications, such as robotic object manipulation and navigation, 3D content creation, as well as AR/VR [47; 9; 92]. The problem is challenging as it requires not only the reconstruction of visible parts but also the hallucination of invisible regions. Consequently, this problem is often ill-posed and corresponds to multiple plausible solutions because of insufficient evidence from a single image. On the other hand, humans can adeptly infer unseen 3D content based on our extensive knowledge of the 3D world. To endow intelligent agents with this ability, many existing methods [31; 19; 25; 11; 87; 91; 16; 83; 39; 10; 37] exploit class-specific priors by training 3D generative networks on 3D shape datasets [4]. However, these methods often fail to generalize to unseen categories, and their reconstruction quality is constrained by the limited size of public 3D datasets. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
702,
|
| 88 |
+
825,
|
| 89 |
+
867
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/e195fed4c2c468a783ee30914ac3691c4d2573c8573559a4b820582d07327ada.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: One-2-3-45 reconstructs a full $3 6 0 ^ { \\circ }$ mesh of any object in 45 seconds given a single image of it. In each example, we showcase the input image in the left column, alongside the generated textured and textureless meshes from three different views. "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
178,
|
| 102 |
+
85,
|
| 103 |
+
825,
|
| 104 |
+
454
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "In this work, we pursue a generic solution to turn an image of any object, regardless of its category, into a high-quality 3D textured mesh. To achieve this, we propose a novel approach that can effectively utilize the strong priors learned by 2D diffusion models for 3D reconstruction. Compared to 3D data, 2D images are more readily available and scalable. Recent 2D generative models (e.g., DALLE [64; 63], Imagen [69], and Stable Diffusion [68]) and visual-language models (e.g., CLIP [61]) have made significant strides by pre-training on Internet-scale image datasets. Since they learn a wide range of visual concepts and possess strong priors about our 3D world, it is natural to marry 3D tasks with them. Consequently, an emerging body of research [26; 23; 52; 60; 36], as exemplified by DreamField [26], DreamFusion [60], and Magic3D [36], employs 2D diffusion models or vision language models to assist 3D generative tasks. The common paradigm of them is to perform per-shape optimization with differentiable rendering and the guidance of the CLIP model or 2D diffusion models. While many other 3D representations have been explored, neural fields are the most commonly used representation during optimization. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
173,
|
| 113 |
+
539,
|
| 114 |
+
825,
|
| 115 |
+
718
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "Although these optimization-based methods have achieved impressive results on both text-to-3D [60; 26; 36] and image-to-3D tasks [48; 72], they face some common dilemmas: (a) time-consuming. Pershape optimization typically involves tens of thousands of iterations of full-image volume rendering and prior model inferences, resulting in typically tens of minutes per shape. (b) memory intensive. Since the full image is required for the 2D prior model, the volume rendering can be memory-intensive when the image resolution goes up. (c) 3D inconsistent. Since the 2D prior model only sees a single view at each iteration and tries to make every view look like the input, they often generate 3D inconsistent shapes (e.g., with two faces, or the Janus problem [48; 60]). (d) poor geometry. Many methods utilize the density field as the representation in volume rendering. It is common that they produce good RGB renderings but extracting high-quality mesh tends to be difficult. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
724,
|
| 125 |
+
825,
|
| 126 |
+
863
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "In this paper, instead of following the common optimization-based paradigm, we propose a novel approach to utilize 2D prior models for 3D modeling. At the heart of our approach is the combination of a 2D diffusion model with a cost-volume-based 3D reconstruction technique, enabling the reconstruction of a high-quality $3 6 0 ^ { \\circ }$ textured mesh from a single image in a feed-forward pass without per-scene optimization. Specifically, we leverage a recent 2D diffusion model, Zero123 [41], which is fine-tuned on Stable Diffusion [68] to predict novel views of the input image given the camera transformation. We utilize it to generate multi-view predictions of the input single image so that we can leverage multi-view 3D reconstruction techniques to obtain a 3D mesh. There are two challenges associated with reconstruction from synthesized multi-view predictions: (a) the inherent lack of perfect consistency within the multi-view predictions, which can lead to severe failures in optimization-based methods such as NeRF methods [53; 5]. (b) the camera pose of the input image is required but unknown. To tackle them, we build our reconstruction module upon a cost volume-based neural surface reconstruction approach, SparseNeuS [45], which is a variant of MVSNeRF [6]. Additionally, we introduce a series of essential training strategies that enable the reconstruction of 360-degree meshes from inherently inconsistent multi-view predictions. We also propose an elevation estimation module that estimates the elevation of the input shape in Zero123’s canonical coordinate system, which is used to compute the camera poses required by the reconstruction module. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
176,
|
| 135 |
+
869,
|
| 136 |
+
825,
|
| 137 |
+
911
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "",
|
| 144 |
+
"bbox": [
|
| 145 |
+
174,
|
| 146 |
+
90,
|
| 147 |
+
825,
|
| 148 |
+
285
|
| 149 |
+
],
|
| 150 |
+
"page_idx": 2
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"type": "text",
|
| 154 |
+
"text": "By integrating the three modules of multi-view synthesis, elevation estimation, and 3D reconstruction, our method can reconstruct 3D meshes of any object from a single image in a feed-forward manner. Without costly optimizations, our method reconstructs 3D shapes in significantly less time, e.g., in just 45 seconds. Our method favors better geometry due to the use of SDF representations, and generates more consistent 3D meshes, thanks to the camera-conditioned multi-view predictions. Moreover, our reconstruction adheres more closely to the input image compared to existing methods. See Figure 1 for some of our example results. We evaluate our method on both synthetic data and real images and demonstrate that our method outperforms existing methods in terms of both quality and efficiency. ",
|
| 155 |
+
"bbox": [
|
| 156 |
+
174,
|
| 157 |
+
291,
|
| 158 |
+
825,
|
| 159 |
+
401
|
| 160 |
+
],
|
| 161 |
+
"page_idx": 2
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"type": "text",
|
| 165 |
+
"text": "2 Related Work ",
|
| 166 |
+
"text_level": 1,
|
| 167 |
+
"bbox": [
|
| 168 |
+
174,
|
| 169 |
+
420,
|
| 170 |
+
321,
|
| 171 |
+
438
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 2
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "2.1 3D Generation Guided by 2D Prior Models ",
|
| 178 |
+
"text_level": 1,
|
| 179 |
+
"bbox": [
|
| 180 |
+
174,
|
| 181 |
+
450,
|
| 182 |
+
513,
|
| 183 |
+
465
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 2
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "Recently, 2D generative models (e.g., DALL-E [64; 63], Imagen [69], and Stable Diffusion [68]) and vision-language models (e.g., CLIP [61]) have learned a wide range of visual concepts by pre-training on Internet-scale image datasets. They possess powerful priors about our 3D world and have inspired a growing body of research to employ 2D prior models for assisting 3D understanding [38; 40] and generative tasks. Exemplified by DreamField [26], DreamFusion [60], and Magic3D [36], a line of works follows the paradigm of per-shape optimization. They typically optimize a 3D representation (i.e., NeRF, mesh, SMPL human model) and utilize differentiable rendering to generate 2D images from various views. The images are then fed to the CLIP model [23; 26; 52; 35; 3; 32; 2; 28; 89; 43] or 2D diffusion model [60; 36; 72; 48; 13; 78; 88; 51; 99; 62; 75] for calculating the loss functions, which are used to guide the 3D shape optimization. In addition to optimization-based 3D shape generation, some works train a 3D generative model but leverage the embedding space of CLIP [8; 44; 71], and some works focus on generating textures or materials for input meshes using 2D models’ prior [52; 82; 7; 51; 67]. ",
|
| 190 |
+
"bbox": [
|
| 191 |
+
173,
|
| 192 |
+
476,
|
| 193 |
+
825,
|
| 194 |
+
656
|
| 195 |
+
],
|
| 196 |
+
"page_idx": 2
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"type": "text",
|
| 200 |
+
"text": "2.2 Single Image to 3D ",
|
| 201 |
+
"text_level": 1,
|
| 202 |
+
"bbox": [
|
| 203 |
+
176,
|
| 204 |
+
671,
|
| 205 |
+
344,
|
| 206 |
+
686
|
| 207 |
+
],
|
| 208 |
+
"page_idx": 2
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"type": "text",
|
| 212 |
+
"text": "Before the emergence of CLIP and large-scale 2D diffusion models, people often learn 3D priors from 3D synthetic data [4] or real scans [65]. Unlike 2D images, 3D data can be represented in various formats and numerous representation-specific 3D generative models have been proposed. By combining 2D image encoder and 3D generators, they generate 3D data in various representations, including 3D voxels [19; 85; 11; 87; 86; 91], point clouds [16; 94; 20; 1; 49; 96], polygon meshes [31; 79; 83; 56], and parametric models [59; 100; 101]. Recently, there has been an increasing number of work on learning to generate a 3D implicit field from a single image [90; 50; 70; 25; 58; 18; 21; 27; 55; 84; 54]. ",
|
| 213 |
+
"bbox": [
|
| 214 |
+
174,
|
| 215 |
+
696,
|
| 216 |
+
825,
|
| 217 |
+
808
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "As previously mentioned, several recent works leverage 2D diffusion models to perform per-shape optimization, allowing for the text-to-3D task [60; 36; 26] given that diffusion models are typically conditioned on text. To enable the generation of 3D models from a single image, some works [48; 13; 51] utilize textual inversion [17], to find the best-matching text embedding for the input image, which is then fed into a diffusion model. NeuralLift-360 [24] adds a CLIP loss to enforce similarity between the rendered image and the input image. 3DFuse [72] finetunes the Stable Diffusion model with LoRA layers [24] and a sparse depth injector to ensure greater 3D consistency. A recent work ",
|
| 224 |
+
"bbox": [
|
| 225 |
+
174,
|
| 226 |
+
814,
|
| 227 |
+
825,
|
| 228 |
+
911
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "image",
|
| 234 |
+
"img_path": "images/9730fafe3a1bcc5aa3f47b4c5057505ff3ad1b38f3e4fa6f0ddfc5fc0b36174f.jpg",
|
| 235 |
+
"image_caption": [
|
| 236 |
+
"Figure 2: Our method consists of three primary components: (a) Multi-view synthesis: we use a view-conditioned 2D diffusion model, Zero123 [41], to generate multi-view images in a two-stage manner. The input of Zero123 includes a single image and a relative camera transformation, which is parameterized by the relative spherical coordinates $( \\Delta \\theta , \\Delta \\phi , \\Delta r )$ . (b) Pose estimation: we estimate the elevation angle $\\theta$ of the input image based on four nearby views generated by Zero123. We then obtain the poses of the multi-view images by combining the specified relative poses with the estimated pose of the input view. (c) 3D reconstruction: We feed the multi-view posed images to an SDF-based generalizable neural surface reconstruction module for $3 6 0 ^ { \\circ }$ mesh reconstruction. "
|
| 237 |
+
],
|
| 238 |
+
"image_footnote": [],
|
| 239 |
+
"bbox": [
|
| 240 |
+
173,
|
| 241 |
+
88,
|
| 242 |
+
825,
|
| 243 |
+
231
|
| 244 |
+
],
|
| 245 |
+
"page_idx": 3
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"type": "text",
|
| 249 |
+
"text": "Zero123 [41; 73] finetunes the Stable Diffusion model [69] to generate a novel view of the input image based on relative camera pose. In addition to these methods, OpenAI trains a 3D native diffusion model Point-E [57], which uses several million internal 3D models to generate point clouds. Very recently, they published another model Shap-E [30] which is trained to generate parameters of implicit functions that can be used for producing textured meshes or neural radiance fields. ",
|
| 250 |
+
"bbox": [
|
| 251 |
+
174,
|
| 252 |
+
381,
|
| 253 |
+
825,
|
| 254 |
+
452
|
| 255 |
+
],
|
| 256 |
+
"page_idx": 3
|
| 257 |
+
},
|
| 258 |
+
{
|
| 259 |
+
"type": "text",
|
| 260 |
+
"text": "2.3 Generalizable Neural Reconstruction ",
|
| 261 |
+
"text_level": 1,
|
| 262 |
+
"bbox": [
|
| 263 |
+
176,
|
| 264 |
+
469,
|
| 265 |
+
472,
|
| 266 |
+
484
|
| 267 |
+
],
|
| 268 |
+
"page_idx": 3
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"type": "text",
|
| 272 |
+
"text": "Traditional NeRF-like methods [53; 80] use a neural network to represent a single scene and require per-scene optimization. However, some approaches aim to learn priors across scenes and generalize to novel scenes. These methods typically take a few source views as input and leverage 2D networks for extracting 2D features. The pixel features are then unprojected into 3D space, and a NeRF-based rendering pipeline is applied on top of them. In this way, they can generate a 3D implicit field given a few source views in a single feed-forward pass. Among the methods, some [81; 65; 22; 95; 93; 42; 34; 76; 77] directly aggregate 2D features with MLPs or transformers, while others explicitly construct the 3D feature/cost volume [6; 29; 98; 45], and utilize the voxel feature for decoding density and color. In addition to the density field representation, some methods such as SparseNeuS [45] and VolRecon [66] utilize SDF representations for geometry reconstruction. ",
|
| 273 |
+
"bbox": [
|
| 274 |
+
173,
|
| 275 |
+
494,
|
| 276 |
+
825,
|
| 277 |
+
633
|
| 278 |
+
],
|
| 279 |
+
"page_idx": 3
|
| 280 |
+
},
|
| 281 |
+
{
|
| 282 |
+
"type": "text",
|
| 283 |
+
"text": "3 Method ",
|
| 284 |
+
"text_level": 1,
|
| 285 |
+
"bbox": [
|
| 286 |
+
174,
|
| 287 |
+
655,
|
| 288 |
+
269,
|
| 289 |
+
671
|
| 290 |
+
],
|
| 291 |
+
"page_idx": 3
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"type": "text",
|
| 295 |
+
"text": "Our overall pipeline is illustrated in Figure 2. In Section 3.1, we introduce a view-conditioned 2D diffusion model, Zero123 [41], which is used to generate multi-view images. In Section 3.2, we show that traditional NeRF-based and SDF-based methods fail to reconstruct high-quality meshes from inconsistent multi-view predictions even given ground truth camera poses. Therefore, in Section 3.3, we propose a cost volume-based neural surface reconstruction module that can be trained to handle inconsistent multi-view predictions and reconstruct a 3D mesh in a single feed-forward pass. Specifically, we build upon the SparseNeuS [45] and introduce several critical training strategies to support $3 6 0 ^ { \\circ }$ mesh reconstruction. Additionally, in Section 3.4, we demonstrate the necessity of estimating the pose of the input view in Zero123’s canonical space for 3D reconstruction. While the azimuth and radius can be arbitrarily specified, we propose a novel module that utilizes four nearby views generated by Zero123 to estimate the elevation of the input view. ",
|
| 296 |
+
"bbox": [
|
| 297 |
+
174,
|
| 298 |
+
686,
|
| 299 |
+
825,
|
| 300 |
+
839
|
| 301 |
+
],
|
| 302 |
+
"page_idx": 3
|
| 303 |
+
},
|
| 304 |
+
{
|
| 305 |
+
"type": "text",
|
| 306 |
+
"text": "3.1 Zero123: View-Conditioned 2D Diffusion ",
|
| 307 |
+
"text_level": 1,
|
| 308 |
+
"bbox": [
|
| 309 |
+
173,
|
| 310 |
+
857,
|
| 311 |
+
500,
|
| 312 |
+
872
|
| 313 |
+
],
|
| 314 |
+
"page_idx": 3
|
| 315 |
+
},
|
| 316 |
+
{
|
| 317 |
+
"type": "text",
|
| 318 |
+
"text": "Recent 2D diffusion models [64; 69; 68] have demonstrated the ability to learn a wide range of visual concepts and strong priors by training on internet-scale data. While the original diffusion models mainly focused on the task of text-to-image, recent work [97; 24] has shown that fine-tuning pretrained models allows us to add various conditional controls to the diffusion models and generate images based on specific conditions. Several conditions, such as canny edges, user scribbles, depth, and normal maps, have already proven effective [97]. ",
|
| 319 |
+
"bbox": [
|
| 320 |
+
174,
|
| 321 |
+
882,
|
| 322 |
+
821,
|
| 323 |
+
911
|
| 324 |
+
],
|
| 325 |
+
"page_idx": 3
|
| 326 |
+
},
|
| 327 |
+
{
|
| 328 |
+
"type": "image",
|
| 329 |
+
"img_path": "images/3859477efc050f954ba77c528f70b013e9252f003f085bb9ea4acd2b6bf5b573.jpg",
|
| 330 |
+
"image_caption": [
|
| 331 |
+
"Figure 3: NeRF-based method [53] and SDF-based method [80] fail to reconstruct high-quality meshes given multi-view images predicted by Zero123. See Figure 1 for our reconstruction results. "
|
| 332 |
+
],
|
| 333 |
+
"image_footnote": [],
|
| 334 |
+
"bbox": [
|
| 335 |
+
173,
|
| 336 |
+
0,
|
| 337 |
+
816,
|
| 338 |
+
155
|
| 339 |
+
],
|
| 340 |
+
"page_idx": 4
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"type": "text",
|
| 344 |
+
"text": "",
|
| 345 |
+
"bbox": [
|
| 346 |
+
174,
|
| 347 |
+
223,
|
| 348 |
+
825,
|
| 349 |
+
279
|
| 350 |
+
],
|
| 351 |
+
"page_idx": 4
|
| 352 |
+
},
|
| 353 |
+
{
|
| 354 |
+
"type": "text",
|
| 355 |
+
"text": "The recent work Zero123 [41] shares a similar spirit and aims to add viewpoint condition control for the Stable Diffusion model [68]. Specifically, given a single RGB image of an object and a relative camera transformation, Zero123 aims to control the diffusion model to synthesize a new image under this transformed camera view. To achieve this, Zero123 fine-tunes the Stable Diffusion on paired images with their relative camera transformations, synthesized from a large-scale 3D dataset [12]. During the creation of the fine-tuning dataset, Zero123 assumes that the object is centered at the origin of the coordinate system and uses a spherical camera, i.e., the camera is placed on the sphere’s surface and always looks at the origin. For two camera poses $( \\theta _ { 1 } , \\phi _ { 1 } , r _ { 1 } )$ and $( \\theta _ { 2 } , \\phi _ { 2 } , r _ { 2 } )$ , where $\\theta _ { i }$ , $\\phi _ { i }$ , and $r _ { i }$ denote the polar angle, azimuth angle, and radius, their relative camera transformation is parameterized as $( \\bar { \\theta _ { 2 } } - \\theta _ { 1 } , \\bar { \\phi } _ { 2 } - \\phi _ { 1 } , r _ { 2 } - r _ { 1 } )$ . They aim to learn a model $f$ , such that $f ( x _ { 1 } , \\theta _ { 2 } - \\theta _ { 1 } , \\phi _ { 2 } - \\phi _ { 1 } , r _ { 2 } - r _ { 1 } )$ is perceptually similar to $x _ { 2 }$ , where $x _ { 1 }$ and $x _ { 2 }$ are two images of an object captured from different views. Zero123 finds that such fine-tuning enables the Stable Diffusion model to learn a generic mechanism for controlling the camera viewpoints, which extrapolates outside of the objects seen in the fine-tuning dataset. ",
|
| 356 |
+
"bbox": [
|
| 357 |
+
173,
|
| 358 |
+
285,
|
| 359 |
+
825,
|
| 360 |
+
478
|
| 361 |
+
],
|
| 362 |
+
"page_idx": 4
|
| 363 |
+
},
|
| 364 |
+
{
|
| 365 |
+
"type": "text",
|
| 366 |
+
"text": "3.2 Can NeRF Optimization Lift Multi-View Predictions to 3D? ",
|
| 367 |
+
"text_level": 1,
|
| 368 |
+
"bbox": [
|
| 369 |
+
174,
|
| 370 |
+
496,
|
| 371 |
+
627,
|
| 372 |
+
511
|
| 373 |
+
],
|
| 374 |
+
"page_idx": 4
|
| 375 |
+
},
|
| 376 |
+
{
|
| 377 |
+
"type": "text",
|
| 378 |
+
"text": "Given a single image of an object, we can utilize Zero123 [41] to generate multi-view images, but can we use traditional NeRF-based or SDF-based methods [5; 80] to reconstruct high-quality 3D meshes from these predictions? We conduct a small experiment to test this hypothesis. Given a single image, we first generate 32 multi-view images using Zero123, with camera poses uniformly sampled from the sphere surface. We then feed the predictions to a NeRF-based method (TensoRF [53]) and an SDF-based method (NeuS [80]), which optimize density and SDF fields, respectively. However, as shown in Figure 3, both methods fail to produce satisfactory results, generating numerous distortions and floaters. This is primarily due to the inconsistency of Zero123’s predictions. In Figure 4, we compare Zero123’s predictions with ground-truth renderings. We can see that the overall PSNR is not very high, particularly when the input relative pose is large or the target pose is at unusual locations (e.g., from the bottom or the top). However, the mask IoU (most regions are greater than 0.95) and CLIP similarity are relatively good. This suggests that Zero123 tends to generate predictions that are perceptually similar to the ground truth and have similar contours or boundaries, but the pixel-level appearance may not be exactly the same. Nevertheless, such inconsistencies between the source views are already fatal to traditional optimization-based methods. Although the original Zero123 paper proposes another method for lifting its multi-view predictions, we will demonstrate in experiments that it also fails to yield perfect results and entails time-consuming optimization. ",
|
| 379 |
+
"bbox": [
|
| 380 |
+
173,
|
| 381 |
+
522,
|
| 382 |
+
825,
|
| 383 |
+
757
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 4
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "3.3 Neural Surface Reconstruction from Imperfect Multi-View Predictions ",
|
| 390 |
+
"text_level": 1,
|
| 391 |
+
"bbox": [
|
| 392 |
+
174,
|
| 393 |
+
773,
|
| 394 |
+
704,
|
| 395 |
+
789
|
| 396 |
+
],
|
| 397 |
+
"page_idx": 4
|
| 398 |
+
},
|
| 399 |
+
{
|
| 400 |
+
"type": "text",
|
| 401 |
+
"text": "Instead of using optimization-based approaches, we base our reconstruction module on a generalizable SDF reconstruction method SparseNeuS [45], which is essentially a variant of the MVSNeRF [6] pipeline that combines multi-view stereo, neural scene representation, and volume rendering. As illustrated in Figure 2, our reconstruction module takes multiple source images with corresponding camera poses as input and generates a textured mesh in a single feed-forward pass. In this section, we will first briefly describe the network pipeline of the module and then explain how we train the module, select the source images, and generate textured meshes. Additionally, in Section 3.4, we will discuss how we generate the camera poses for the source images. ",
|
| 402 |
+
"bbox": [
|
| 403 |
+
174,
|
| 404 |
+
800,
|
| 405 |
+
825,
|
| 406 |
+
911
|
| 407 |
+
],
|
| 408 |
+
"page_idx": 4
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"type": "image",
|
| 412 |
+
"img_path": "images/afb3b5398aab25a99ee78c8ff3910e86209424d3887c1c60cdc0e0887a5fcbf7.jpg",
|
| 413 |
+
"image_caption": [
|
| 414 |
+
"Figure 4: We analyze the prediction quality of Zero123 by comparing its predictions to ground truth renderings across various view transformations. For each view transformation, we report the average PSNR, mask IoU, and CLIP similarity of 100 shapes from the Objaverse [12] dataset. The prediction mask is calculated by considering foreground objects (i.e., non-white regions). Zero123 provides more accurate predictions when the view transformation is small. "
|
| 415 |
+
],
|
| 416 |
+
"image_footnote": [],
|
| 417 |
+
"bbox": [
|
| 418 |
+
174,
|
| 419 |
+
88,
|
| 420 |
+
823,
|
| 421 |
+
191
|
| 422 |
+
],
|
| 423 |
+
"page_idx": 5
|
| 424 |
+
},
|
| 425 |
+
{
|
| 426 |
+
"type": "text",
|
| 427 |
+
"text": "As shown in Figure 2, our reconstruction module takes $m$ posed source images as input. The module begins by extracting $m$ 2D feature maps using a 2D feature network. Next, the module builds a 3D cost volume whose contents are computed by first projecting each 3D voxel to $m$ 2D feature planes and then fetching the variance of the features across the $m$ projected 2D locations. The cost volume is then processed using a sparse 3D CNN to obtain a geometry volume that encodes the underlying geometry of the input shape. To predict the SDF at an arbitrary 3D point, an MLP network takes the 3D coordinate and its corresponding interpolated features from the geometry encoding volume as input. To predict the color of a 3D point, another MLP network takes as input the 2D features at the projected locations, interpolated features from the geometry volume, and the viewing direction of the query ray relative to the viewing direction of the source images. The network predicts the blending weights for each source view, and the color of the 3D point is predicted as the weighted sum of its projected colors. Finally, an SDF-based rendering technique is applied on top of the two MLP networks for RGB and mask rendering [80]. In each iteration, we randomly choose one view to build the cost volume and another view for rendering supervision. ",
|
| 428 |
+
"bbox": [
|
| 429 |
+
173,
|
| 430 |
+
332,
|
| 431 |
+
825,
|
| 432 |
+
525
|
| 433 |
+
],
|
| 434 |
+
"page_idx": 5
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"type": "text",
|
| 438 |
+
"text": "2-Stage Source View Selection and Groundtruth-Prediction Mixed Training. Although the original SparseNeuS [45] paper only demonstrated frontal view reconstruction, we have extended it to reconstruct 360-degree meshes in a single feed-forward pass by selecting source views in a particular way. Specifically, our reconstruction model is trained on a 3D object dataset while freezing Zero123. We follow Zero123 to normalize the training shapes and use a spherical camera model. For each shape, we first render $n$ ground-truth RGB images from $n$ camera poses uniformly placed on the sphere. For each of the $n$ views, we use Zero123 to predict four nearby views. During training, we feed all $4 \\times n$ predictions with ground-truth poses into the reconstruction module and randomly choose one of the $n$ ground-truth RGB images views as the target view. We call this view selection strategy as 2-stage source view selection. We supervise the training with both the ground-truth RGB and mask values. In this way, the module can learn to handle the inconsistent predictions from Zero123 and reconstruct a consistent $3 6 0 ^ { \\circ }$ mesh. We argue that our two-stage source view selection strategy is critical since uniformly choosing $n \\times 4$ source views from the sphere surface would result in larger distances between the camera poses. However, cost volume-based methods [45; 29; 6] typically rely on very close source views to find local correspondences. Furthermore, as shown in Figure 4, when the relative pose is small (e.g., 10 degrees apart), Zero123 can provide very accurate and consistent predictions and thus can be used to find local correspondences and infer the geometry. ",
|
| 439 |
+
"bbox": [
|
| 440 |
+
174,
|
| 441 |
+
532,
|
| 442 |
+
825,
|
| 443 |
+
767
|
| 444 |
+
],
|
| 445 |
+
"page_idx": 5
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"type": "text",
|
| 449 |
+
"text": "During training, we utilize $n$ ground-truth renderings in the initial stage. We find that employing $n$ predicted images at this stage would suffer from notable inconsistencies across different views, complicating the network’s ability to learn sharp details (see examples in ablation study). However, during inference, we can replace the $n$ ground-truth renderings with Zero123 predictions, as shown in Figure 2, the network can automatically generalize to some extent. We will show in the experiments that this groundtruth-prediction mixed training strategy is also important. To export the textured mesh, we use marching cubes [46] to extract the mesh from the predicted SDF field and query the color of the mesh vertices as described in [80]. Although our reconstruction module is trained on a 3D dataset, we find that it mainly relies on local correspondences and can generalize to unseen shapes very well. ",
|
| 450 |
+
"bbox": [
|
| 451 |
+
174,
|
| 452 |
+
772,
|
| 453 |
+
825,
|
| 454 |
+
911
|
| 455 |
+
],
|
| 456 |
+
"page_idx": 5
|
| 457 |
+
},
|
| 458 |
+
{
|
| 459 |
+
"type": "image",
|
| 460 |
+
"img_path": "images/10e583600c6824d1055f6191a328c798b6f94f3d61f563debc0a3189835096b2.jpg",
|
| 461 |
+
"image_caption": [
|
| 462 |
+
"Figure 5: Qualitative examples of One-2-3-45 for both synthetic and real images. Each triplet showcases an input image, a textured mesh, and a textureless mesh. "
|
| 463 |
+
],
|
| 464 |
+
"image_footnote": [],
|
| 465 |
+
"bbox": [
|
| 466 |
+
174,
|
| 467 |
+
84,
|
| 468 |
+
825,
|
| 469 |
+
305
|
| 470 |
+
],
|
| 471 |
+
"page_idx": 6
|
| 472 |
+
},
|
| 473 |
+
{
|
| 474 |
+
"type": "text",
|
| 475 |
+
"text": "3.4 Camera Pose Estimation ",
|
| 476 |
+
"text_level": 1,
|
| 477 |
+
"bbox": [
|
| 478 |
+
174,
|
| 479 |
+
372,
|
| 480 |
+
383,
|
| 481 |
+
386
|
| 482 |
+
],
|
| 483 |
+
"page_idx": 6
|
| 484 |
+
},
|
| 485 |
+
{
|
| 486 |
+
"type": "text",
|
| 487 |
+
"text": "Our reconstruction module requires camera poses for the $4 \\times n$ source view images. Note that we adopt Zero123 for image synthesis, which parameterizes cameras in a canonical spherical coordinate frame, $( \\theta , \\phi , r )$ , where $\\theta$ , $\\phi$ and $r$ represent the elevation, azimuth, and radius. While we can arbitrarily adjust the azimuth angle $\\phi$ and the radius $r$ of all source view images simultaneously, resulting in the rotation and scaling of the reconstructed object accordingly, this parameterization requires knowing the absolute elevation angle $\\theta$ of one camera to determine the relative poses of all cameras in a standard XYZ frame. More specifically, the relative poses between camera $( \\theta _ { 0 } , \\phi _ { 0 } , r _ { 0 } )$ and camera $( \\theta _ { 0 } + \\Delta \\theta , \\phi _ { 0 } + \\Delta \\phi , r _ { 0 } )$ vary for different $\\theta _ { 0 }$ even when $\\Delta \\theta$ and $\\Delta \\phi$ are the same. Because of this, changing the elevation angles of all source images together (e.g., by 30 degrees up or 30 degrees down) will lead to the distortion of the reconstructed shape (see Figure 10 for examples). ",
|
| 488 |
+
"bbox": [
|
| 489 |
+
174,
|
| 490 |
+
397,
|
| 491 |
+
825,
|
| 492 |
+
536
|
| 493 |
+
],
|
| 494 |
+
"page_idx": 6
|
| 495 |
+
},
|
| 496 |
+
{
|
| 497 |
+
"type": "text",
|
| 498 |
+
"text": "Therefore, we propose an elevation estimation module to infer the elevation angle of the input image. First, we use Zero123 to predict four nearby views of the input image. Then we enumerate all possible elevation angles in a coarse-to-fine manner. For each elevation candidate angle, we compute the corresponding camera poses for the four images and calculate a reprojection error for this set of camera poses to measure the consistency between the images and the camera poses. The elevation angle with the smallest reprojection error is used to generate the camera poses for all $4 \\times n$ source views by combining the pose of the input view and the relative poses. Please refer to the appendix for details on how we calculate the reprojection error for a set of posed images. ",
|
| 499 |
+
"bbox": [
|
| 500 |
+
174,
|
| 501 |
+
542,
|
| 502 |
+
825,
|
| 503 |
+
654
|
| 504 |
+
],
|
| 505 |
+
"page_idx": 6
|
| 506 |
+
},
|
| 507 |
+
{
|
| 508 |
+
"type": "text",
|
| 509 |
+
"text": "4 Experiments ",
|
| 510 |
+
"text_level": 1,
|
| 511 |
+
"bbox": [
|
| 512 |
+
174,
|
| 513 |
+
672,
|
| 514 |
+
312,
|
| 515 |
+
690
|
| 516 |
+
],
|
| 517 |
+
"page_idx": 6
|
| 518 |
+
},
|
| 519 |
+
{
|
| 520 |
+
"type": "text",
|
| 521 |
+
"text": "4.1 Implementation Details ",
|
| 522 |
+
"text_level": 1,
|
| 523 |
+
"bbox": [
|
| 524 |
+
176,
|
| 525 |
+
704,
|
| 526 |
+
377,
|
| 527 |
+
719
|
| 528 |
+
],
|
| 529 |
+
"page_idx": 6
|
| 530 |
+
},
|
| 531 |
+
{
|
| 532 |
+
"type": "text",
|
| 533 |
+
"text": "For each input image, we generate $n = 8$ images by choosing camera poses uniformly placed on the sphere surface and then generate 4 local images $1 0 ^ { \\circ }$ apart) for each of the 8 views, resulting in 32 source-view images for reconstruction. During training, we freeze the Zero123 [41] model and train our reconstruction module on the Objaverse-LVIS [12] dataset, which contains 46K 3D models in 1,156 categories. We use BlenderProc [14] to render ground-truth RGB images. For images with background, we utilize an off-the-shelf segmentation network SAM [33] with bounding-box prompts for background removal. Please refer to the appendix for more details. ",
|
| 534 |
+
"bbox": [
|
| 535 |
+
174,
|
| 536 |
+
729,
|
| 537 |
+
825,
|
| 538 |
+
827
|
| 539 |
+
],
|
| 540 |
+
"page_idx": 6
|
| 541 |
+
},
|
| 542 |
+
{
|
| 543 |
+
"type": "text",
|
| 544 |
+
"text": "4.2 Single Image to 3D Mesh ",
|
| 545 |
+
"text_level": 1,
|
| 546 |
+
"bbox": [
|
| 547 |
+
174,
|
| 548 |
+
843,
|
| 549 |
+
387,
|
| 550 |
+
858
|
| 551 |
+
],
|
| 552 |
+
"page_idx": 6
|
| 553 |
+
},
|
| 554 |
+
{
|
| 555 |
+
"type": "text",
|
| 556 |
+
"text": "We present qualitative examples of our method in Figures 1 and 5, illustrating its effectiveness in handling both synthetic images and real images. We also compare One-2-3-45 with existing zero-shot single image 3D reconstruction approaches, including Point-E [57], Shap-E [30], Zero123 (Stable ",
|
| 557 |
+
"bbox": [
|
| 558 |
+
174,
|
| 559 |
+
869,
|
| 560 |
+
825,
|
| 561 |
+
911
|
| 562 |
+
],
|
| 563 |
+
"page_idx": 6
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"type": "image",
|
| 567 |
+
"img_path": "images/2345160747284d8de57498cc586db11845594856f89bd599e6fef3186fcec258.jpg",
|
| 568 |
+
"image_caption": [
|
| 569 |
+
"Figure 6: We compare One-2-3-45 with Point-E [57], Shap-E [30], Zero123 (Stable Dreamfusion version) [41], 3DFuse [72], and RealFusion [48]. In each example, we present both the textured and textureless meshes. As 3DFuse [72] and RealFusion [48] do not natively support the export of textured meshes, we showcase the results of volume rendering instead. "
|
| 570 |
+
],
|
| 571 |
+
"image_footnote": [],
|
| 572 |
+
"bbox": [
|
| 573 |
+
173,
|
| 574 |
+
84,
|
| 575 |
+
823,
|
| 576 |
+
435
|
| 577 |
+
],
|
| 578 |
+
"page_idx": 7
|
| 579 |
+
},
|
| 580 |
+
{
|
| 581 |
+
"type": "table",
|
| 582 |
+
"img_path": "images/20b00e7ecf87d255e752c778da28af649c559dc9d0c95dd9af4c316f403e0293.jpg",
|
| 583 |
+
"table_caption": [
|
| 584 |
+
"Table 1: Quantitative Comparison on GSO [15] and Objaverse [12] datasets. "
|
| 585 |
+
],
|
| 586 |
+
"table_footnote": [],
|
| 587 |
+
"table_body": "<table><tr><td></td><td>Prior Source</td><td>F-Score GSO Obj.</td><td>avg.</td><td>CLIP Similarity GSO Obj. avg.</td><td></td><td>Time</td></tr><tr><td>Point-E [57] Shap-E [30]</td><td>internal 3D data</td><td>81.0 81.0 83.4</td><td>81.0</td><td>74.3</td><td>78.5 76.4</td><td>78s 27s</td></tr><tr><td>Zero123+SD [41]</td><td></td><td></td><td>81.2 82.3</td><td>79.6</td><td>82.1 80.9</td><td></td></tr><tr><td></td><td>2D</td><td>75.1</td><td>69.9 72.5</td><td>71.0</td><td>72.7 71.9</td><td>~15min</td></tr><tr><td>RealFusion [48]</td><td>diffusion</td><td>66.7</td><td>59.3 63.0</td><td>69.3</td><td>69.5 69.4</td><td>~90min</td></tr><tr><td>3DFuse [72]</td><td></td><td>60.7 60.2</td><td>60.4</td><td>71.4</td><td>74.072.7</td><td>~30min</td></tr><tr><td></td><td>models</td><td>84.0</td><td></td><td></td><td>79.778.1</td><td>45s</td></tr><tr><td>Ours</td><td></td><td>83.1</td><td>83.5</td><td>76.4</td><td></td><td></td></tr></table>",
|
| 588 |
+
"bbox": [
|
| 589 |
+
178,
|
| 590 |
+
566,
|
| 591 |
+
534,
|
| 592 |
+
670
|
| 593 |
+
],
|
| 594 |
+
"page_idx": 7
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"type": "image",
|
| 598 |
+
"img_path": "images/f7032038382ca1d96dc1d7b0e999c17660b4e0d2d2cae29c6076a139017551eb.jpg",
|
| 599 |
+
"image_caption": [
|
| 600 |
+
"Figure 7: Error distribution of predicted elevations. The median and average are 5.4 and 9.7 degrees. "
|
| 601 |
+
],
|
| 602 |
+
"image_footnote": [],
|
| 603 |
+
"bbox": [
|
| 604 |
+
604,
|
| 605 |
+
522,
|
| 606 |
+
766,
|
| 607 |
+
621
|
| 608 |
+
],
|
| 609 |
+
"page_idx": 7
|
| 610 |
+
},
|
| 611 |
+
{
|
| 612 |
+
"type": "text",
|
| 613 |
+
"text": "Dreamfusion version) [41], 3DFuse [72], and RealFusion [48]. Among them, Point-E and Shap-E are two 3D native diffusion models released by OpenAI, which are trained on several million internal 3D data, while others are optimization-based approaches leveraging priors from Stable Diffusion [68]. ",
|
| 614 |
+
"bbox": [
|
| 615 |
+
174,
|
| 616 |
+
708,
|
| 617 |
+
825,
|
| 618 |
+
750
|
| 619 |
+
],
|
| 620 |
+
"page_idx": 7
|
| 621 |
+
},
|
| 622 |
+
{
|
| 623 |
+
"type": "text",
|
| 624 |
+
"text": "Figure 6 presents the qualitative comparison. While most methods can generate plausible 3D meshes from a single image, notable differences exist among them in terms of geometry quality, adherence to the input, and overall 3D consistency. In terms of geometry quality, approaches like RealFusion [48] and 3DFuse [72], which optimize a neural radiance field, face challenges in extracting high-quality meshes. Likewise, Point-E [57] produces a sparse point cloud as its output, resulting in numerous holes on the reconstructed meshes. In contrast, our approach utilizes an SDF presentation and favors better geometry. Regarding adherence to the input, we observe that most baseline methods struggle to preserve the similarity to the input image. Although Shap-E performs slightly better, it still produces lots of failure cases (see the backpack without shoulder straps, distorted shoe, and stool with three legs). In contrast, our approach leverages a powerful 2D diffusion model to directly produce high-quality multi-view images, rather than relying on 3D space hallucination. This strategy provides better adherence to the input views, alleviates the burden of the 3D reconstruction module, and yields results that are more finely attuned to the input. Furthermore, many approaches encounter challenges in achieving consistent 3D results (also known as the Janus problem [48; 60]), as highlighted in the right figure (two-handle mug, multi-face Mario, and two-face backpack). One of the contributing factors to this issue is that several methods optimize each view independently, striving to make each view resemble the input. In contrast, our method capitalizes on the view-conditioned 2D diffusion model, inherently enhancing 3D consistency. ",
|
| 625 |
+
"bbox": [
|
| 626 |
+
173,
|
| 627 |
+
756,
|
| 628 |
+
825,
|
| 629 |
+
909
|
| 630 |
+
],
|
| 631 |
+
"page_idx": 7
|
| 632 |
+
},
|
| 633 |
+
{
|
| 634 |
+
"type": "image",
|
| 635 |
+
"img_path": "images/ed0b10d751e50a7e47e3f085deb2e467e192ecf2f43b1f76786a1cd84c6b8265.jpg",
|
| 636 |
+
"image_caption": [
|
| 637 |
+
"Figure 8: Ablations on training strategies of the reconstruction module and the number of views. "
|
| 638 |
+
],
|
| 639 |
+
"image_footnote": [],
|
| 640 |
+
"bbox": [
|
| 641 |
+
179,
|
| 642 |
+
90,
|
| 643 |
+
821,
|
| 644 |
+
354
|
| 645 |
+
],
|
| 646 |
+
"page_idx": 8
|
| 647 |
+
},
|
| 648 |
+
{
|
| 649 |
+
"type": "text",
|
| 650 |
+
"text": "",
|
| 651 |
+
"bbox": [
|
| 652 |
+
174,
|
| 653 |
+
414,
|
| 654 |
+
614,
|
| 655 |
+
551
|
| 656 |
+
],
|
| 657 |
+
"page_idx": 8
|
| 658 |
+
},
|
| 659 |
+
{
|
| 660 |
+
"type": "image",
|
| 661 |
+
"img_path": "images/02a28683343a2951871939a3565781ffec559d945d96544c8edc5d3e4b7d0211.jpg",
|
| 662 |
+
"image_caption": [],
|
| 663 |
+
"image_footnote": [],
|
| 664 |
+
"bbox": [
|
| 665 |
+
625,
|
| 666 |
+
420,
|
| 667 |
+
825,
|
| 668 |
+
521
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 8
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "text",
|
| 674 |
+
"text": "We also quantitatively compare the approaches on Objaverse [12] and GoogleScannedObjects (GSO) [15] datasets. For each dataset, we randomly choose 20 shapes and render a single image per shape for evaluation. To align the predictions with the ground-truth mesh, we linearly search the scaling factor and the rotation angle, apply Iterative Closest Point (ICP) for sampled point clouds, and select the one with the most number of inliers. We follow RealFusion [48] to report F-score (with a threshold of 0.05) and CLIP similarity, and the runtime on an A100 GPU. As shown in Table 1, our method outperforms all baseline approaches in terms of F-Score. As for CLIP similarity, we surpass all methods except a concurrent work Shap-E [30]. We find that CLIP similarity is very sensitive to the color distribution and less discriminative in local geometry variations (i.e., the number of legs of a stool, the number of handles of a mug). Regarding running time, our method demonstrates a notable advantage over optimization-based approaches and performs on par with 3D native diffusion models, such as Point-E [57] and Shap-E [30]. Specifically, our 3D reconstruction module reconstructs a 3D mesh in approximately 5 seconds, with the remaining time primarily spent on Zero123 predictions, which takes roughly 1 second per image on an A100 GPU. ",
|
| 675 |
+
"bbox": [
|
| 676 |
+
174,
|
| 677 |
+
559,
|
| 678 |
+
825,
|
| 679 |
+
751
|
| 680 |
+
],
|
| 681 |
+
"page_idx": 8
|
| 682 |
+
},
|
| 683 |
+
{
|
| 684 |
+
"type": "text",
|
| 685 |
+
"text": "4.3 Ablation Study ",
|
| 686 |
+
"text_level": 1,
|
| 687 |
+
"bbox": [
|
| 688 |
+
174,
|
| 689 |
+
772,
|
| 690 |
+
316,
|
| 691 |
+
787
|
| 692 |
+
],
|
| 693 |
+
"page_idx": 8
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"type": "text",
|
| 697 |
+
"text": "Training strategies. We ablate our training strategies in Figure 8. We found that without our 2-stage source view selection strategy, a network trained to consume 31 uniformly posed Zero123 predictions (fourth column) suffers from severe inconsistency among source views, causing the reconstruction module to fail completely. If we feed only 7 source views (sixth column) without the four nearby views, the reconstruction fails to capture local correspondence and cannot reconstruct fine-grained geometry. During training, we first render $n$ ground-truth renderings and then use Zero123 to predict four nearby views for each of them. If we train directly on $8 \\times 4$ ground-truth renderings without Zero123 prediction during training (second column), it fails to generalize well to Zero123 predictions during inference, with many missing regions. Instead, if we replace the $n$ ground-truth renderings with $n$ Zero123 predictions during training (first column), the network fail to generate sharp details (see the strips of the backpack). ",
|
| 698 |
+
"bbox": [
|
| 699 |
+
174,
|
| 700 |
+
800,
|
| 701 |
+
825,
|
| 702 |
+
911
|
| 703 |
+
],
|
| 704 |
+
"page_idx": 8
|
| 705 |
+
},
|
| 706 |
+
{
|
| 707 |
+
"type": "image",
|
| 708 |
+
"img_path": "images/e9f02912f18dc6a0538f98aa1ad796c21973649bd53e56bc984b211aa33e564a.jpg",
|
| 709 |
+
"image_caption": [
|
| 710 |
+
"Figure 9: $3 6 0 ^ { \\circ }$ reconstruction vs. multiview fusion. Meshes from different views are in different colors. "
|
| 711 |
+
],
|
| 712 |
+
"image_footnote": [],
|
| 713 |
+
"bbox": [
|
| 714 |
+
189,
|
| 715 |
+
89,
|
| 716 |
+
457,
|
| 717 |
+
195
|
| 718 |
+
],
|
| 719 |
+
"page_idx": 9
|
| 720 |
+
},
|
| 721 |
+
{
|
| 722 |
+
"type": "image",
|
| 723 |
+
"img_path": "images/6ca597811c38b79885a0ecc1752aa25edaf27f83d734bcade31e336c59ff0290.jpg",
|
| 724 |
+
"image_caption": [
|
| 725 |
+
"Figure 10: Incorrect elevations lead to distorted reconstruction. Our elevation estimation module can predict an accurate elevation of the input view. "
|
| 726 |
+
],
|
| 727 |
+
"image_footnote": [],
|
| 728 |
+
"bbox": [
|
| 729 |
+
486,
|
| 730 |
+
88,
|
| 731 |
+
807,
|
| 732 |
+
198
|
| 733 |
+
],
|
| 734 |
+
"page_idx": 9
|
| 735 |
+
},
|
| 736 |
+
{
|
| 737 |
+
"type": "image",
|
| 738 |
+
"img_path": "images/917c7233445de66d2fbaaf138f12d7bed5cd37de61c8abcb4fc1db1ed49327a4.jpg",
|
| 739 |
+
"image_caption": [
|
| 740 |
+
"Figure 11: Text to 3D. First row: “a bear in cowboy suit.” Second row: “a kungfu cat.” We utilize DALL-E 2 [63] to generate an image conditioned on the text and then lift it to 3D. We compare our method with Stable Dreamfusion [60] and 3DFuse [72]. For baselines, volume renderings are shown. "
|
| 741 |
+
],
|
| 742 |
+
"image_footnote": [],
|
| 743 |
+
"bbox": [
|
| 744 |
+
179,
|
| 745 |
+
263,
|
| 746 |
+
825,
|
| 747 |
+
372
|
| 748 |
+
],
|
| 749 |
+
"page_idx": 9
|
| 750 |
+
},
|
| 751 |
+
{
|
| 752 |
+
"type": "text",
|
| 753 |
+
"text": "",
|
| 754 |
+
"bbox": [
|
| 755 |
+
173,
|
| 756 |
+
450,
|
| 757 |
+
826,
|
| 758 |
+
492
|
| 759 |
+
],
|
| 760 |
+
"page_idx": 9
|
| 761 |
+
},
|
| 762 |
+
{
|
| 763 |
+
"type": "text",
|
| 764 |
+
"text": "Elevation estimation. Our reconstruction module relies on accurate elevation angles of the input view. In Figure 10, we demonstrate the impact of providing incorrect elevation angles (e.g., altering the elevation angles of source views by $\\pm 3 0 ^ { \\circ }$ ), which results in distorted reconstruction results. Instead, utilizing our predicted elevation angles can perfectly match results with ground truth elevations. We also quantitatively test our elevation estimation module by rendering 1,700 images from random camera poses. As shown in Figure 7, our elevation estimation module predicts accurate elevations. ",
|
| 765 |
+
"bbox": [
|
| 766 |
+
173,
|
| 767 |
+
498,
|
| 768 |
+
825,
|
| 769 |
+
583
|
| 770 |
+
],
|
| 771 |
+
"page_idx": 9
|
| 772 |
+
},
|
| 773 |
+
{
|
| 774 |
+
"type": "text",
|
| 775 |
+
"text": "Number of source views. In Figure 8, we also investigate the impact of varying the number of source views on 3D reconstruction. We observe that our method is not very sensitive to the number of views as long as the reconstruction module is retrained with the corresponding setting. ",
|
| 776 |
+
"bbox": [
|
| 777 |
+
176,
|
| 778 |
+
588,
|
| 779 |
+
821,
|
| 780 |
+
630
|
| 781 |
+
],
|
| 782 |
+
"page_idx": 9
|
| 783 |
+
},
|
| 784 |
+
{
|
| 785 |
+
"type": "text",
|
| 786 |
+
"text": "$3 6 0 ^ { \\circ }$ reconstruction vs. multi-view fusion. While our method reconstructs a $3 6 0 ^ { \\circ }$ mesh in a single pass, most existing generalizable neural reconstruction approaches [45; 29; 6] primarily focus on frontal view reconstruction. An alternative approach is to independently infer the geometry for each view and subsequently fuse them together. However, we have observed that this strategy often struggles with multi-view fusion due to inconsistent Zero123 predictions, as illustrated in Figure 9. ",
|
| 787 |
+
"bbox": [
|
| 788 |
+
174,
|
| 789 |
+
636,
|
| 790 |
+
825,
|
| 791 |
+
707
|
| 792 |
+
],
|
| 793 |
+
"page_idx": 9
|
| 794 |
+
},
|
| 795 |
+
{
|
| 796 |
+
"type": "text",
|
| 797 |
+
"text": "4.4 Text to 3D Mesh ",
|
| 798 |
+
"text_level": 1,
|
| 799 |
+
"bbox": [
|
| 800 |
+
174,
|
| 801 |
+
715,
|
| 802 |
+
326,
|
| 803 |
+
731
|
| 804 |
+
],
|
| 805 |
+
"page_idx": 9
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "text",
|
| 809 |
+
"text": "As shown in Figure 11, by integrating with off-the-shelf text-to-image 2D diffusion models [68; 63], our method can be naturally extended to support text-to-image-3D tasks and generate high-quality textured meshes in a short time. See supplementary for more examples. ",
|
| 810 |
+
"bbox": [
|
| 811 |
+
176,
|
| 812 |
+
734,
|
| 813 |
+
825,
|
| 814 |
+
777
|
| 815 |
+
],
|
| 816 |
+
"page_idx": 9
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "text",
|
| 820 |
+
"text": "5 Conclusion ",
|
| 821 |
+
"text_level": 1,
|
| 822 |
+
"bbox": [
|
| 823 |
+
174,
|
| 824 |
+
790,
|
| 825 |
+
299,
|
| 826 |
+
808
|
| 827 |
+
],
|
| 828 |
+
"page_idx": 9
|
| 829 |
+
},
|
| 830 |
+
{
|
| 831 |
+
"type": "text",
|
| 832 |
+
"text": "In this paper, we present a novel method for reconstructing a high-quality $3 6 0 ^ { \\circ }$ mesh of any object from a single image of it. In comparison to existing zero-shot approaches, our results exhibit superior geometry, enhanced 3D consistency, and a remarkable adherence to the input image. Notably, our approach reconstructs meshes in a single forward pass without the need for time-consuming optimization, resulting in significantly reduced processing time. Furthermore, our method can be effortlessly extended to support the text-to-3D task. ",
|
| 833 |
+
"bbox": [
|
| 834 |
+
174,
|
| 835 |
+
815,
|
| 836 |
+
825,
|
| 837 |
+
898
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 9
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "text",
|
| 843 |
+
"text": "Acknowledgments ",
|
| 844 |
+
"text_level": 1,
|
| 845 |
+
"bbox": [
|
| 846 |
+
176,
|
| 847 |
+
89,
|
| 848 |
+
328,
|
| 849 |
+
106
|
| 850 |
+
],
|
| 851 |
+
"page_idx": 10
|
| 852 |
+
},
|
| 853 |
+
{
|
| 854 |
+
"type": "text",
|
| 855 |
+
"text": "This work is supported in part by gifts from Qualcomm. We would like to thank Ruoxi Shi, Xinyue Wei, Hansheng Chen, Jiayuan Gu, Fanbo Xiang, Xiaoshuai Zhang, and Yulin Liu for their helpful discussions and manuscript proofreading. ",
|
| 856 |
+
"bbox": [
|
| 857 |
+
174,
|
| 858 |
+
121,
|
| 859 |
+
825,
|
| 860 |
+
162
|
| 861 |
+
],
|
| 862 |
+
"page_idx": 10
|
| 863 |
+
},
|
| 864 |
+
{
|
| 865 |
+
"type": "text",
|
| 866 |
+
"text": "We would like to thank the following sketchfab users for the models used for the demo images in this paper: dimaponomar2019 (backpack), danielpeng (bag), pmlzbt233 (wooden barrel), felixyadomi (cactus), avianinda (burger), shedmon (robocat), ie-niels (stool), phucn (armchair), techCIR (mug), sabriny (fox). All models are CC-By licensed. ",
|
| 867 |
+
"bbox": [
|
| 868 |
+
174,
|
| 869 |
+
169,
|
| 870 |
+
825,
|
| 871 |
+
224
|
| 872 |
+
],
|
| 873 |
+
"page_idx": 10
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "text",
|
| 877 |
+
"text": "References ",
|
| 878 |
+
"text_level": 1,
|
| 879 |
+
"bbox": [
|
| 880 |
+
174,
|
| 881 |
+
244,
|
| 882 |
+
266,
|
| 883 |
+
261
|
| 884 |
+
],
|
| 885 |
+
"page_idx": 10
|
| 886 |
+
},
|
| 887 |
+
{
|
| 888 |
+
"type": "text",
|
| 889 |
+
"text": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40–49. PMLR, 2018. \n[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022. \n[3] Zehranaz Canfes, M Furkan Atasoy, Alara Dirik, and Pinar Yanardag. Text and image guided 3d avatar generation and manipulation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4421–4431, 2023. \n[4] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. \n[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, pages 333–350. Springer, 2022. \n[6] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124– 14133, 2021. \n[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. \n[8] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. arXiv preprint arXiv:2212.04493, 2022. \n[9] Han-Pang Chiu, Leslie Pack Kaelbling, and Tomas Lozano-P ´ erez. Automatic class-specific 3d ´ reconstruction from a single image. CSAIL, pages 1–9, 2009. \n[10] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-sdf: Conditional generative modeling of signed distance functions. 2023. \n[11] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628–644. Springer, 2016. \n[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. \n[13] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv preprint arXiv:2212.03267, 2022. \n[14] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. \n[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A highquality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553–2560. IEEE, 2022. \n[16] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017. \n[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. \n[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841–31854, 2022. \n[19] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 484–499. Springer, 2016. \n[20] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mach ˆ e approach to learning 3d surface generation. In ´ Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216–224, 2018. \n[21] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oguz. 3dgen: Triplane latent ˘ diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. \n[22] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700–4709, 2021. \n[23] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. \n[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. \n[25] Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Planes vs. chairs: Category-guided 3d shape learning without any 3d cues. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part I, pages 727–744. Springer, 2022. \n[26] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867–876, 2022. \n[27] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12949–12958, 2021. \n[28] Nikolay Jetchev. Clipmatrix: Text-controlled creation of 3d textured meshes. arXiv preprint arXiv:2109.12922, 2021. \n[29] Mohammad Mahdi Johari, Yann Lepoittevin, and Franc¸ois Fleuret. Geonerf: Generalizing nerf with geometry priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18365–18375, 2022. \n[30] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. \n[31] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning categoryspecific mesh reconstruction from image collectionsgirdhar2016learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371–386, 2018. \n[32] Nasir Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Text to mesh without 3d supervision using limit subdivision. arXiv preprint arXiv:2203.13333, 2022. \n[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. \n[34] Jona´s Kulh ˇ anek, Erik Derner, Torsten Sattler, and Robert Babu ´ ska. Viewformer: Nerf-free ˇ neural rendering from few images using transformers. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XV, pages 198–216. Springer, 2022. \n[35] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. \n[36] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. \n[37] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11596–11603, 2020. \n[38] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. arXiv preprint arXiv:2305.10764, 2023. \n[39] Minghua Liu, Minhyuk Sung, Radomir Mech, and Hao Su. Deepmetahandles: Learning deformation meta-handles of 3d meshes with biharmonic coordinates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12–21, 2021. \n[40] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21736–21746, 2023. \n[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. \n[42] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824–7833, 2022. \n[43] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss: Image as stetting stone for text-guided 3d shape generation. arXiv preprint arXiv:2209.04145, 2022. \n[44] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss++: Image as stepping stone for text-guided 3d shape generation. arXiv preprint arXiv:2303.15181, 2023. \n[45] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXXII, pages 210–227. Springer, 2022. \n[46] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163–169, 1987. \n[47] Oier Mees, Maxim Tatarchenko, Thomas Brox, and Wolfram Burgard. Self-supervised 3d shape and viewpoint estimation from single images for robotics. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6083–6089. IEEE, 2019. \n[48] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: 360 $\\{ \\backslash \\deg \\}$ reconstruction of any object from a single image. arXiv preprint arXiv:2302.10663, 2023. \n[49] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. pc2: Projection—conditioned point cloud diffusion for single-image 3d reconstruction. arXiv preprint arXiv:2302.10668, 2023. \n[50] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460–4470, 2019. \n[51] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. \n[52] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492–13502, 2022. \n[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021. \n[54] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 306–315, 2022. \n[55] Norman Muller, Andrea Simonelli, Lorenzo Porzi, Samuel Rota Bul ¨ o, Matthias Nießner, and \\` Peter Kontschieder. Autorf: Learning 3d object radiance fields from single view observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971–3980, 2022. \n[56] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. Polygen: An autoregressive generative model of 3d meshes. In International conference on machine learning, pages 7220– 7229. PMLR, 2020. \n[57] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. \n[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165–174, 2019. \n[59] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975–10985, 2019. \n[60] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. \n[61] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021. \n[62] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. \n[63] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. \n[64] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821–8831. PMLR, 2021. \n[65] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. \n[66] Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine Susstrunk. Volrecon: ¨ Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. arXiv preprint arXiv:2212.08067, 2022. \n[67] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. \n[68] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. ¨ High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. \n[69] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. \n[70] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304–2314, 2019. \n[71] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards zero-shot text-to-shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18603–18613, 2022. \n[72] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Jaehoon Ko, Hyeonsu Kim, Junho Kim, JinHwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. \n[73] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero $^ { 1 2 3 + + }$ : a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. \n[74] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021. \n[75] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. arXiv preprint arXiv:2303.14184, 2023. rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182–15192, 2021. \n[77] Mukund Varma, Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, and Zhangyang Wang. Is attention all that nerf needs? In The Eleventh International Conference on Learning Representations, 2022. \n[78] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. \n[79] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52–67, 2018. \n[80] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. \n[81] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690–4699, 2021. \n[82] Jiacheng Wei, Hao Wang, Jiashi Feng, Guosheng Lin, and Kim-Hui Yap. Taps3d: Text-guided 3d textured shape generation from pseudo supervision, 2023. \n[83] Chao Wen, Yinda Zhang, Zhuwen Li, and Yanwei Fu. Pixel2mesh $^ { + + }$ : Multi-view 3d mesh generation via deformation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1042–1051, 2019. \n[84] Chao-Yuan Wu, Justin Johnson, Jitendra Malik, Christoph Feichtenhofer, and Georgia Gkioxari. Multiview compressive coding for 3d reconstruction. arXiv preprint arXiv:2301.08247, 2023. \n[85] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. \n[86] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690–2698, 2019. \n[87] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox $^ { + + }$ : Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12):2919–2935, 2020. \n[88] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift360: Lifting an in-the-wild 2d photo to a 3d object with $\\mathrm { 3 6 0 ~ \\{ \\backslash { d e g } \\} }$ views. arXiv preprint arXiv:2211.16431, 2022. \n[89] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. arXiv preprint arXiv:2212.14704, 2022. \n[90] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. \n[91] Farid Yagubbayli, Yida Wang, Alessio Tonioni, and Federico Tombari. Legoformer: Transformers for block-by-block multi-view 3d reconstruction. arXiv preprint arXiv:2106.12102, 2021. \n[92] Daniel Yang, Tarik Tosun, Benjamin Eisner, Volkan Isler, and Daniel Lee. Robotic grasping through combined image-based grasp proposal and 3d reconstruction. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6350–6356. IEEE, 2021. \n[93] Hao Yang, Lanqing Hong, Aoxue Li, Tianyang Hu, Zhenguo Li, Gim Hee Lee, and Liwei Wang. Contranerf: Generalizable neural radiance fields for synthetic-to-real novel view synthesis via contrastive learning. arXiv preprint arXiv:2303.11052, 2023. \n[94] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. \n[95] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578–4587, 2021. \n[96] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. [97] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. [98] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5449–5458, 2022. [99] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In CVPR, 2023. \n[100] Silvia Zuffi, Angjoo Kanazawa, and Michael J Black. Lions and tigers and bears: Capturing non-rigid, 3d, articulated shape from images. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 3955–3963, 2018. \n[101] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365–6373, 2017. ",
|
| 890 |
+
"bbox": [
|
| 891 |
+
184,
|
| 892 |
+
267,
|
| 893 |
+
826,
|
| 894 |
+
912
|
| 895 |
+
],
|
| 896 |
+
"page_idx": 10
|
| 897 |
+
},
|
| 898 |
+
{
|
| 899 |
+
"type": "text",
|
| 900 |
+
"text": "",
|
| 901 |
+
"bbox": [
|
| 902 |
+
178,
|
| 903 |
+
93,
|
| 904 |
+
826,
|
| 905 |
+
892
|
| 906 |
+
],
|
| 907 |
+
"page_idx": 11
|
| 908 |
+
},
|
| 909 |
+
{
|
| 910 |
+
"type": "text",
|
| 911 |
+
"text": "",
|
| 912 |
+
"bbox": [
|
| 913 |
+
179,
|
| 914 |
+
89,
|
| 915 |
+
826,
|
| 916 |
+
916
|
| 917 |
+
],
|
| 918 |
+
"page_idx": 12
|
| 919 |
+
},
|
| 920 |
+
{
|
| 921 |
+
"type": "text",
|
| 922 |
+
"text": "",
|
| 923 |
+
"bbox": [
|
| 924 |
+
179,
|
| 925 |
+
85,
|
| 926 |
+
826,
|
| 927 |
+
904
|
| 928 |
+
],
|
| 929 |
+
"page_idx": 13
|
| 930 |
+
},
|
| 931 |
+
{
|
| 932 |
+
"type": "text",
|
| 933 |
+
"text": "",
|
| 934 |
+
"bbox": [
|
| 935 |
+
179,
|
| 936 |
+
113,
|
| 937 |
+
826,
|
| 938 |
+
905
|
| 939 |
+
],
|
| 940 |
+
"page_idx": 14
|
| 941 |
+
},
|
| 942 |
+
{
|
| 943 |
+
"type": "text",
|
| 944 |
+
"text": "",
|
| 945 |
+
"bbox": [
|
| 946 |
+
174,
|
| 947 |
+
92,
|
| 948 |
+
826,
|
| 949 |
+
279
|
| 950 |
+
],
|
| 951 |
+
"page_idx": 15
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"type": "text",
|
| 955 |
+
"text": "Appendix ",
|
| 956 |
+
"text_level": 1,
|
| 957 |
+
"bbox": [
|
| 958 |
+
174,
|
| 959 |
+
89,
|
| 960 |
+
258,
|
| 961 |
+
107
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 16
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "text",
|
| 967 |
+
"text": "We first show more qualitative comparison in Section A, which is followed by a demonstration of additional examples on real-world images and the text-to-3D task in Sections B and C respectively. Furthermore, we present the details of our elevation estimation module in Section D, training and evaluation details in Section E. We finally show the failure cases and discuss the limitations in Section F. ",
|
| 968 |
+
"bbox": [
|
| 969 |
+
173,
|
| 970 |
+
121,
|
| 971 |
+
826,
|
| 972 |
+
189
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 16
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "text",
|
| 978 |
+
"text": "A More Qualitative Comparison ",
|
| 979 |
+
"text_level": 1,
|
| 980 |
+
"bbox": [
|
| 981 |
+
176,
|
| 982 |
+
203,
|
| 983 |
+
460,
|
| 984 |
+
220
|
| 985 |
+
],
|
| 986 |
+
"page_idx": 16
|
| 987 |
+
},
|
| 988 |
+
{
|
| 989 |
+
"type": "image",
|
| 990 |
+
"img_path": "images/5f66056c1114dc1d71b7ce2d1909cd5c83a36b54b689f3ae3e0bea4dd975d55a.jpg",
|
| 991 |
+
"image_caption": [
|
| 992 |
+
"Figure 12: We compare One-2-3-45 with Point-E [57], Shap-E [30], Zero123 (Stable Dreamfusion version) [41], 3DFuse [72], and RealFusion [48]. In each example, we present both the textured and textureless meshes. As 3DFuse [72] and RealFusion [48] do not natively support the export of textured meshes, we showcase the results of volume rendering instead. "
|
| 993 |
+
],
|
| 994 |
+
"image_footnote": [],
|
| 995 |
+
"bbox": [
|
| 996 |
+
171,
|
| 997 |
+
232,
|
| 998 |
+
825,
|
| 999 |
+
561
|
| 1000 |
+
],
|
| 1001 |
+
"page_idx": 16
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"type": "text",
|
| 1005 |
+
"text": "In Figure 12, we demonstrate more qualitative comparison on Objaverse [12] and GoogleScannedObjects (GSO) [15] datasets. Note that all test shapes are not seen during the training of our 3D reconstruction module. ",
|
| 1006 |
+
"bbox": [
|
| 1007 |
+
173,
|
| 1008 |
+
642,
|
| 1009 |
+
825,
|
| 1010 |
+
684
|
| 1011 |
+
],
|
| 1012 |
+
"page_idx": 16
|
| 1013 |
+
},
|
| 1014 |
+
{
|
| 1015 |
+
"type": "text",
|
| 1016 |
+
"text": "B More Examples on Real-World Images ",
|
| 1017 |
+
"text_level": 1,
|
| 1018 |
+
"bbox": [
|
| 1019 |
+
174,
|
| 1020 |
+
704,
|
| 1021 |
+
532,
|
| 1022 |
+
722
|
| 1023 |
+
],
|
| 1024 |
+
"page_idx": 16
|
| 1025 |
+
},
|
| 1026 |
+
{
|
| 1027 |
+
"type": "text",
|
| 1028 |
+
"text": "In Figure 13, we showcase more examples on real-world images and compare our method with the concurrent method Shap-E [30]. The input images are from unsplash.com or captured by ourselves. Note that our results exhibit a closer adherence to the input image. ",
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
174,
|
| 1031 |
+
736,
|
| 1032 |
+
826,
|
| 1033 |
+
777
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 16
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "text",
|
| 1039 |
+
"text": "C More Examples on Text-to-3D ",
|
| 1040 |
+
"text_level": 1,
|
| 1041 |
+
"bbox": [
|
| 1042 |
+
176,
|
| 1043 |
+
796,
|
| 1044 |
+
460,
|
| 1045 |
+
814
|
| 1046 |
+
],
|
| 1047 |
+
"page_idx": 16
|
| 1048 |
+
},
|
| 1049 |
+
{
|
| 1050 |
+
"type": "text",
|
| 1051 |
+
"text": "In Figure 14, we present additional examples for the text-to-3D task. It is evident that existing approaches struggle to capture fine-grained details, such as a tree hollow, or achieve compositionality, as seen in examples like an orange stool with green legs, a pineapple-shaped Havana hat, or a rocking horse chair. In contrast, our method produces superior results that adhere more closely to the input text. We hypothesize that controlling such fine-grained attributes in the 3D space using existing optimization strategies is inherently challenging. However, by leveraging established 2D text-to-image diffusion models, our method becomes more effective in lifting a single 2D image to a corresponding 3D textured mesh. ",
|
| 1052 |
+
"bbox": [
|
| 1053 |
+
174,
|
| 1054 |
+
827,
|
| 1055 |
+
825,
|
| 1056 |
+
911
|
| 1057 |
+
],
|
| 1058 |
+
"page_idx": 16
|
| 1059 |
+
},
|
| 1060 |
+
{
|
| 1061 |
+
"type": "image",
|
| 1062 |
+
"img_path": "images/7d32058f86abe5b1c8460fb35f0a1baf13cfcaaa1afacef3ed1fd8dda3aaca86.jpg",
|
| 1063 |
+
"image_caption": [
|
| 1064 |
+
"Figure 13: We compare One-2-3-45 with Shap-E [30] on real-world images. In each example, we present the input image, generated textured and textureless meshes. "
|
| 1065 |
+
],
|
| 1066 |
+
"image_footnote": [],
|
| 1067 |
+
"bbox": [
|
| 1068 |
+
176,
|
| 1069 |
+
77,
|
| 1070 |
+
825,
|
| 1071 |
+
736
|
| 1072 |
+
],
|
| 1073 |
+
"page_idx": 17
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "text",
|
| 1077 |
+
"text": "",
|
| 1078 |
+
"bbox": [
|
| 1079 |
+
176,
|
| 1080 |
+
790,
|
| 1081 |
+
823,
|
| 1082 |
+
818
|
| 1083 |
+
],
|
| 1084 |
+
"page_idx": 17
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"type": "text",
|
| 1088 |
+
"text": "D Details of Elevation Estimation ",
|
| 1089 |
+
"text_level": 1,
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
176,
|
| 1092 |
+
837,
|
| 1093 |
+
467,
|
| 1094 |
+
854
|
| 1095 |
+
],
|
| 1096 |
+
"page_idx": 17
|
| 1097 |
+
},
|
| 1098 |
+
{
|
| 1099 |
+
"type": "text",
|
| 1100 |
+
"text": "To estimate the elevation angle $\\theta$ of the input image, we first utilize Zero123 [41] to predict four nearby views (10 degrees apart) of the input view. With these predicted views, we proceed to enumerate all possible elevation angles and compute the re-projection error for each candidate angle. ",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
174,
|
| 1103 |
+
869,
|
| 1104 |
+
826,
|
| 1105 |
+
911
|
| 1106 |
+
],
|
| 1107 |
+
"page_idx": 17
|
| 1108 |
+
},
|
| 1109 |
+
{
|
| 1110 |
+
"type": "image",
|
| 1111 |
+
"img_path": "images/2042d5a4cef0cee41474ca6d36d6e259fcf445ff2389553c21f5e490be52a969.jpg",
|
| 1112 |
+
"image_caption": [
|
| 1113 |
+
"Figure 14: Text-to-3D: We compare our method against two native text-to-3D approaches Stable DreamFusion [60] and 3DFuse [72]. To enable text-to-3D, our method first uses a pretrained text-toimage model DALL-E 2 [63] to generate an image from input text (prompted with “3d model, long shot”), and then uplifts the image to a 3D textured mesh. "
|
| 1114 |
+
],
|
| 1115 |
+
"image_footnote": [],
|
| 1116 |
+
"bbox": [
|
| 1117 |
+
179,
|
| 1118 |
+
77,
|
| 1119 |
+
818,
|
| 1120 |
+
720
|
| 1121 |
+
],
|
| 1122 |
+
"page_idx": 18
|
| 1123 |
+
},
|
| 1124 |
+
{
|
| 1125 |
+
"type": "text",
|
| 1126 |
+
"text": "The re-projection error assesses the consistency between camera poses and image observations, akin to the bundle adjustment module employed in the Structure-from-Motion (SfM) pipeline. ",
|
| 1127 |
+
"bbox": [
|
| 1128 |
+
174,
|
| 1129 |
+
820,
|
| 1130 |
+
823,
|
| 1131 |
+
849
|
| 1132 |
+
],
|
| 1133 |
+
"page_idx": 18
|
| 1134 |
+
},
|
| 1135 |
+
{
|
| 1136 |
+
"type": "text",
|
| 1137 |
+
"text": "Specifically, we enumerate all candidate elevation angles in a coarse-to-fine manner. In the coarse stage, we enumerate elevation angles with a 10-degree interval. Once we have determined the elevation angle $e ^ { * }$ associated with the smallest re-projection error, we proceed to the fine stage. In this stage, we enumerate elevation angle candidates ranging from $e ^ { * } - 1 0 ^ { \\circ }$ to $e ^ { * } + 1 0 ^ { \\circ }$ with a 1-degree interval. This coarse-to-fine design facilitates rapid estimation, completing the elevation estimation module in under 1 second for each shape. ",
|
| 1138 |
+
"bbox": [
|
| 1139 |
+
174,
|
| 1140 |
+
856,
|
| 1141 |
+
825,
|
| 1142 |
+
911
|
| 1143 |
+
],
|
| 1144 |
+
"page_idx": 18
|
| 1145 |
+
},
|
| 1146 |
+
{
|
| 1147 |
+
"type": "text",
|
| 1148 |
+
"text": "",
|
| 1149 |
+
"bbox": [
|
| 1150 |
+
174,
|
| 1151 |
+
90,
|
| 1152 |
+
823,
|
| 1153 |
+
119
|
| 1154 |
+
],
|
| 1155 |
+
"page_idx": 19
|
| 1156 |
+
},
|
| 1157 |
+
{
|
| 1158 |
+
"type": "text",
|
| 1159 |
+
"text": "Given a set of four predicted nearby views, we perform feature matching to identify corresponding keypoints across each pair of images (a total of six pairs) using an off-the-shelf module LoFTR [74]. For each elevation angle candidate, we calculate the camera pose for the input image by employing the spherical coordinate system with a radius of 1.2 and an azimuth angle of 0. Note that the azimuth angle $\\phi$ and the radius $r$ can be arbitrarily adjusted, resulting in the rotation and scaling of the reconstructed object accordingly. Subsequently, we obtain the camera poses for the four predicted views by incorporating the specified delta poses. ",
|
| 1160 |
+
"bbox": [
|
| 1161 |
+
174,
|
| 1162 |
+
126,
|
| 1163 |
+
825,
|
| 1164 |
+
223
|
| 1165 |
+
],
|
| 1166 |
+
"page_idx": 19
|
| 1167 |
+
},
|
| 1168 |
+
{
|
| 1169 |
+
"type": "text",
|
| 1170 |
+
"text": "Once we have the four posed images, we compute the re-projection error by enumerating triplet images. For each triplet of images $( a , b , c )$ sharing a set of keypoints $P$ , we consider each point $p \\in P$ . Utilizing images $a$ and $b$ , we perform triangulation to determine the 3D location of $p$ . We then project the 3D point onto the third image $c$ and calculate the reprojection error, which is defined as the $l 1$ distance between the reprojected 2D pixel and the estimated keypoint in image $c$ . By enumerating all image triplets and their corresponding shared keypoints, we obtain the mean projection error for each elevation angle candidate. ",
|
| 1171 |
+
"bbox": [
|
| 1172 |
+
173,
|
| 1173 |
+
229,
|
| 1174 |
+
825,
|
| 1175 |
+
327
|
| 1176 |
+
],
|
| 1177 |
+
"page_idx": 19
|
| 1178 |
+
},
|
| 1179 |
+
{
|
| 1180 |
+
"type": "text",
|
| 1181 |
+
"text": "E Details of Training and Evaluation ",
|
| 1182 |
+
"text_level": 1,
|
| 1183 |
+
"bbox": [
|
| 1184 |
+
173,
|
| 1185 |
+
344,
|
| 1186 |
+
496,
|
| 1187 |
+
362
|
| 1188 |
+
],
|
| 1189 |
+
"page_idx": 19
|
| 1190 |
+
},
|
| 1191 |
+
{
|
| 1192 |
+
"type": "text",
|
| 1193 |
+
"text": "Training We train the reconstruction module using the following loss function: ",
|
| 1194 |
+
"bbox": [
|
| 1195 |
+
173,
|
| 1196 |
+
376,
|
| 1197 |
+
705,
|
| 1198 |
+
390
|
| 1199 |
+
],
|
| 1200 |
+
"page_idx": 19
|
| 1201 |
+
},
|
| 1202 |
+
{
|
| 1203 |
+
"type": "equation",
|
| 1204 |
+
"img_path": "images/058aab9ab0eb326754c023b080fb831a455f0b4821e25cbfe696420695975afd.jpg",
|
| 1205 |
+
"text": "$$\n\\mathcal { L } = \\mathcal { L } _ { r g b } + \\lambda _ { 1 } \\mathcal { L } _ { e i k o n a l } + \\lambda _ { 2 } \\mathcal { L } _ { s p a r s i t y }\n$$",
|
| 1206 |
+
"text_format": "latex",
|
| 1207 |
+
"bbox": [
|
| 1208 |
+
369,
|
| 1209 |
+
396,
|
| 1210 |
+
627,
|
| 1211 |
+
412
|
| 1212 |
+
],
|
| 1213 |
+
"page_idx": 19
|
| 1214 |
+
},
|
| 1215 |
+
{
|
| 1216 |
+
"type": "text",
|
| 1217 |
+
"text": "where $\\mathcal { L } _ { r g b }$ represents the $l 1$ loss between the rendered and ground truth color, weighted by the sum of accumulated weights; $\\mathcal { L } _ { e i k o n a l }$ and $\\mathcal { L } _ { s p a r s i t y }$ are the Eikonal and sparsity terms, respectively, following SparseNeuS [45]. We empirically set the weights as $\\lambda _ { 0 } = 1$ , $\\lambda _ { 1 } = 0 . 1$ , and $\\lambda _ { 2 } = 0 . 0 2$ . For $\\lambda _ { 2 }$ , we adopt a linear warm-up strategy following SparseNeuS [45]. To train our reconstruction module, we utilize the LVIS subset of the Objaverse [12] dataset, which consists of 46k 3D models across 1,156 categories. The reconstruction module is trained for $3 0 0 \\mathrm { k }$ iterations using two A10 GPUs, with the training process lasting approximately 6 days. It is important to note that our reconstruction module does not heavily rely on large-scale training data, as it primarily leverages local correspondence to infer the geometry, which is relatively easier to learn and generalize. ",
|
| 1218 |
+
"bbox": [
|
| 1219 |
+
173,
|
| 1220 |
+
416,
|
| 1221 |
+
825,
|
| 1222 |
+
542
|
| 1223 |
+
],
|
| 1224 |
+
"page_idx": 19
|
| 1225 |
+
},
|
| 1226 |
+
{
|
| 1227 |
+
"type": "text",
|
| 1228 |
+
"text": "Evaluation We evaluate all baseline approaches using their official codebase. Since the approaches take only a single image as input, the predicted mesh may not have the same scale and transformation as the ground-truth mesh. To ensure a fair comparison, we employ the following process to align the predicted mesh with the ground-truth mesh. First, we align the up direction for the results generated by each approach. Next, for each generated mesh, we perform a linear search over scales and rotation angles along the up direction. After applying each pair of scale and $\\mathbf { Z }$ -rotation, we utilize the Iterative Closest Point (ICP) algorithm to align the transformed mesh to the ground-truth mesh. Finally, we select the mesh with the largest number of inliers as the final alignment. This alignment process helps us establish a consistent reference frame for evaluating the predicted meshes across different approaches. To calculate CLIP similarity, we render both ground-truth and generated meshes, capturing 24 views around the 3D shape from fixed viewpoints - 12 views at $3 0 ^ { \\circ }$ elevation and 12 views at $0 ^ { \\circ }$ elevation. ",
|
| 1229 |
+
"bbox": [
|
| 1230 |
+
174,
|
| 1231 |
+
556,
|
| 1232 |
+
825,
|
| 1233 |
+
723
|
| 1234 |
+
],
|
| 1235 |
+
"page_idx": 19
|
| 1236 |
+
},
|
| 1237 |
+
{
|
| 1238 |
+
"type": "text",
|
| 1239 |
+
"text": "F Failure Cases and Limitations ",
|
| 1240 |
+
"text_level": 1,
|
| 1241 |
+
"bbox": [
|
| 1242 |
+
176,
|
| 1243 |
+
742,
|
| 1244 |
+
459,
|
| 1245 |
+
758
|
| 1246 |
+
],
|
| 1247 |
+
"page_idx": 19
|
| 1248 |
+
},
|
| 1249 |
+
{
|
| 1250 |
+
"type": "text",
|
| 1251 |
+
"text": "Our method relies on Zero123 for generating multi-view images, which introduces challenges due to its occasional production of inconsistent results. In Figure 15, we present two typical cases that exemplify such inconsistencies. The first case involves an input view that lacks sufficient information, such as the back view of a fox. In this scenario, Zero123 struggles to generate consistent predictions for the invisible regions, such as the face of the fox. As a consequence, our method may encounter difficulties in accurately inferring the geometry for those regions. The second case involves an input view with ambiguous or complex structures, such as the pulp and peel of a banana. In such situations, Zero123’s ability to accurately infer the underlying geometry becomes limited. As a result, our method may be affected by the inconsistent predictions generated by Zero123. It is important to acknowledge that these limitations arise from the occasional scenarios, and they can impact the performance of our method in certain cases. Addressing these challenges and refining the reliability of Zero123’s predictions remain areas for further investigation and improvement. ",
|
| 1252 |
+
"bbox": [
|
| 1253 |
+
174,
|
| 1254 |
+
772,
|
| 1255 |
+
825,
|
| 1256 |
+
911
|
| 1257 |
+
],
|
| 1258 |
+
"page_idx": 19
|
| 1259 |
+
},
|
| 1260 |
+
{
|
| 1261 |
+
"type": "image",
|
| 1262 |
+
"img_path": "images/c7302d6a7bae15a457b8e2746da194f4466d363a8ae3c79deb2b1c0b9fd41acc.jpg",
|
| 1263 |
+
"image_caption": [
|
| 1264 |
+
"Figure 15: Failure cases. Our method relies on Zero123 to generate multi-view images, and we encounter challenges when Zero123 generates inconsistent results. (a) The input view lacks sufficient information. (b) The input view contains ambiguous or complicated structures. "
|
| 1265 |
+
],
|
| 1266 |
+
"image_footnote": [],
|
| 1267 |
+
"bbox": [
|
| 1268 |
+
173,
|
| 1269 |
+
87,
|
| 1270 |
+
823,
|
| 1271 |
+
161
|
| 1272 |
+
],
|
| 1273 |
+
"page_idx": 20
|
| 1274 |
+
},
|
| 1275 |
+
{
|
| 1276 |
+
"type": "text",
|
| 1277 |
+
"text": "",
|
| 1278 |
+
"bbox": [
|
| 1279 |
+
171,
|
| 1280 |
+
241,
|
| 1281 |
+
823,
|
| 1282 |
+
268
|
| 1283 |
+
],
|
| 1284 |
+
"page_idx": 20
|
| 1285 |
+
},
|
| 1286 |
+
{
|
| 1287 |
+
"type": "text",
|
| 1288 |
+
"text": "We have also noticed slight artifacts on the back side of our generated results. As one of the first works in combining view-conditioned 2D diffusion models with generalizable multi-view reconstruction, we believe that there is still ample room for exploring more advanced reconstruction techniques and incorporating additional regularizations. By doing so, we expect to significantly mitigate the minor artifacts and further enhance results in the future. ",
|
| 1289 |
+
"bbox": [
|
| 1290 |
+
174,
|
| 1291 |
+
275,
|
| 1292 |
+
825,
|
| 1293 |
+
344
|
| 1294 |
+
],
|
| 1295 |
+
"page_idx": 20
|
| 1296 |
+
}
|
| 1297 |
+
]
|
parse/dev/A6X9y8n4sT/A6X9y8n4sT_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/A6X9y8n4sT/A6X9y8n4sT_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/B4tkwuzeiY/B4tkwuzeiY_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/B4tkwuzeiY/B4tkwuzeiY_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/B4tkwuzeiY/B4tkwuzeiY_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/DAxQXzdq8SF/DAxQXzdq8SF_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc.md
ADDED
|
@@ -0,0 +1,304 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering
|
| 2 |
+
|
| 3 |
+
Pan $\mathbf { L u ^ { 1 , 3 } }$ , Swaroop Mishra2,3, Tony $\mathbf { X i a } ^ { 1 }$ , Liang $\mathbf { Q i u } ^ { 1 }$ , Kai-Wei Chang1, Song-Chun $\mathbf { Z } \mathbf { h } \mathbf { u } ^ { 1 }$ , Oyvind Tafjord3, Peter Clark3, Ashwin Kalyan3 1University of California, Los Angeles, 2Arizona State University, 3Allen Institute for AI {lupantech, kwchang.cs}@gmail.com, sczhu@stat.ucla.edu, {oyvindt, peterc, ashwinkv}@allenai.org
|
| 4 |
+
|
| 5 |
+
# Abstract
|
| 6 |
+
|
| 7 |
+
When answering a question, humans utilize the information available across different modalities to synthesize a consistent and complete chain of thought (CoT). This process is normally a black box in the case of deep learning models like large-scale language models. Recently, science question benchmarks have been used to diagnose the multi-hop reasoning ability and interpretability of an AI system. However, existing datasets fail to provide annotations for the answers, or are restricted to the textual-only modality, small scales, and limited domain diversity. To this end, we present Science Question Answering (SCIENCEQA), a new benchmark that consists of ${ \sim } 2 1 \mathrm { k }$ multimodal multiple choice questions with diverse science topics and annotations of their answers with corresponding lectures and explanations. We further design language models to learn to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process when answering SCIENCEQA questions. SCIENCEQA demonstrates the utility of CoT in language models, as CoT improves the question answering performance by $1 . 2 0 \%$ in fewshot GPT-3 and $3 . 9 9 \%$ in fine-tuned UnifiedQA. We also explore the upper bound for models to leverage explanations by feeding those in the input; we observe that it improves the few-shot performance of GPT-3 by $1 8 . 9 6 \%$ . Our analysis further shows that language models, similar to humans, benefit from explanations to learn from fewer data and achieve the same performance with just $40 \%$ of the data.1
|
| 8 |
+
|
| 9 |
+
# 1 Introduction
|
| 10 |
+
|
| 11 |
+
A long-standing goal of AI systems is to act reliably and learn complex tasks efficiently like human beings. In the process of reliable decision making, humans follow an explicit chain-of-thought (CoT) reasoning process that is typically expressed as an explanation. However, machine learning models are trained mostly using a large number of input-output examples to perform a specific task. These black-box models only generate the final decision without reliably revealing the underlying reasoning process. Not surprisingly, it is unclear if they understand the task and can generalize even though they perform well on the benchmark. On the other hand, humans are able to learn from instructions or explanations from past experience and generalize them to novel and unseen problems. This helps them learn more quickly with fewer data. In this work, we explore if machines can be endowed with such reasoning abilities in the context of science-based question answering.
|
| 12 |
+
|
| 13 |
+
Recently, science problem solving benchmarks [18] have been used to diagnose the multi-hop reasoning ability and interpretability of AI systems. To answer science questions, a model needs to not only understand multimodal contents but also extract external knowledge to arrive at the correct answer. Since these tasks require domain-specific knowledge and explicit multi-hop reasoning, a model would be not interpretable if it fails to provide explanations to reveal the reasoning process. However, current science question datasets [18, 17, 52] mostly lack annotated explanations for the answers. To address this issue, other science datasets annotate the explanations, but they are restricted to the textual only modality and limited to small data scales [13, 6, 37] or a small set of topics [20, 14]. Therefore, we collect Science Question Answering (SCIENCEQA), a large-scale multi-choice dataset that contains multimodal science questions with explanations and features rich domain diversity.
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
Figure 1: We construct the SCIENCEQA dataset where a data example consists of multimodal question answering information and the grounded lecture and explanation. We study if QA models can generate a reasonable explanation to reveal the chain-of-thought reasoning.
|
| 17 |
+
|
| 18 |
+
SCIENCEQA is collected from elementary and high school science curricula, and contains 21,208 examples along with lectures and explanations. Different from existing datasets [17, 18, 52], SCIENCEQA has richer domain diversity from three different subjects: natural science, social science, and language science. A typical example consists of a question, multiple choices, multimodal contexts, a correct answer, as well as a lecture and an explanation. The lecture and explanation provide general external knowledge and specific reasons, respectively, for arriving at the correct answer.
|
| 19 |
+
|
| 20 |
+
Consider the thoughts one person might have when answering the question in Figure 1. One first recalls the knowledge regarding the definition of a force learned from textbooks: “A force is a push or a pull that ... The direction of a push is ... The direction of a pull is ...”, then forms a line of reasoning: “The baby’s hand applies a force to the cabinet door. This force causes the door to open. $ T h e$ direction of this force is toward the baby’s hand.”, and finally arrives at the correct answer: “This force is a pull.”. Following [41], we formulate the task to output a natural explanation alongside the predicted answer. In this paper, we train language models to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process to answer SCIENCEQA questions.
|
| 21 |
+
|
| 22 |
+
Our experiments show that current multimodal methods [55, 1, 21, 9, 26, 35] fail to achieve satisfactory performance on SCIENCEQA and do not generate correct explanations. Instead, we find that CoT can help large language models not only in the few-shot learning setting but also in the fine-tuning setting. When combined with CoT to generate the lecture and explanation, the fine-tuned UnifiedQA [19] achieves an improvement of $3 . 9 9 \%$ as opposed to not using CoT in the fine-tuning stage. The few-shot GPT-3 model [4] via chain-of-thought prompting can obtain $7 5 . 1 7 \%$ on SCIENCEQA with an improvement of $1 . 2 0 \%$ compared to the few-shot GPT-3 without CoT. Prompted with CoT, GPT-3 can generate reasonable explanations as evaluated by automated metrics, and promisingly, $6 5 . 2 \%$ of explanations meet the gold standard of human evaluations. We also investigate the upper bound for models to harness explanations by including them in the input. We find that doing so improves GPT-3’s few-shot performance by $1 8 . 9 6 \%$ , suggesting that explanations do aid models and are currently underutilized in the CoT framework. Further analysis shows that, like humans, language models benefit from explanations to learn with less data: UnifiedQA with CoT obtains the same results as UnifiedQA without CoT with only $40 \%$ of the training data.
|
| 23 |
+
|
| 24 |
+
To sum up, our contributions are three-fold: (a) To bridge the gap in existing datasets in the scientific domain, we build Science Question Answering (SCIENCEQA), a new dataset containing 21,208 multimodal science questions with rich domain diversity. To the best of our knowledge, SCIENCEQA is the first large-scale multimodal dataset that annotates lectures and explanations for the answers.
|
| 25 |
+
|
| 26 |
+
(b) We show that CoT benefits large language models in both few-shot and fine-tuning learning by improving model performance and reliability via generating explanations. (c) We further explore the upper bound of GPT-3 and show that CoT helps language models learn from fewer data.
|
| 27 |
+
|
| 28 |
+
# 2 Related Work
|
| 29 |
+
|
| 30 |
+
Visual question answering. Since the task of visual question answering (VQA) was first proposed in [2], there have been plenty of VQA datasets [56, 58, 23, 11, 15, 12] conducted to facilitate the research work. Although our SCIENCEQA dataset shares some features with VQA, there are several main differences between them. First, SCIENCEQA is more challenging than existing VQA datasets because it contains multimodal contexts and diverse topics in the scientific domain. In addition, most answers are annotated with lectures and explanations, which makes SCIENCEQA a suitable dataset for multi-modal question answering and multi-hop reasoning for AI systems. Inspired by the recent remarkable performance achieved for VQA [33, 32, 10, 9, 26, 7, 8], in this paper, we further extensively benchmark SCIENCEQA with a wide range of attention-based [1, 33, 21, 9] and Transformer-based [30, 26, 27, 7] methods.
|
| 31 |
+
|
| 32 |
+
Datasets for science problems. Science problem solving is a challenging task that requires an AI system not only to understand the multimodal information from the science curriculum but also to reason about how to answer the domain-specific questions. Current science problem datasets such as AI2D [17], DVQA [16], VLQA [52], and FOODWEDS [24] have contributed to multimodal reasoning in the scientific domain. For example, a portion of VLQA contains multimodal questions on science subjects. These datasets, however, lack annotated explanations for the answers to reveal the reasoning steps. Some other datasets annotate the answers in the forms of supporting facts [37, 20], entailment trees [6], explanation graphs [13], reasoning chains [14]. However, these datasets are restricted to the single text modality with small data scales and limited topics. Instead, our SCIENCEQA annotates the answers with grounded lectures and explanations. Besides, SCIENCEQA features a richer domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills.
|
| 33 |
+
|
| 34 |
+
Learning from explanations and few-shot learning. Explanations help humans understand a task better, and there have been several attempts to show the same for models. For example, the learning from instruction paradigm [40, 43, 53, 39, 45, 25], where the task level explanation is provided in the form of instruction, improves model performance significantly. An example of learning from explanations in the scientific domain is proposed in [51] where the model interprets demonstrative solutions to solve geometry problems. Recently, there has been a surge of interest in few-shot learning, where language models learn a specific task from a few examples [46, 3]. For instance, [42, 54, 34] find that explanations in the format of the chain of thought can improve language models’ reasoning ability in few-shot learning. In this paper, we show that the chain of thought boosts the performance of large language models like UnifiedQA [19] if the models generate explanations along with the answer in a fine-tuning way. Furthermore, a few-shot GPT-3 model via chain-of-thought prompting is able to improve the reasoning performance on SCIENCEQA and generate reasonable explanations.
|
| 35 |
+
|
| 36 |
+
# 3 Dataset
|
| 37 |
+
|
| 38 |
+
We collect SCIENCEQA, which is a multimodal multiple-choice science question dataset containing 21,208 examples. An example in SCIENCEQA is shown in Figure 1. Given the science question and multimodal contexts, the task is to select the correct answer from multiple options. Different from existing datasets [50, 17, 52, 31, 24], SCIENCEQA covers diverse topics across three subjects: natural science, social science, and language science. Moreover, most questions are annotated with grounded lectures and detailed explanations. The lecture provides general knowledge that introduces the background information for solving problems of a similar class. The explanation reveals a specific reason for the answer. To effectively answer the questions, a model often needs to be able to understand the multimodal content in the input and extract external knowledge, similar to how humans do. More importantly, the goal of SCIENCEQA is to aid development of a reliable model that is capable of generating a coherent chain of thought when arriving at the correct answer to reveal the multi-step reasoning process. For data collection details, see Appendix A.1.
|
| 39 |
+
|
| 40 |
+
Table 1: Main statistics in SCIENCEQA.
|
| 41 |
+
|
| 42 |
+
<table><tr><td>Statistic</td><td>Number</td></tr><tr><td>Total questions</td><td>21,208</td></tr><tr><td>Questions with text context Questions with image context</td><td>10,220 (48.2%) 10,332 (48.7%)</td></tr><tr><td>* Image of natural format</td><td>~2,960 (14.0%)</td></tr><tr><td>* Image of diagram format</td><td>~7,372 (34.8%)</td></tr><tr><td>Questions with both contexts</td><td></td></tr><tr><td></td><td>6,532 (30.8%)</td></tr><tr><td>Questions without any context</td><td>7,188 (33.9%)</td></tr><tr><td>Questions with a lecture</td><td>17,798 (83.9%)</td></tr><tr><td>Questions with a explanation</td><td>19,202 (90.5%)</td></tr><tr><td>Different questions</td><td>9,122</td></tr><tr><td>Different lectures</td><td>261</td></tr><tr><td>Topic classes</td><td>26</td></tr><tr><td>Category classes</td><td>127</td></tr><tr><td>Skill classes</td><td>379</td></tr><tr><td>Average question length</td><td>12.11</td></tr><tr><td>Average choice length</td><td>4.40</td></tr><tr><td>Average lecture length</td><td>125.06</td></tr><tr><td></td><td></td></tr><tr><td>Average explanation length</td><td>47.66</td></tr></table>
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
Figure 2: Question distribution in SCIENCEQA.
|
| 46 |
+
|
| 47 |
+
# 3.1 Data Analysis
|
| 48 |
+
|
| 49 |
+
Key statistics. We randomly split the dataset into training, validation, and test splits with a ratio of 60:20:20. Each split has 12,726, 4,241, and 4,241 examples, respectively. Table 1 shows the main statistics of SCIENCEQA. SCIENCEQA has a large set of different questions, totaling up to 9,122. Out of the 21,208 questions in SCIENCEQA, 10,332 $( 4 8 . 7 \% )$ have an image context, 10,220 $( 4 8 . 2 \% )$ have a text context, and 6,532 $( 3 0 . 8 \% )$ have both. $8 3 . 9 \%$ of the questions are annotated with a lecture, while $9 0 . 5 \%$ of the questions feature an explanation. The cross-combination of these information sources diversifies the problem scenario: sometimes the model is given a lot of information from multiple sources, while at other times, the only source of information is the question itself. This level of complexity is very common in grade-level science exams.
|
| 50 |
+
|
| 51 |
+
(a) Question length distribution of related datasets. SCIENCEQA is distributed more evenly in terms of the number of question words than other datasets.
|
| 52 |
+
|
| 53 |
+
|
| 54 |
+
Figure 3: Question length distribution (a) and context distribution in SCIENCEQA (b).
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
(b) Question distribution with different context formats. $6 6 . 1 1 \%$ of the questions in SCIENCEQA have either an image or text context, while $3 0 . 8 0 \%$ have both.
|
| 58 |
+
|
| 59 |
+
Question analysis. SCIENCEQA has a diverse set of science questions. Figure 2 shows a distribution of the first four words in the question text. A large number of question lengths and formats highlight the diversity of SCIENCEQA. The question lengths range from 3 words to 141 words, and the questions in SCIENCEQA have an average length of 12.11 words. The question length distribution is visualized against other VQA datasets in Figure 3 (a). As shown in the diagram, SCIENCEQA’s distribution is flatter than other datasets, spanning more evenly across different question lengths.
|
| 60 |
+
|
| 61 |
+
Context analysis. Figure 3 (b) shows the number and percentage of questions with either an image context, a text context, or both. There are a total of 7,803 unique image contexts and 4,651 unique text contexts. $6 6 . 1 1 \%$ of the questions have at least one type of context information. The image context is in the format of diagrams or natural images, which visualize the critical scenario necessary for question answering or simply illustrate the question for better understanding. Similarly, the textual context can provide either semantically rich information or a simple hint to the question. Therefore, models need to be flexible and general to understand these diverse types of contexts.
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
Figure 4: Domain diversity in SCIENCEQA. Each color corresponds to one subject: natural science, social science, and language science. For visual clarity, only the most frequent classes are shown.
|
| 65 |
+
|
| 66 |
+
Domain diversity. Each SCIENCEQA question belongs to one of the three subjects: natural science, social science, and language science. With each subject, questions are categorized first by the topic (Biology, Physics, Chemistry, etc.), then by the category (Plants, Cells, Animals, etc.), and finally by the specific skill (Classify fruits and vegetables as plant parts, Identify countries of Africa, etc.). SCIENCEQA has a total of 26 topics, 127 categories, and 379 skills. The treemap in Figure 4 visualizes the different subjects, topics, and categories and shows that SCIENCEQA questions are very diverse, spanning a wide range of domains.
|
| 67 |
+
|
| 68 |
+
# 3.2 Comparisons with Existing Datasets
|
| 69 |
+
|
| 70 |
+
Table 2 shows a comparison of SCIENCEQA and other science problem datasets. As shown in the table, SCIENCEQA is much larger than most other datasets. SCIENCEQA also has the largest set of images, spans across all 12 grades, contains the longest questions, and has the most diverse input sources. As opposed to limiting the subject to only natural science, SCIENCEQA also includes social science and language science, largely adding to the domain diversity of the dataset. Furthermore, most of the questions in SCIENCEQA are annotated with textual lectures $( 8 3 . 9 \% )$ and explanations $( 9 0 . 5 \% )$ , which reveal the reasoning path to the correct answer. To the best of our knowledge, SCIENCEQA is the first large-scale multimodal science question dataset that annotates the answers with detailed lectures and explanations.
|
| 71 |
+
|
| 72 |
+
Table 2: Statistics for SCIENCEQA and comparisons with existing datasets. #Q: number of questions, #I: number of images, AvgQ: average question length; MaxQ: maximum question length.
|
| 73 |
+
|
| 74 |
+
<table><tr><td></td><td>#Q</td><td>#</td><td>AvgQ MaxQ Grades</td><td></td><td></td><td>Science subjects</td><td>Contexts</td><td>Images</td><td>Lecture Explanation</td><td></td></tr><tr><td>Geometry3K [31]</td><td>3,002</td><td>2,342</td><td>10.1</td><td>46</td><td>6-12</td><td>natural (geometry)</td><td>image</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>AI2D [17]</td><td>4,563</td><td>4,903</td><td>9.8</td><td>64</td><td>1-6</td><td>natural</td><td>image</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>FOODWEBS [24]</td><td>~5,000</td><td>~5,00</td><td></td><td>-</td><td>8</td><td>natural (foodweb only)</td><td>image</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>ARC [5]</td><td>7,787</td><td>0</td><td>20.4</td><td>128</td><td>3-9</td><td>natural</td><td>×</td><td>×</td><td>×</td><td>×</td></tr><tr><td>TQA [18]</td><td>26,260</td><td>3,455</td><td>9.2</td><td>57</td><td>6-8</td><td>natural</td><td>image,text</td><td>diagram</td><td>?</td><td>×</td></tr><tr><td>IconQA [35]</td><td>107,439 96,817</td><td></td><td>8.4</td><td>73</td><td>PreK-3</td><td>math</td><td>visual</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>WorldTree [13]</td><td>1,680</td><td>0</td><td>-</td><td>1</td><td>3-5</td><td>natural</td><td>×</td><td>×</td><td>×</td><td>√</td></tr><tr><td>OpenBookQA [37]</td><td>5,957</td><td>0</td><td>10.6</td><td>68</td><td>1-6</td><td>natural</td><td>×</td><td>×</td><td>×</td><td>?</td></tr><tr><td>QASC [20]</td><td>9,980</td><td>0</td><td>8.0</td><td>25</td><td>1-9</td><td>natural</td><td>X</td><td>×</td><td>×</td><td>L</td></tr><tr><td>SCIENCEQA (ours)</td><td>21,208</td><td>10,332</td><td>12.1</td><td>141</td><td>1-12</td><td>natural, social,language image,text natural,diagram</td><td></td><td></td><td>√</td><td></td></tr></table>
|
| 75 |
+
|
| 76 |
+
# 4 Baselines and Chain-of-Thought Models
|
| 77 |
+
|
| 78 |
+
In this section, we establish baselines and develop two chain-of-thought models on SCIENCEQA.
|
| 79 |
+
|
| 80 |
+
# 4.1 Baselines
|
| 81 |
+
|
| 82 |
+
Heuristic baselines. The first heuristic baseline is random chance: we randomly select one from the multiple options. Each trial is completed on the whole test set, and we take three different trials for an average result. The second heuristic baseline is human performance. We post the task to Amazon Mechanical Turk and ask workers to answer SCIENCEQA questions. Only workers who obtain a high school or higher degree and pass the qualification examples are qualified for the study. Each worker needs to answer a set of 10 test questions, and each question is answered by three different workers. For more details of the human performance study, see Appendix B.2.
|
| 83 |
+
|
| 84 |
+
Zero-shot and few-shot baselines. We establish the zero-shot baselines on top of UnifiedQA [19] and GPT-3 [4]. The zero-shot setup follows the format of $\mathrm { Q C M } { } \mathbf { A }$ where the input is the concatenation of tokens of the question text (Q), the context text (C), and multiple options (M), while the output is to predict the answer (A) from the option set. We extract the caption from the captioning model based on ViT [7] and GPT-2 [47] for the image as the visual context. In the few-shot setting, we follow the standard prompting [4] where in-context examples from the training set are concatenated before the test instance. These in-context examples serve as an instruction for the language model to adjust to the specific task in SCIENCEQA.
|
| 85 |
+
|
| 86 |
+
Fine-tuning baselines. We first consider the fine-tuning baselines from VQA models [1, 21, 55, 9, 22, 35, 26] proposed in recent years. These VQA baselines take the question, the context, and choices as the textual input, take the image as the visual input, and predict the score distribution over choice candidates via a linear classifier. In addition, we build the fine-tuning baseline on top of the large language model UnifiedQA [19]. UnifiedQA takes the textual information as the input and outputs the answer option. Similarly, the image is converted into a caption that provides the visual semantics for the language model.
|
| 87 |
+
|
| 88 |
+
# 4.2 Language Models with the Chain of Thought
|
| 89 |
+
|
| 90 |
+
A chain of thought refers to a coherent flow of sentences that reveals the premises and conclusion of a reasoning problem [54]. A chain of thought clearly decomposes a multi-hop reasoning task into intermediate steps instead of solving the task in a black-box way. The chain of thought can be the step-by-step thought process [54] before arriving at the final answer or explanations [41] that come after the answer. The annotated lectures and explanations in SCIENCEQA serve as demonstrations of the chain of thought that mimics the multi-step reasoning steps of human beings. In this paper, we study if large language models can generate reasonable explanations as the chain of thought to reveal the thought process when answering SCIENCEQA questions. Further, we explore how the chain of thought can improve the reasoning ability of language models on SCIENCEQA in both few-shot and fine-tuning learning.
|
| 91 |
+
|
| 92 |
+
UnifiedQA with the chain of thought. UnifiedQA [19] is a state of the art model for multi-option question answering. The original architecture of UnifiedQA takes the question and options as the input and outputs a short phrase as the final answer. We make a format modification to develop UnifiedQA with the chain of thought (CoT), i.e., UnifiedQA is fine-tuned to generate a long sequence of text which consists of the answer followed by the lecture and explanation.
|
| 93 |
+
|
| 94 |
+
GPT-3 via chain-of-thought prompting. Recent research work [4, 38, 34] has shown that GPT-3 [4] can perform various tasks when provided with in-context examples in a standard prompt. Take multi-option question answering as an example, the standard prompt [36, 57, 29] builds instructions using in-context examples with components of the question text, options, and the correct answer text. This style of few-shot learning enables the GPT-3 model to answer specific questions without parameter updates. Different from standard prompting, we build GPT-3 via chain-of-thought (CoT) prompting, as shown in Figure 5. To be specific, for each test problem $t$ , we map the prompt instruction $I : \{ I _ { i } \} _ { n } , I _ { t }$ into a textual format where $\{ I _ { i } \} _ { n }$ refers to the instruction set of $n$ -shot in-context examples from the training set, while $I _ { t }$ denotes the test instruction. Instead of the way where the explanation comes before the answer [54], we feed the instruction $I$ into the encoderdecoder model GPT-3 to generate the answer $a$ followed by the lecture lect and explanation exp: $M : \{ I _ { i } \} _ { n } , I _ { t } \to a , l e c t , e x p$ .
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
Figure 5: Prompt instruction encoding for the test example $t$ in GPT-3 (CoT). The prompt above consists of the instruction $\{ I _ { i } \} _ { 1 }$ for the 1-shot training example and $I _ { t }$ for the test example.
|
| 98 |
+
|
| 99 |
+
# 5 Experiments
|
| 100 |
+
|
| 101 |
+
# 5.1 Experimental Setup
|
| 102 |
+
|
| 103 |
+
Evaluation metrics. The heuristics and VQA baselines treat our SCIENCEQA task as a multi-class classification problem with multiple options and are evaluated with the accuracy metrics. UnifiedQA and GPT-3 treat SCIENCEQA as a text generation problem. So the most similar option is selected as the final prediction to evaluate the question answering accuracy. The generated lectures and explanations are evaluated by automatic metrics [44, 28, 49] and human scores by annotators.
|
| 104 |
+
|
| 105 |
+
Implementation details. The VQA baselines are trained for a maximum number of 50 epochs with a learning rate of $5 e { - 5 }$ . We fine-tune the UnifiedQA for $5 0 k$ iterations and evaluate every $1 k$ iteration. The training process is stopped following the early stopping strategy with a patience period of three evaluations. For GPT-3, we use the text-davinci-002 engine, which is the most capable model version suggested in the official documentation. More details can be found in Appendix B.1.
|
| 106 |
+
|
| 107 |
+
# 5.2 Results for Question Answering
|
| 108 |
+
|
| 109 |
+
Table 3 demonstrates the empirical results for Science Question Answering.
|
| 110 |
+
|
| 111 |
+
VQA baselines. We feed the VQA baseline models with the input of QCM format to predict answers A. Out of all the VQA models we benchmarked, VisualBERT [26, 27] performs the best on average $( 6 1 . 8 7 \% )$ . Interestingly, Patch-TRM [35] beats VisualBERT in natural science (NAT) and language science (LAN), and it also performs better in higher-grade questions $6 7 . 5 0 \%$ v.s. $5 9 . 9 2 \%$ ). However, in the subject of social science (SOC), VisualBERT outperforms Patch-TRM by a large margin $( + 2 2 . 3 9 \% )$ . Such drastic changes in performance might imply that current VQA models are not generalized to process the challenging questions in SCIENCEQA.
|
| 112 |
+
|
| 113 |
+
Language models. We evaluate whether large-scale pretraining on text can help language models learn scientific knowledge and thus perform better on the SCIENCEQA task. For this purpose, we have tried two of the state-of-the-art pre-trained language models: UnifiedQA and GPT-3.
|
| 114 |
+
|
| 115 |
+
(i) UnifiedQA. The results show that without any supervised fine-tuning (zero-shot), UnifiedQA cannot beat any VQA baseline model, while the pretraining does help the model obtain some scientific knowledge to outperform the random baseline. When fine-tuned with the answer labels in SCIENCEQA, UnifiedQABASE reports an accuracy of $7 0 . 1 2 \%$ on average. By further teaching the model to generate the answer along with lecture and explanation, the developed language model with chain-of-thought (UnifiedQABASE (CoT)) brings additional improvements of $+ 3 . 2 1 \%$ $\mathrm { ( Q C M \to A E }$ ) and $+ 3 . 9 9 \%$ ( $\mathrm { Q C M } $ ALE). These results show that generating the chain of thought along with the answer benefits the reasoning ability of language models.
|
| 116 |
+
|
| 117 |
+
(ii) GPT-3. The positive effect of pretraining is also proved by the surprisingly good results from GPT-3 in the same zero-shot setting as UnifiedQA. Without any fine-tuning, GPT-3 already reaches almost the best performance we can get. Interestingly, prompting the GPT-3 with two training examples with only answers results in a negligible difference. However, if we prompt GPT-3 with chain-of-thought prompting (QCM ALE), we obtain the state-of-the-art result so far $( 7 5 . 1 7 \% )$ .
|
| 118 |
+
|
| 119 |
+
Figure 6: One example of the predicted answer along with the chain of thought from GPT-3 (CoT).
|
| 120 |
+
|
| 121 |
+
<table><tr><td>Model</td><td>Learning</td><td>Format</td><td>NAT</td><td>SOC</td><td>LAN</td><td>TXT</td><td>IMG</td><td>NO</td><td>G1-6</td><td>G7-12</td><td>Avg</td></tr><tr><td>Random chance</td><td>-</td><td>M→A</td><td>40.28</td><td>46.13</td><td>29.25</td><td>47.45</td><td>40.08</td><td>33.66</td><td>39.35</td><td>40.67</td><td>39.83</td></tr><tr><td>Qonly [1]</td><td>train set</td><td>Q→A</td><td>41.34</td><td>27.22</td><td>47.00</td><td>41.79</td><td>35.15</td><td>44.60</td><td>39.28</td><td>40.87</td><td>39.85</td></tr><tr><td>C1 only[1]</td><td>train set</td><td>C1→A</td><td>41.34</td><td>29.25</td><td>45.45</td><td>42.33</td><td>36.09</td><td>42.93</td><td>39.21</td><td>41.07</td><td>39.87</td></tr><tr><td>Q+Monly [1]</td><td>train set</td><td>QM→A</td><td>52.66</td><td>51.86</td><td>60.18</td><td>55.57</td><td>50.37</td><td>57.42</td><td>52.53</td><td>57.88</td><td>54.44</td></tr><tr><td>Q+Cr+Monly [1]</td><td>train set</td><td>QCrM→A</td><td>57.28</td><td>49.04</td><td>61.36</td><td>60.46</td><td>52.80</td><td>58.82</td><td>54.44</td><td>60.51</td><td>56.61</td></tr><tr><td>Q+C1+Monly [1]</td><td>train set</td><td>QCiM→A</td><td>58.97</td><td>53.77</td><td>60.45</td><td>62.85</td><td>54.49</td><td>57.63</td><td>56.72</td><td>61.04</td><td>58.26</td></tr><tr><td>MCAN [55]</td><td>train set</td><td>QCM→A</td><td>56.08</td><td>46.23</td><td>58.09</td><td>59.43</td><td>51.17</td><td>55.40</td><td>51.65</td><td>59.72</td><td>54.54</td></tr><tr><td>Top-Down [1]</td><td>train set</td><td>QCM→A</td><td>59.50</td><td>54.33</td><td>61.82</td><td>62.90</td><td>54.88</td><td>59.79</td><td>57.27</td><td>62.16</td><td>59.02</td></tr><tr><td>BAN [21]</td><td>train set</td><td>QCM→A</td><td>60.88</td><td>46.57</td><td>66.64</td><td>62.61</td><td>52.60</td><td>65.51</td><td>56.83</td><td>63.94</td><td>59.37</td></tr><tr><td>DFAF [9]</td><td>train set</td><td>QCM→A</td><td>64.03</td><td>48.82</td><td>63.55</td><td>65.88</td><td>54.49</td><td>64.11</td><td>57.12</td><td>67.17</td><td>60.72</td></tr><tr><td>ViLT [22]</td><td>train set</td><td>QCM→A</td><td>60.48</td><td>63.89</td><td>60.27</td><td>63.20</td><td>61.38</td><td>57.00</td><td>60.72</td><td>61.90</td><td>61.14</td></tr><tr><td>Patch-TRM[35]</td><td>train set</td><td>QCM→A</td><td>65.19</td><td>46.79</td><td>65.55</td><td>66.96</td><td>55.28</td><td>64.95</td><td>58.04</td><td>67.50</td><td>61.42</td></tr><tr><td>VisualBERT [26,27]</td><td>train set</td><td>QCM→A</td><td>59.33</td><td>69.18</td><td>61.18</td><td>62.71</td><td>62.17</td><td>58.54</td><td>62.96</td><td>59.92</td><td>61.87</td></tr><tr><td>UnifiedQAsMALL [48]</td><td>zero-shot</td><td>QCM→A</td><td>47.78</td><td>40.49</td><td>46.00</td><td>50.24</td><td>44.12</td><td>44.39</td><td>45.56</td><td>46.21</td><td>45.79</td></tr><tr><td>UnifiedQABASE [48]</td><td>zero-shot</td><td>QCM→A</td><td>50.13</td><td>44.54</td><td>48.18</td><td>53.08</td><td>48.09</td><td>46.69</td><td>47.58</td><td>50.03</td><td>48.46</td></tr><tr><td>UnifiedQAsMALL [48]</td><td>train set</td><td>QCM→A</td><td>53.77</td><td>58.04</td><td>61.09</td><td>52.10</td><td>51.51</td><td>61.46</td><td>58.22</td><td>53.59</td><td>56.57</td></tr><tr><td>UnifiedQABASE [48]</td><td>train set</td><td>QCM→A</td><td>68.16</td><td>69.18</td><td>74.91</td><td>63.78</td><td>61.38</td><td>77.84</td><td>72.98</td><td>65.00</td><td>70.12</td></tr><tr><td>UnifiedQABASE (CoT)</td><td>train set</td><td>QCM→AE</td><td>70.60</td><td>74.02</td><td>78.36</td><td>65.69</td><td>64.80</td><td>81.53</td><td>75.48</td><td>69.48</td><td>73.333.21↑</td></tr><tr><td>UnifiedQABASE (CoT)</td><td>train set</td><td>QCM→ALE</td><td>71.00</td><td>76.04</td><td>78.91</td><td>66.42</td><td>66.53</td><td>81.81</td><td>77.06</td><td>68.82</td><td>74.113.99↑</td></tr><tr><td>GPT-3 [4]</td><td>zero-shot</td><td>QCM→A</td><td>75.04</td><td>66.59</td><td>78.00</td><td>74.24</td><td>65.74</td><td>79.58</td><td>76.36</td><td>69.87</td><td>74.04</td></tr><tr><td>GPT-3 [4]</td><td>2-shot</td><td>QCM→A</td><td>74.64</td><td>69.74</td><td>76.00</td><td>74.44</td><td>67.28</td><td>77.42</td><td>76.80</td><td>68.89</td><td>73.97</td></tr><tr><td>GPT-3 (CoT)</td><td>2-shot</td><td>QCM→AE</td><td>76.60</td><td>65.92</td><td>77.55</td><td>75.51</td><td>66.09</td><td>79.58</td><td>78.49</td><td>67.63</td><td>74.610.64↑</td></tr><tr><td>GPT-3 (CoT)</td><td>2-shot</td><td>QCM→ALE</td><td>75.44</td><td>70.87</td><td>78.09</td><td>74.68</td><td>67.43</td><td>79.93</td><td>78.23</td><td>69.68</td><td>75.171.20↑</td></tr><tr><td>Human</td><td>-</td><td>QCM→A</td><td>90.23</td><td>84.97</td><td>87.48</td><td>89.60</td><td>87.50</td><td>88.10</td><td>91.59</td><td>82.42</td><td>88.40</td></tr></table>
|
| 122 |
+
|
| 123 |
+
Table 3: Evaluation of baselines over different classes in accuracy $( \% )$ . Model names: $\mathrm { Q } =$ question, $\mathbf { M } =$ multiple options, $\mathbf { C } =$ context, $\mathrm { C } _ { T } =$ text context, $\mathrm { C } _ { I } =$ image context, $\mathbf { C o T = }$ chain of thought. Format names: $\mathbf { A } =$ answer, $\mathbf { A } \mathbf { E } =$ answer with explanation, $\mathbf { A L E = }$ answer with lecture and explanation. Question classes: $\mathbf { N A T } =$ natural science, ${ \bf S O C = }$ social science, $\mathrm { L A N } =$ language science, $\mathrm { T X T = }$ text context, $\mathbf { I M G } =$ image context, ${ \mathrm { N O } } =$ no context, G1- $\cdot 6 =$ grades 1-6, G7- $\cdot 1 2 =$ grades 7-12. Segments 1: Random chance; Segment 2: Ablation studies on top of Top-Down; Segment 3: VQA baselines; Segment 4: UnifiedQA baselines and UnifiedQA with CoT; Segment 5: GPT-3 baselines and GPT-3 with CoT; Segment 6: Average human performance.
|
| 124 |
+
|
| 125 |
+
Question: What do these two changes have in common?\n
|
| 126 |
+
bleaching clothes\n
|
| 127 |
+
a piece of apple turning brown\n
|
| 128 |
+
Options: (A) Both are only physical changes. (B) Both are caused by cooling. (C) Both are chemical changes. (D) Both are caused by heating. $\backslash \mathsf { n }$
|
| 129 |
+
Answer: The answer is C. BECAUSE: A physical change is a change to a substance that does not change what the substance is. A physical change only changes the substance’s physical properties, such as its size, shape, or color.\n
|
| 130 |
+
A chemical change is a change to a substance that changes what the substance is. A chemical change changes the substance’s chemical properties, such as its chemical structure. $\backslash \mathsf { n }$
|
| 131 |
+
The changes that you listed are both chemical changes.\n
|
| 132 |
+
When you bleach clothes, the chemical structure of the fabric changes. This change makes the fabric less likely to hold color. $\backslash \mathsf { n }$ When a piece of fruit turns brown, the chemical structure of the fruit changes. This change makes the fruit taste different.
|
| 133 |
+
|
| 134 |
+
Human performance. Humans outperform all benchmarks consistently across question classes, context types, and grades, e.g., a $2 0 . 0 7 \%$ gap for questions with the image context (IMG) between humans and our best performing model. The gap is to be filled by future research on multimodal reasoning for scientific question answering.
|
| 135 |
+
|
| 136 |
+
# 5.3 Results for Generated Explanations
|
| 137 |
+
|
| 138 |
+
One prediction example of GPT-3 (CoT) is visualized in Figure 6. We can see that GPT-3 (CoT) predicts the correct answer and generates a reasonable lecture and explanation to mimic the human thought process. We further report automatic metrics (BLEU-1/4 [44], ROUGE-L [44], and (sentence)
|
| 139 |
+
|
| 140 |
+
Similarity [49] to evaluate the generated lectures and explanations, as shown in Table 4. The Similarity metric computes the cosine-similarity of semantic embeddings between two sentences based on the Sentence-BERT network [49]. The results show that UnifiedQABASE (CoT) generates the most similar explanations to the given ones. However, it’s commonly agreed that automatic evaluation of generated texts only provides a partial view and has to be complemented by a human study. By asking annotators to rate the relevance, correctness, and completeness of generated explanations, we find that the explanations generated by GPT-3 (CoT) conform best to human judgment.
|
| 141 |
+
|
| 142 |
+
<table><tr><td>Model</td><td>Format</td><td>BLEU-1</td><td>BLEU-4 ROUGE-L</td><td></td><td>Similarity</td><td>Relevant Correct Complete</td><td>Gold</td></tr><tr><td>UnifiedQABASE ：(CoT)</td><td>QCM→ALE</td><td>0.397 0.370</td><td>0.714</td><td>0.811</td><td>80.4%</td><td>76.6% 76.1%</td><td>56.9%</td></tr><tr><td>GPT-3 (CoT)</td><td>QCM→AE</td><td>0.234 0.048</td><td>0.351</td><td>0.561</td><td>76.9%</td><td>73.0% 70.5%</td><td>52.5%</td></tr><tr><td>GPT-3 (CoT)</td><td>QCM→ALE</td><td>0.192 0.052</td><td>0.323</td><td>0.595</td><td>88.5%</td><td>78.8% 84.5%</td><td>65.2%</td></tr></table>
|
| 143 |
+
|
| 144 |
+
Table 4: Automatic metrics (BLEU-1/4, ROUGE-L, Similarity) and human evaluation of generated explanations. Note that a gold explanation refers to one that is relevant, correct, and complete.
|
| 145 |
+
|
| 146 |
+
# 5.4 Analysis
|
| 147 |
+
|
| 148 |
+
Blind studies. Blind studies are conducted on top of the modification of the full model, Top-Down [1]. The results achieved in blind studies of Q only and $\mathrm { C } _ { I }$ only are close to random chance, showing that the SCIENCEQA dataset is robust and reliable in distribution. The performance drops in $\mathbf { Q } { + } \mathbf { M }$ only, $\mathrm { Q + C } _ { T } { + } \mathrm { M }$ only, and $\mathrm { Q + C } _ { I } + \mathrm { M }$ only indicate that all input components provide critical information for answering SCIENCEQA questions.
|
| 149 |
+
|
| 150 |
+
Prompt types. We study the effect of prompt types and visualize the comparison in Figure 7 (a). It shows that prompting the GPT-3 model with both lectures and explanations $( \mathrm { Q C M } { } \mathrm { A L E }$ ) results in the highest accuracy on average and the smallest variance. In contrast, prompting with only explanations $( \mathrm { Q C M } { } \mathrm { A E }$ ) gives the largest variance, resulting in a less stable model.
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
(a) Acc. v.s. different prompts with 4-shot examples.
|
| 154 |
+
(b) Acc. v.s. different # of training examples.
|
| 155 |
+
Figure 7: Accuracy of GPT-3 (CoT) cross different prompt types (a) and # of training examples (b).
|
| 156 |
+
|
| 157 |
+
Number of in-context examples. In Figure 7 (b), we further investigate how different numbers of training examples encoded in prompts can affect the prediction accuracy. The $\mathrm { Q C M } { } \mathrm { A L E }$ prompt type outperforms or performs comparably to the $\mathrm { Q C M } { } \mathbf { A }$ type with all numbers of examples. And we observe the peak performance of $\mathrm { Q C M } .$ ALE with 2 training examples being prompted. After that, the accuracy goes down as more training examples are added to the model.
|
| 158 |
+
|
| 159 |
+
Dynamic sampling. In Table 5, instead of random sampling, we try to dynamically select the in-context examples to prompt with the same class as the test sample. However, slight differences in prediction accuracy are observed when comparing them to simple random sampling.
|
| 160 |
+
|
| 161 |
+
<table><tr><td>Prompt type</td><td>Sampling</td><td>Acc.(%)</td></tr><tr><td>QCM→ALE</td><td>Dynamic (same topic)</td><td>75.15</td></tr><tr><td>QCM→ALE</td><td>Dynamic (same category)</td><td>74.58</td></tr><tr><td>QCM→ALE</td><td>Dynamic (same skill)</td><td>75.10</td></tr></table>
|
| 162 |
+
|
| 163 |
+
Table 5: Dynamic sampling for GPT-3 (CoT).
|
| 164 |
+
|
| 165 |
+
Upper bound. We search the upper bound of the GPT-3 accuracy by feeding the gold lecture and explanation in the test prompt. As reported in Table 6, $\mathrm { Q C M E ^ { * } { } A }$ outperforms the $\mathrm { Q C M } { } \mathrm { A L E }$ baseline by $1 8 . 8 6 \%$ and $\mathrm { Q C M L E ^ { * } { } A }$ outperforms $\mathrm { Q C M } { } \mathrm { A L E }$ by $1 8 . 9 6 \%$ , indicating a potential improvement direction by generating correct explanations before answering science questions.
|
| 166 |
+
|
| 167 |
+
Table 6: Upper bound of GPT-3 (CoT).
|
| 168 |
+
|
| 169 |
+
<table><tr><td>Prompt type</td><td>Sampling</td><td>Acc.(%)</td></tr><tr><td>QCML*→A</td><td>Random</td><td>73.59</td></tr><tr><td>QCML*→AE</td><td>Random</td><td>74.32</td></tr><tr><td>QCME*→A</td><td>Random</td><td>94.0318.86↑</td></tr><tr><td>QCMLE*→A</td><td>Random</td><td>94.1318.96↑</td></tr><tr><td>QCM→ALE</td><td>Random</td><td>75.17</td></tr></table>
|
| 170 |
+
|
| 171 |
+
<table><tr><td>Prompt type</td><td>Sampling</td><td>Acc.(%)</td></tr><tr><td>QCM→LA</td><td>Random</td><td>60.6</td></tr><tr><td>QCM→EA</td><td>Random</td><td>56.0</td></tr><tr><td>QCM→LEA</td><td>Random</td><td>55.4</td></tr><tr><td>QCM→ELA</td><td>Random</td><td>51.5</td></tr><tr><td>QCM→ALE</td><td>Random</td><td>73.6</td></tr></table>
|
| 172 |
+
|
| 173 |
+
Table 7: Different positions of L/E for GPT-3 (CoT).
|
| 174 |
+
|
| 175 |
+
Positions of lectures and explanations. We study the performance of GPT-3 (CoT) in terms of different positions of lectures and explanations on 1,000 test examples. The results are shown in Table 7. There could be huge accuracy decreases if GPT-3 (CoT) predicts lectures and explanations before answers. It is mainly because if GPT-3 (CoT) is formulated to generate the long lecture and explanation first, there is a greater chance that it will stop generating the prediction early or use up the maximum token limits before obtaining the required answer.
|
| 176 |
+
|
| 177 |
+
CoT learns with fewer data. To study if the chain of thought helps language models learn more efficiently, we report the accuracies of UnifiedQA and UnifiedQA (CoT) fine-tuned on different sizes of the training set in Figure 8. UnifiedQA (CoT) benefits language models by learning the coherent reasoning path when answering questions, resulting in similar accuracy with fewer training examples.
|
| 178 |
+
|
| 179 |
+
Error analysis. GPT-3 via chain-of-thought prompting obtains promising results but still fails to answer a wide range of challenging questions in SCIENCEQA. See examples of failure cases in Appendix B.4. The failure cases can be classified into two types: (a) the model fails to understand the multimodal inputs and lacks domain-specific knowledge to arrive at the correct answer; (b) the model generates the wrong chain of thought with irrelevant, incorrect, or incomplete information.
|
| 180 |
+
|
| 181 |
+
|
| 182 |
+
Figure 8: UnifiedQA (CoT) learns efficiently with fewer training examples.
|
| 183 |
+
|
| 184 |
+
# 6 Discussion and Conclusion
|
| 185 |
+
|
| 186 |
+
In this paper, we propose SCIENCEQA, a dataset that features 21,208 multi-option questions with multimodal contexts from the science curriculum. To the best of our knowledge, SCIENCEQA is the first large-scale multimodal science dataset where most questions are annotated with corresponding lectures and explanations. We establish various baselines, including recent VQA models and large language models on SCIENCEQA. We further study if language models can generate reasonable explanations and then benefit the reasoning ability. Experiments show that UnifiedQA with the chain of thought can achieve an improvement of $3 . 9 9 \%$ and few-shot GPT-3 via chain-of-thought (CoT) prompting can obtain a satisfactory accuracy of $7 5 . 1 7 \%$ on SCIENCEQA. $6 5 . 2 \%$ of the generated explanations from GPT-3 (CoT) meet the gold standard by human evaluations.
|
| 187 |
+
|
| 188 |
+
# 7 Acknowledgment
|
| 189 |
+
|
| 190 |
+
We would like to thank the anonymous reviewers for their valuable comments and suggestions. We would also like to thank Xiaodan Liang for insightful discussions on dataset collection. We thank our colleagues at The Allen Institute of AI (AI2), Jiasen Lu and Jungo Kasai for helpful discussions. The work does not relate to Liang Qiu’s position at Amazon Alexa.
|
| 191 |
+
|
| 192 |
+
# References
|
| 193 |
+
|
| 194 |
+
[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.
|
| 195 |
+
[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision (CVPR), pages 2425–2433, 2015.
|
| 196 |
+
|
| 197 |
+
[3] Jonathan Bragg, Arman Cohan, Kyle Lo, and Iz Beltagy. Flex: Unifying evaluation for few-shot nlp. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021.
|
| 198 |
+
|
| 199 |
+
[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877–1901, 2020.
|
| 200 |
+
|
| 201 |
+
[5] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018.
|
| 202 |
+
|
| 203 |
+
[6] Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. Explaining answers with entailment trees. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021.
|
| 204 |
+
|
| 205 |
+
[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations (ICLR), 2021.
|
| 206 |
+
|
| 207 |
+
[8] Feng Gao, Qing Ping, Govind Thattai, Aishwarya Reganti, Ying Nian Wu, and Prem Natarajan. Transformretrieve-generate: Natural language-centric outside-knowledge visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5067–5077, June 2022.
|
| 208 |
+
|
| 209 |
+
[9] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6639–6648, 2019.
|
| 210 |
+
|
| 211 |
+
[10] Peng Gao, Hongsheng Li, Shuang Li, Pan Lu, Yikang Li, Steven CH Hoi, and Xiaogang Wang. Questionguided hybrid convolution for visual question answering. In The European Conference on Computer Vision (ECCV), pages 469–485, 2018.
|
| 212 |
+
|
| 213 |
+
[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
|
| 214 |
+
|
| 215 |
+
[12] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6700–6709, 2019.
|
| 216 |
+
|
| 217 |
+
[13] Peter A Jansen, Elizabeth Wainwright, Steven Marmorstein, and Clayton T Morrison. Worldtree: A corpus of explanation graphs for elementary science questions supporting multi-hop inference. arXiv preprint arXiv:1802.03052, 2018.
|
| 218 |
+
|
| 219 |
+
[14] Harsh Jhamtani and Peter Clark. Learning to explain: Datasets and models for identifying valid reasoning chains in multihop question-answering. arXiv preprint arXiv:2010.03274, 2020.
|
| 220 |
+
|
| 221 |
+
[15] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2901–2910, 2017.
|
| 222 |
+
|
| 223 |
+
[16] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5648–5656, 2018.
|
| 224 |
+
|
| 225 |
+
[17] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Min Joon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Proceedings of the European Conference on Computer Vision (ECCV), 2016.
|
| 226 |
+
|
| 227 |
+
[18] Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4999–5007, 2017.
|
| 228 |
+
|
| 229 |
+
[19] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings of the Association for Computational Linguistics (EMNLP), pages 1896–1907, 2020.
|
| 230 |
+
|
| 231 |
+
[20] Tushar Khot, Peter Clark, Michal Guerquin, Peter Alexander Jansen, and Ashish Sabharwal. Qasc: A dataset for question answering via sentence composition. ArXiv, abs/1910.11473, 2020.
|
| 232 |
+
[21] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 1571–1581, 2018.
|
| 233 |
+
[22] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 5583–5594, 2021.
|
| 234 |
+
[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision (IJCV), pages 32–73, 2017.
|
| 235 |
+
[24] Jayant Krishnamurthy, Oyvind Tafjord, and Aniruddha Kembhavi. Semantic parsing to probabilistic programs for situated question answering. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 160–170, 2016.
|
| 236 |
+
[25] Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329, 2022.
|
| 237 |
+
[26] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019.
|
| 238 |
+
[27] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5265–5275, 2020.
|
| 239 |
+
[28] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81, 2004.
|
| 240 |
+
[29] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021.
|
| 241 |
+
[30] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems (NeurIPS), pages 13–23, 2019.
|
| 242 |
+
[31] Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. Intergps: Interpretable geometry problem solving with formal language and symbolic reasoning. In The 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021.
|
| 243 |
+
[32] Pan Lu, Lei Ji, Wei Zhang, Nan Duan, Ming Zhou, and Jianyong Wang. R-vqa: learning visual relation facts with semantic attention for visual question answering. In The ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (SIGKDD), pages 1880–1889, 2018.
|
| 244 |
+
[33] Pan Lu, Hongsheng Li, Wei Zhang, Jianyong Wang, and Xiaogang Wang. Co-attending free-form regions and detections with multi-modal multiplicative feature embedding for visual question answering. In The AAAI Conference on Artificial Intelligence (AAAI), 2018.
|
| 245 |
+
[34] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022.
|
| 246 |
+
[35] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021.
|
| 247 |
+
[36] Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786, 2021.
|
| 248 |
+
[37] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2018.
|
| 249 |
+
[38] Swaroop Mishra, Matthew Finlayson, Pan Lu, Leonard Tang, Sean Welleck, Chitta Baral, Tanmay Rajpurohit, Oyvind Tafjord, Ashish Sabharwal, Peter Clark, and Ashwin Kalyan. Lila: A unified benchmark for mathematical reasoning. In The 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2022.
|
| 250 |
+
[39] Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Reframing instructional prompts to gptk’s language. ACL Findings, 2021.
|
| 251 |
+
[40] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. The 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021.
|
| 252 |
+
[41] Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. Wt5?! training text-to-text models to explain their predictions. arXiv preprint arXiv:2004.14546, 2020.
|
| 253 |
+
[42] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021.
|
| 254 |
+
[43] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022.
|
| 255 |
+
[44] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (ACL), pages 311–318, 2002.
|
| 256 |
+
[45] Mihir Parmar, Swaroop Mishra, Mirali Purohit, Man Luo, Murad Mohammad, and Chitta Baral. InBoXBART: Get instructions into biomedical multi-task learning. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 112–128, Seattle, United States, July 2022. Association for Computational Linguistics.
|
| 257 |
+
[46] Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021.
|
| 258 |
+
[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
|
| 259 |
+
[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21:1–67, 2020.
|
| 260 |
+
[49] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 11 2019.
|
| 261 |
+
[50] Mrinmaya Sachan, Kumar Dubey, and Eric Xing. From textbooks to knowledge: A case study in harvesting axiomatic knowledge from textbooks to solve geometry problems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 773–784, 2017.
|
| 262 |
+
[51] Mrinmaya Sachan and Eric Xing. Learning to solve geometry problems from natural language demonstrations in textbooks. In Proceedings of the 6th Joint Conference on Lexical and Computational Semantics (\* SEM 2017), pages 251–261, 2017.
|
| 263 |
+
[52] Shailaja Keyur Sampat, Yezhou Yang, and Chitta Baral. Visuo-lingustic question answering (vlqa) challenge. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings (EMNLP), pages 4606–4616, 2020.
|
| 264 |
+
[53] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. The International Conference on Learning Representations (ICLR), 2021.
|
| 265 |
+
[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022.
|
| 266 |
+
[55] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6281–6290, 2019.
|
| 267 |
+
[56] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and Yang: Balancing and answering binary visual questions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
|
| 268 |
+
[57] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning (ICML), pages 12697–12706. PMLR, 2021.
|
| 269 |
+
[58] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
|
| 270 |
+
|
| 271 |
+
# Checklist
|
| 272 |
+
|
| 273 |
+
1. For all authors...
|
| 274 |
+
|
| 275 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 276 |
+
(b) Did you describe the limitations of your work? [Yes] Yes, we did the error analysis in Section 5.4 and discussed the limitations of the work in Appendix B.4.
|
| 277 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] We discussed the broader impacts in Appendix B.5.
|
| 278 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 279 |
+
|
| 280 |
+
2. If you are including theoretical results...
|
| 281 |
+
|
| 282 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
|
| 283 |
+
|
| 284 |
+
3. If you ran experiments...
|
| 285 |
+
|
| 286 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We included 100 data examples and the data visualizer tool in the supplemental material. The whole dataset and code will be available at https://scienceqa.github.io.
|
| 287 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section 5.1 and Appendix B.1 for experimental details.
|
| 288 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] We reported the error bars for GPT-3 (CoT) experiments in Figure 7, where each experiment was repeated four times.
|
| 289 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] We discussed compute resources in Appendix B.1.
|
| 290 |
+
|
| 291 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 292 |
+
|
| 293 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes] We collected the SCIENCEQA dataset from https://www.ixl.com/. The copyright belongs to IXL.
|
| 294 |
+
(b) Did you mention the license of the assets? [Yes] SCIENCEQA is under the CC BY-NCSA 4.0 license and is used for non-commercial research purposes.
|
| 295 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We included data examples and a visualizer tool in the supplemental material. The dataset will be available at https://scienceqa.github.io.
|
| 296 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
|
| 297 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] The collected data does not contain personally identifiable information or offensive content.
|
| 298 |
+
|
| 299 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 300 |
+
|
| 301 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes] We included screenshots of the instructions in Appendix B.2 and B.3.
|
| 302 |
+
|
| 303 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 304 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [Yes] We included the monetary compensation details in Appendix B.2 and B.3.
|
parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc_content_list.json
ADDED
|
@@ -0,0 +1,1379 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
205,
|
| 8 |
+
122,
|
| 9 |
+
794,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Pan $\\mathbf { L u ^ { 1 , 3 } }$ , Swaroop Mishra2,3, Tony $\\mathbf { X i a } ^ { 1 }$ , Liang $\\mathbf { Q i u } ^ { 1 }$ , Kai-Wei Chang1, Song-Chun $\\mathbf { Z } \\mathbf { h } \\mathbf { u } ^ { 1 }$ , Oyvind Tafjord3, Peter Clark3, Ashwin Kalyan3 1University of California, Los Angeles, 2Arizona State University, 3Allen Institute for AI {lupantech, kwchang.cs}@gmail.com, sczhu@stat.ucla.edu, {oyvindt, peterc, ashwinkv}@allenai.org ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
207,
|
| 19 |
+
223,
|
| 20 |
+
795,
|
| 21 |
+
296
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Abstract ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
462,
|
| 31 |
+
332,
|
| 32 |
+
535,
|
| 33 |
+
348
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "When answering a question, humans utilize the information available across different modalities to synthesize a consistent and complete chain of thought (CoT). This process is normally a black box in the case of deep learning models like large-scale language models. Recently, science question benchmarks have been used to diagnose the multi-hop reasoning ability and interpretability of an AI system. However, existing datasets fail to provide annotations for the answers, or are restricted to the textual-only modality, small scales, and limited domain diversity. To this end, we present Science Question Answering (SCIENCEQA), a new benchmark that consists of ${ \\sim } 2 1 \\mathrm { k }$ multimodal multiple choice questions with diverse science topics and annotations of their answers with corresponding lectures and explanations. We further design language models to learn to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process when answering SCIENCEQA questions. SCIENCEQA demonstrates the utility of CoT in language models, as CoT improves the question answering performance by $1 . 2 0 \\%$ in fewshot GPT-3 and $3 . 9 9 \\%$ in fine-tuned UnifiedQA. We also explore the upper bound for models to leverage explanations by feeding those in the input; we observe that it improves the few-shot performance of GPT-3 by $1 8 . 9 6 \\%$ . Our analysis further shows that language models, similar to humans, benefit from explanations to learn from fewer data and achieve the same performance with just $40 \\%$ of the data.1 ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
233,
|
| 42 |
+
364,
|
| 43 |
+
766,
|
| 44 |
+
627
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 Introduction ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
174,
|
| 54 |
+
655,
|
| 55 |
+
310,
|
| 56 |
+
671
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "A long-standing goal of AI systems is to act reliably and learn complex tasks efficiently like human beings. In the process of reliable decision making, humans follow an explicit chain-of-thought (CoT) reasoning process that is typically expressed as an explanation. However, machine learning models are trained mostly using a large number of input-output examples to perform a specific task. These black-box models only generate the final decision without reliably revealing the underlying reasoning process. Not surprisingly, it is unclear if they understand the task and can generalize even though they perform well on the benchmark. On the other hand, humans are able to learn from instructions or explanations from past experience and generalize them to novel and unseen problems. This helps them learn more quickly with fewer data. In this work, we explore if machines can be endowed with such reasoning abilities in the context of science-based question answering. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
686,
|
| 66 |
+
825,
|
| 67 |
+
825
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Recently, science problem solving benchmarks [18] have been used to diagnose the multi-hop reasoning ability and interpretability of AI systems. To answer science questions, a model needs to not only understand multimodal contents but also extract external knowledge to arrive at the correct answer. Since these tasks require domain-specific knowledge and explicit multi-hop reasoning, a model would be not interpretable if it fails to provide explanations to reveal the reasoning process. However, current science question datasets [18, 17, 52] mostly lack annotated explanations for the answers. To address this issue, other science datasets annotate the explanations, but they are restricted to the textual only modality and limited to small data scales [13, 6, 37] or a small set of topics [20, 14]. Therefore, we collect Science Question Answering (SCIENCEQA), a large-scale multi-choice dataset that contains multimodal science questions with explanations and features rich domain diversity. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
176,
|
| 76 |
+
832,
|
| 77 |
+
820,
|
| 78 |
+
859
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "image",
|
| 84 |
+
"img_path": "images/bbe5265bbdcce0ffbf85cef9fd612ed120c97964ad993f721b8ee4dda5d131a6.jpg",
|
| 85 |
+
"image_caption": [
|
| 86 |
+
"Figure 1: We construct the SCIENCEQA dataset where a data example consists of multimodal question answering information and the grounded lecture and explanation. We study if QA models can generate a reasonable explanation to reveal the chain-of-thought reasoning. "
|
| 87 |
+
],
|
| 88 |
+
"image_footnote": [],
|
| 89 |
+
"bbox": [
|
| 90 |
+
173,
|
| 91 |
+
88,
|
| 92 |
+
821,
|
| 93 |
+
241
|
| 94 |
+
],
|
| 95 |
+
"page_idx": 1
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"type": "text",
|
| 99 |
+
"text": "",
|
| 100 |
+
"bbox": [
|
| 101 |
+
174,
|
| 102 |
+
332,
|
| 103 |
+
825,
|
| 104 |
+
443
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "SCIENCEQA is collected from elementary and high school science curricula, and contains 21,208 examples along with lectures and explanations. Different from existing datasets [17, 18, 52], SCIENCEQA has richer domain diversity from three different subjects: natural science, social science, and language science. A typical example consists of a question, multiple choices, multimodal contexts, a correct answer, as well as a lecture and an explanation. The lecture and explanation provide general external knowledge and specific reasons, respectively, for arriving at the correct answer. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
449,
|
| 114 |
+
825,
|
| 115 |
+
532
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "Consider the thoughts one person might have when answering the question in Figure 1. One first recalls the knowledge regarding the definition of a force learned from textbooks: “A force is a push or a pull that ... The direction of a push is ... The direction of a pull is ...”, then forms a line of reasoning: “The baby’s hand applies a force to the cabinet door. This force causes the door to open. $ T h e$ direction of this force is toward the baby’s hand.”, and finally arrives at the correct answer: “This force is a pull.”. Following [41], we formulate the task to output a natural explanation alongside the predicted answer. In this paper, we train language models to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process to answer SCIENCEQA questions. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
173,
|
| 124 |
+
539,
|
| 125 |
+
825,
|
| 126 |
+
650
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "Our experiments show that current multimodal methods [55, 1, 21, 9, 26, 35] fail to achieve satisfactory performance on SCIENCEQA and do not generate correct explanations. Instead, we find that CoT can help large language models not only in the few-shot learning setting but also in the fine-tuning setting. When combined with CoT to generate the lecture and explanation, the fine-tuned UnifiedQA [19] achieves an improvement of $3 . 9 9 \\%$ as opposed to not using CoT in the fine-tuning stage. The few-shot GPT-3 model [4] via chain-of-thought prompting can obtain $7 5 . 1 7 \\%$ on SCIENCEQA with an improvement of $1 . 2 0 \\%$ compared to the few-shot GPT-3 without CoT. Prompted with CoT, GPT-3 can generate reasonable explanations as evaluated by automated metrics, and promisingly, $6 5 . 2 \\%$ of explanations meet the gold standard of human evaluations. We also investigate the upper bound for models to harness explanations by including them in the input. We find that doing so improves GPT-3’s few-shot performance by $1 8 . 9 6 \\%$ , suggesting that explanations do aid models and are currently underutilized in the CoT framework. Further analysis shows that, like humans, language models benefit from explanations to learn with less data: UnifiedQA with CoT obtains the same results as UnifiedQA without CoT with only $40 \\%$ of the training data. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
656,
|
| 136 |
+
825,
|
| 137 |
+
849
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "To sum up, our contributions are three-fold: (a) To bridge the gap in existing datasets in the scientific domain, we build Science Question Answering (SCIENCEQA), a new dataset containing 21,208 multimodal science questions with rich domain diversity. To the best of our knowledge, SCIENCEQA is the first large-scale multimodal dataset that annotates lectures and explanations for the answers. ",
|
| 144 |
+
"bbox": [
|
| 145 |
+
176,
|
| 146 |
+
856,
|
| 147 |
+
825,
|
| 148 |
+
911
|
| 149 |
+
],
|
| 150 |
+
"page_idx": 1
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"type": "text",
|
| 154 |
+
"text": "(b) We show that CoT benefits large language models in both few-shot and fine-tuning learning by improving model performance and reliability via generating explanations. (c) We further explore the upper bound of GPT-3 and show that CoT helps language models learn from fewer data. ",
|
| 155 |
+
"bbox": [
|
| 156 |
+
173,
|
| 157 |
+
90,
|
| 158 |
+
823,
|
| 159 |
+
133
|
| 160 |
+
],
|
| 161 |
+
"page_idx": 2
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"type": "text",
|
| 165 |
+
"text": "2 Related Work ",
|
| 166 |
+
"text_level": 1,
|
| 167 |
+
"bbox": [
|
| 168 |
+
174,
|
| 169 |
+
165,
|
| 170 |
+
321,
|
| 171 |
+
183
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 2
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "Visual question answering. Since the task of visual question answering (VQA) was first proposed in [2], there have been plenty of VQA datasets [56, 58, 23, 11, 15, 12] conducted to facilitate the research work. Although our SCIENCEQA dataset shares some features with VQA, there are several main differences between them. First, SCIENCEQA is more challenging than existing VQA datasets because it contains multimodal contexts and diverse topics in the scientific domain. In addition, most answers are annotated with lectures and explanations, which makes SCIENCEQA a suitable dataset for multi-modal question answering and multi-hop reasoning for AI systems. Inspired by the recent remarkable performance achieved for VQA [33, 32, 10, 9, 26, 7, 8], in this paper, we further extensively benchmark SCIENCEQA with a wide range of attention-based [1, 33, 21, 9] and Transformer-based [30, 26, 27, 7] methods. ",
|
| 178 |
+
"bbox": [
|
| 179 |
+
174,
|
| 180 |
+
204,
|
| 181 |
+
825,
|
| 182 |
+
343
|
| 183 |
+
],
|
| 184 |
+
"page_idx": 2
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"type": "text",
|
| 188 |
+
"text": "Datasets for science problems. Science problem solving is a challenging task that requires an AI system not only to understand the multimodal information from the science curriculum but also to reason about how to answer the domain-specific questions. Current science problem datasets such as AI2D [17], DVQA [16], VLQA [52], and FOODWEDS [24] have contributed to multimodal reasoning in the scientific domain. For example, a portion of VLQA contains multimodal questions on science subjects. These datasets, however, lack annotated explanations for the answers to reveal the reasoning steps. Some other datasets annotate the answers in the forms of supporting facts [37, 20], entailment trees [6], explanation graphs [13], reasoning chains [14]. However, these datasets are restricted to the single text modality with small data scales and limited topics. Instead, our SCIENCEQA annotates the answers with grounded lectures and explanations. Besides, SCIENCEQA features a richer domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills. ",
|
| 189 |
+
"bbox": [
|
| 190 |
+
174,
|
| 191 |
+
349,
|
| 192 |
+
825,
|
| 193 |
+
501
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "Learning from explanations and few-shot learning. Explanations help humans understand a task better, and there have been several attempts to show the same for models. For example, the learning from instruction paradigm [40, 43, 53, 39, 45, 25], where the task level explanation is provided in the form of instruction, improves model performance significantly. An example of learning from explanations in the scientific domain is proposed in [51] where the model interprets demonstrative solutions to solve geometry problems. Recently, there has been a surge of interest in few-shot learning, where language models learn a specific task from a few examples [46, 3]. For instance, [42, 54, 34] find that explanations in the format of the chain of thought can improve language models’ reasoning ability in few-shot learning. In this paper, we show that the chain of thought boosts the performance of large language models like UnifiedQA [19] if the models generate explanations along with the answer in a fine-tuning way. Furthermore, a few-shot GPT-3 model via chain-of-thought prompting is able to improve the reasoning performance on SCIENCEQA and generate reasonable explanations. ",
|
| 200 |
+
"bbox": [
|
| 201 |
+
174,
|
| 202 |
+
507,
|
| 203 |
+
825,
|
| 204 |
+
674
|
| 205 |
+
],
|
| 206 |
+
"page_idx": 2
|
| 207 |
+
},
|
| 208 |
+
{
|
| 209 |
+
"type": "text",
|
| 210 |
+
"text": "3 Dataset ",
|
| 211 |
+
"text_level": 1,
|
| 212 |
+
"bbox": [
|
| 213 |
+
174,
|
| 214 |
+
705,
|
| 215 |
+
267,
|
| 216 |
+
723
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "We collect SCIENCEQA, which is a multimodal multiple-choice science question dataset containing 21,208 examples. An example in SCIENCEQA is shown in Figure 1. Given the science question and multimodal contexts, the task is to select the correct answer from multiple options. Different from existing datasets [50, 17, 52, 31, 24], SCIENCEQA covers diverse topics across three subjects: natural science, social science, and language science. Moreover, most questions are annotated with grounded lectures and detailed explanations. The lecture provides general knowledge that introduces the background information for solving problems of a similar class. The explanation reveals a specific reason for the answer. To effectively answer the questions, a model often needs to be able to understand the multimodal content in the input and extract external knowledge, similar to how humans do. More importantly, the goal of SCIENCEQA is to aid development of a reliable model that is capable of generating a coherent chain of thought when arriving at the correct answer to reveal the multi-step reasoning process. For data collection details, see Appendix A.1. ",
|
| 223 |
+
"bbox": [
|
| 224 |
+
174,
|
| 225 |
+
744,
|
| 226 |
+
825,
|
| 227 |
+
911
|
| 228 |
+
],
|
| 229 |
+
"page_idx": 2
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"type": "table",
|
| 233 |
+
"img_path": "images/90d0ffac2fdbb8174736257b90aa1be2827914648c177c6ae689822b5e43190c.jpg",
|
| 234 |
+
"table_caption": [
|
| 235 |
+
"Table 1: Main statistics in SCIENCEQA. "
|
| 236 |
+
],
|
| 237 |
+
"table_footnote": [],
|
| 238 |
+
"table_body": "<table><tr><td>Statistic</td><td>Number</td></tr><tr><td>Total questions</td><td>21,208</td></tr><tr><td>Questions with text context Questions with image context</td><td>10,220 (48.2%) 10,332 (48.7%)</td></tr><tr><td>* Image of natural format</td><td>~2,960 (14.0%)</td></tr><tr><td>* Image of diagram format</td><td>~7,372 (34.8%)</td></tr><tr><td>Questions with both contexts</td><td></td></tr><tr><td></td><td>6,532 (30.8%)</td></tr><tr><td>Questions without any context</td><td>7,188 (33.9%)</td></tr><tr><td>Questions with a lecture</td><td>17,798 (83.9%)</td></tr><tr><td>Questions with a explanation</td><td>19,202 (90.5%)</td></tr><tr><td>Different questions</td><td>9,122</td></tr><tr><td>Different lectures</td><td>261</td></tr><tr><td>Topic classes</td><td>26</td></tr><tr><td>Category classes</td><td>127</td></tr><tr><td>Skill classes</td><td>379</td></tr><tr><td>Average question length</td><td>12.11</td></tr><tr><td>Average choice length</td><td>4.40</td></tr><tr><td>Average lecture length</td><td>125.06</td></tr><tr><td></td><td></td></tr><tr><td>Average explanation length</td><td>47.66</td></tr></table>",
|
| 239 |
+
"bbox": [
|
| 240 |
+
191,
|
| 241 |
+
84,
|
| 242 |
+
464,
|
| 243 |
+
351
|
| 244 |
+
],
|
| 245 |
+
"page_idx": 3
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"type": "image",
|
| 249 |
+
"img_path": "images/b2b497f521cd368b69a1f31122293c104dce22ad961c7e9dfc676b45c8d7885f.jpg",
|
| 250 |
+
"image_caption": [
|
| 251 |
+
"Figure 2: Question distribution in SCIENCEQA. "
|
| 252 |
+
],
|
| 253 |
+
"image_footnote": [],
|
| 254 |
+
"bbox": [
|
| 255 |
+
503,
|
| 256 |
+
90,
|
| 257 |
+
812,
|
| 258 |
+
334
|
| 259 |
+
],
|
| 260 |
+
"page_idx": 3
|
| 261 |
+
},
|
| 262 |
+
{
|
| 263 |
+
"type": "text",
|
| 264 |
+
"text": "3.1 Data Analysis ",
|
| 265 |
+
"text_level": 1,
|
| 266 |
+
"bbox": [
|
| 267 |
+
174,
|
| 268 |
+
395,
|
| 269 |
+
310,
|
| 270 |
+
410
|
| 271 |
+
],
|
| 272 |
+
"page_idx": 3
|
| 273 |
+
},
|
| 274 |
+
{
|
| 275 |
+
"type": "text",
|
| 276 |
+
"text": "Key statistics. We randomly split the dataset into training, validation, and test splits with a ratio of 60:20:20. Each split has 12,726, 4,241, and 4,241 examples, respectively. Table 1 shows the main statistics of SCIENCEQA. SCIENCEQA has a large set of different questions, totaling up to 9,122. Out of the 21,208 questions in SCIENCEQA, 10,332 $( 4 8 . 7 \\% )$ have an image context, 10,220 $( 4 8 . 2 \\% )$ have a text context, and 6,532 $( 3 0 . 8 \\% )$ have both. $8 3 . 9 \\%$ of the questions are annotated with a lecture, while $9 0 . 5 \\%$ of the questions feature an explanation. The cross-combination of these information sources diversifies the problem scenario: sometimes the model is given a lot of information from multiple sources, while at other times, the only source of information is the question itself. This level of complexity is very common in grade-level science exams. ",
|
| 277 |
+
"bbox": [
|
| 278 |
+
173,
|
| 279 |
+
421,
|
| 280 |
+
826,
|
| 281 |
+
546
|
| 282 |
+
],
|
| 283 |
+
"page_idx": 3
|
| 284 |
+
},
|
| 285 |
+
{
|
| 286 |
+
"type": "text",
|
| 287 |
+
"text": "(a) Question length distribution of related datasets. SCIENCEQA is distributed more evenly in terms of the number of question words than other datasets. ",
|
| 288 |
+
"bbox": [
|
| 289 |
+
173,
|
| 290 |
+
717,
|
| 291 |
+
501,
|
| 292 |
+
755
|
| 293 |
+
],
|
| 294 |
+
"page_idx": 3
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"type": "image",
|
| 298 |
+
"img_path": "images/a8fc123565ea6029cb4df49cf573e7d0e325e0fd955edb451d415a2ed5f6b8d1.jpg",
|
| 299 |
+
"image_caption": [
|
| 300 |
+
"Figure 3: Question length distribution (a) and context distribution in SCIENCEQA (b). "
|
| 301 |
+
],
|
| 302 |
+
"image_footnote": [],
|
| 303 |
+
"bbox": [
|
| 304 |
+
184,
|
| 305 |
+
559,
|
| 306 |
+
488,
|
| 307 |
+
710
|
| 308 |
+
],
|
| 309 |
+
"page_idx": 3
|
| 310 |
+
},
|
| 311 |
+
{
|
| 312 |
+
"type": "image",
|
| 313 |
+
"img_path": "images/d728d925e4b6375221b45f3f2e9d01c3b689d03fbaaa6e6abe88d77158aeac8f.jpg",
|
| 314 |
+
"image_caption": [
|
| 315 |
+
"(b) Question distribution with different context formats. $6 6 . 1 1 \\%$ of the questions in SCIENCEQA have either an image or text context, while $3 0 . 8 0 \\%$ have both. "
|
| 316 |
+
],
|
| 317 |
+
"image_footnote": [],
|
| 318 |
+
"bbox": [
|
| 319 |
+
586,
|
| 320 |
+
560,
|
| 321 |
+
799,
|
| 322 |
+
695
|
| 323 |
+
],
|
| 324 |
+
"page_idx": 3
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"type": "text",
|
| 328 |
+
"text": "Question analysis. SCIENCEQA has a diverse set of science questions. Figure 2 shows a distribution of the first four words in the question text. A large number of question lengths and formats highlight the diversity of SCIENCEQA. The question lengths range from 3 words to 141 words, and the questions in SCIENCEQA have an average length of 12.11 words. The question length distribution is visualized against other VQA datasets in Figure 3 (a). As shown in the diagram, SCIENCEQA’s distribution is flatter than other datasets, spanning more evenly across different question lengths. ",
|
| 329 |
+
"bbox": [
|
| 330 |
+
173,
|
| 331 |
+
794,
|
| 332 |
+
825,
|
| 333 |
+
877
|
| 334 |
+
],
|
| 335 |
+
"page_idx": 3
|
| 336 |
+
},
|
| 337 |
+
{
|
| 338 |
+
"type": "text",
|
| 339 |
+
"text": "Context analysis. Figure 3 (b) shows the number and percentage of questions with either an image context, a text context, or both. There are a total of 7,803 unique image contexts and 4,651 unique text contexts. $6 6 . 1 1 \\%$ of the questions have at least one type of context information. The image context is in the format of diagrams or natural images, which visualize the critical scenario necessary for question answering or simply illustrate the question for better understanding. Similarly, the textual context can provide either semantically rich information or a simple hint to the question. Therefore, models need to be flexible and general to understand these diverse types of contexts. ",
|
| 340 |
+
"bbox": [
|
| 341 |
+
173,
|
| 342 |
+
883,
|
| 343 |
+
820,
|
| 344 |
+
911
|
| 345 |
+
],
|
| 346 |
+
"page_idx": 3
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"type": "image",
|
| 350 |
+
"img_path": "images/3be2ddf228b4c036aeb745f000cd23f0863b810bbd988dd9e5d2ae35c73eee02.jpg",
|
| 351 |
+
"image_caption": [
|
| 352 |
+
"Figure 4: Domain diversity in SCIENCEQA. Each color corresponds to one subject: natural science, social science, and language science. For visual clarity, only the most frequent classes are shown. "
|
| 353 |
+
],
|
| 354 |
+
"image_footnote": [],
|
| 355 |
+
"bbox": [
|
| 356 |
+
176,
|
| 357 |
+
90,
|
| 358 |
+
823,
|
| 359 |
+
301
|
| 360 |
+
],
|
| 361 |
+
"page_idx": 4
|
| 362 |
+
},
|
| 363 |
+
{
|
| 364 |
+
"type": "text",
|
| 365 |
+
"text": "",
|
| 366 |
+
"bbox": [
|
| 367 |
+
174,
|
| 368 |
+
366,
|
| 369 |
+
825,
|
| 370 |
+
436
|
| 371 |
+
],
|
| 372 |
+
"page_idx": 4
|
| 373 |
+
},
|
| 374 |
+
{
|
| 375 |
+
"type": "text",
|
| 376 |
+
"text": "Domain diversity. Each SCIENCEQA question belongs to one of the three subjects: natural science, social science, and language science. With each subject, questions are categorized first by the topic (Biology, Physics, Chemistry, etc.), then by the category (Plants, Cells, Animals, etc.), and finally by the specific skill (Classify fruits and vegetables as plant parts, Identify countries of Africa, etc.). SCIENCEQA has a total of 26 topics, 127 categories, and 379 skills. The treemap in Figure 4 visualizes the different subjects, topics, and categories and shows that SCIENCEQA questions are very diverse, spanning a wide range of domains. ",
|
| 377 |
+
"bbox": [
|
| 378 |
+
174,
|
| 379 |
+
443,
|
| 380 |
+
825,
|
| 381 |
+
540
|
| 382 |
+
],
|
| 383 |
+
"page_idx": 4
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"type": "text",
|
| 387 |
+
"text": "3.2 Comparisons with Existing Datasets ",
|
| 388 |
+
"text_level": 1,
|
| 389 |
+
"bbox": [
|
| 390 |
+
176,
|
| 391 |
+
558,
|
| 392 |
+
464,
|
| 393 |
+
574
|
| 394 |
+
],
|
| 395 |
+
"page_idx": 4
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"type": "text",
|
| 399 |
+
"text": "Table 2 shows a comparison of SCIENCEQA and other science problem datasets. As shown in the table, SCIENCEQA is much larger than most other datasets. SCIENCEQA also has the largest set of images, spans across all 12 grades, contains the longest questions, and has the most diverse input sources. As opposed to limiting the subject to only natural science, SCIENCEQA also includes social science and language science, largely adding to the domain diversity of the dataset. Furthermore, most of the questions in SCIENCEQA are annotated with textual lectures $( 8 3 . 9 \\% )$ and explanations $( 9 0 . 5 \\% )$ , which reveal the reasoning path to the correct answer. To the best of our knowledge, SCIENCEQA is the first large-scale multimodal science question dataset that annotates the answers with detailed lectures and explanations. ",
|
| 400 |
+
"bbox": [
|
| 401 |
+
173,
|
| 402 |
+
584,
|
| 403 |
+
826,
|
| 404 |
+
709
|
| 405 |
+
],
|
| 406 |
+
"page_idx": 4
|
| 407 |
+
},
|
| 408 |
+
{
|
| 409 |
+
"type": "table",
|
| 410 |
+
"img_path": "images/3facd8eece2f0de8a670cebf930ecf7669bf6115fe46f4fd3e7bef2a144a0a66.jpg",
|
| 411 |
+
"table_caption": [
|
| 412 |
+
"Table 2: Statistics for SCIENCEQA and comparisons with existing datasets. #Q: number of questions, #I: number of images, AvgQ: average question length; MaxQ: maximum question length. "
|
| 413 |
+
],
|
| 414 |
+
"table_footnote": [],
|
| 415 |
+
"table_body": "<table><tr><td></td><td>#Q</td><td>#</td><td>AvgQ MaxQ Grades</td><td></td><td></td><td>Science subjects</td><td>Contexts</td><td>Images</td><td>Lecture Explanation</td><td></td></tr><tr><td>Geometry3K [31]</td><td>3,002</td><td>2,342</td><td>10.1</td><td>46</td><td>6-12</td><td>natural (geometry)</td><td>image</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>AI2D [17]</td><td>4,563</td><td>4,903</td><td>9.8</td><td>64</td><td>1-6</td><td>natural</td><td>image</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>FOODWEBS [24]</td><td>~5,000</td><td>~5,00</td><td></td><td>-</td><td>8</td><td>natural (foodweb only)</td><td>image</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>ARC [5]</td><td>7,787</td><td>0</td><td>20.4</td><td>128</td><td>3-9</td><td>natural</td><td>×</td><td>×</td><td>×</td><td>×</td></tr><tr><td>TQA [18]</td><td>26,260</td><td>3,455</td><td>9.2</td><td>57</td><td>6-8</td><td>natural</td><td>image,text</td><td>diagram</td><td>?</td><td>×</td></tr><tr><td>IconQA [35]</td><td>107,439 96,817</td><td></td><td>8.4</td><td>73</td><td>PreK-3</td><td>math</td><td>visual</td><td>diagram</td><td>×</td><td>×</td></tr><tr><td>WorldTree [13]</td><td>1,680</td><td>0</td><td>-</td><td>1</td><td>3-5</td><td>natural</td><td>×</td><td>×</td><td>×</td><td>√</td></tr><tr><td>OpenBookQA [37]</td><td>5,957</td><td>0</td><td>10.6</td><td>68</td><td>1-6</td><td>natural</td><td>×</td><td>×</td><td>×</td><td>?</td></tr><tr><td>QASC [20]</td><td>9,980</td><td>0</td><td>8.0</td><td>25</td><td>1-9</td><td>natural</td><td>X</td><td>×</td><td>×</td><td>L</td></tr><tr><td>SCIENCEQA (ours)</td><td>21,208</td><td>10,332</td><td>12.1</td><td>141</td><td>1-12</td><td>natural, social,language image,text natural,diagram</td><td></td><td></td><td>√</td><td></td></tr></table>",
|
| 416 |
+
"bbox": [
|
| 417 |
+
173,
|
| 418 |
+
724,
|
| 419 |
+
831,
|
| 420 |
+
862
|
| 421 |
+
],
|
| 422 |
+
"page_idx": 4
|
| 423 |
+
},
|
| 424 |
+
{
|
| 425 |
+
"type": "text",
|
| 426 |
+
"text": "4 Baselines and Chain-of-Thought Models ",
|
| 427 |
+
"text_level": 1,
|
| 428 |
+
"bbox": [
|
| 429 |
+
174,
|
| 430 |
+
88,
|
| 431 |
+
542,
|
| 432 |
+
107
|
| 433 |
+
],
|
| 434 |
+
"page_idx": 5
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"type": "text",
|
| 438 |
+
"text": "In this section, we establish baselines and develop two chain-of-thought models on SCIENCEQA. ",
|
| 439 |
+
"bbox": [
|
| 440 |
+
169,
|
| 441 |
+
123,
|
| 442 |
+
808,
|
| 443 |
+
137
|
| 444 |
+
],
|
| 445 |
+
"page_idx": 5
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"type": "text",
|
| 449 |
+
"text": "4.1 Baselines ",
|
| 450 |
+
"text_level": 1,
|
| 451 |
+
"bbox": [
|
| 452 |
+
174,
|
| 453 |
+
159,
|
| 454 |
+
277,
|
| 455 |
+
174
|
| 456 |
+
],
|
| 457 |
+
"page_idx": 5
|
| 458 |
+
},
|
| 459 |
+
{
|
| 460 |
+
"type": "text",
|
| 461 |
+
"text": "Heuristic baselines. The first heuristic baseline is random chance: we randomly select one from the multiple options. Each trial is completed on the whole test set, and we take three different trials for an average result. The second heuristic baseline is human performance. We post the task to Amazon Mechanical Turk and ask workers to answer SCIENCEQA questions. Only workers who obtain a high school or higher degree and pass the qualification examples are qualified for the study. Each worker needs to answer a set of 10 test questions, and each question is answered by three different workers. For more details of the human performance study, see Appendix B.2. ",
|
| 462 |
+
"bbox": [
|
| 463 |
+
174,
|
| 464 |
+
186,
|
| 465 |
+
825,
|
| 466 |
+
284
|
| 467 |
+
],
|
| 468 |
+
"page_idx": 5
|
| 469 |
+
},
|
| 470 |
+
{
|
| 471 |
+
"type": "text",
|
| 472 |
+
"text": "Zero-shot and few-shot baselines. We establish the zero-shot baselines on top of UnifiedQA [19] and GPT-3 [4]. The zero-shot setup follows the format of $\\mathrm { Q C M } { } \\mathbf { A }$ where the input is the concatenation of tokens of the question text (Q), the context text (C), and multiple options (M), while the output is to predict the answer (A) from the option set. We extract the caption from the captioning model based on ViT [7] and GPT-2 [47] for the image as the visual context. In the few-shot setting, we follow the standard prompting [4] where in-context examples from the training set are concatenated before the test instance. These in-context examples serve as an instruction for the language model to adjust to the specific task in SCIENCEQA. ",
|
| 473 |
+
"bbox": [
|
| 474 |
+
174,
|
| 475 |
+
290,
|
| 476 |
+
825,
|
| 477 |
+
400
|
| 478 |
+
],
|
| 479 |
+
"page_idx": 5
|
| 480 |
+
},
|
| 481 |
+
{
|
| 482 |
+
"type": "text",
|
| 483 |
+
"text": "Fine-tuning baselines. We first consider the fine-tuning baselines from VQA models [1, 21, 55, 9, 22, 35, 26] proposed in recent years. These VQA baselines take the question, the context, and choices as the textual input, take the image as the visual input, and predict the score distribution over choice candidates via a linear classifier. In addition, we build the fine-tuning baseline on top of the large language model UnifiedQA [19]. UnifiedQA takes the textual information as the input and outputs the answer option. Similarly, the image is converted into a caption that provides the visual semantics for the language model. ",
|
| 484 |
+
"bbox": [
|
| 485 |
+
173,
|
| 486 |
+
406,
|
| 487 |
+
825,
|
| 488 |
+
503
|
| 489 |
+
],
|
| 490 |
+
"page_idx": 5
|
| 491 |
+
},
|
| 492 |
+
{
|
| 493 |
+
"type": "text",
|
| 494 |
+
"text": "4.2 Language Models with the Chain of Thought ",
|
| 495 |
+
"text_level": 1,
|
| 496 |
+
"bbox": [
|
| 497 |
+
174,
|
| 498 |
+
525,
|
| 499 |
+
526,
|
| 500 |
+
540
|
| 501 |
+
],
|
| 502 |
+
"page_idx": 5
|
| 503 |
+
},
|
| 504 |
+
{
|
| 505 |
+
"type": "text",
|
| 506 |
+
"text": "A chain of thought refers to a coherent flow of sentences that reveals the premises and conclusion of a reasoning problem [54]. A chain of thought clearly decomposes a multi-hop reasoning task into intermediate steps instead of solving the task in a black-box way. The chain of thought can be the step-by-step thought process [54] before arriving at the final answer or explanations [41] that come after the answer. The annotated lectures and explanations in SCIENCEQA serve as demonstrations of the chain of thought that mimics the multi-step reasoning steps of human beings. In this paper, we study if large language models can generate reasonable explanations as the chain of thought to reveal the thought process when answering SCIENCEQA questions. Further, we explore how the chain of thought can improve the reasoning ability of language models on SCIENCEQA in both few-shot and fine-tuning learning. ",
|
| 507 |
+
"bbox": [
|
| 508 |
+
173,
|
| 509 |
+
553,
|
| 510 |
+
825,
|
| 511 |
+
691
|
| 512 |
+
],
|
| 513 |
+
"page_idx": 5
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"type": "text",
|
| 517 |
+
"text": "UnifiedQA with the chain of thought. UnifiedQA [19] is a state of the art model for multi-option question answering. The original architecture of UnifiedQA takes the question and options as the input and outputs a short phrase as the final answer. We make a format modification to develop UnifiedQA with the chain of thought (CoT), i.e., UnifiedQA is fine-tuned to generate a long sequence of text which consists of the answer followed by the lecture and explanation. ",
|
| 518 |
+
"bbox": [
|
| 519 |
+
174,
|
| 520 |
+
696,
|
| 521 |
+
825,
|
| 522 |
+
766
|
| 523 |
+
],
|
| 524 |
+
"page_idx": 5
|
| 525 |
+
},
|
| 526 |
+
{
|
| 527 |
+
"type": "text",
|
| 528 |
+
"text": "GPT-3 via chain-of-thought prompting. Recent research work [4, 38, 34] has shown that GPT-3 [4] can perform various tasks when provided with in-context examples in a standard prompt. Take multi-option question answering as an example, the standard prompt [36, 57, 29] builds instructions using in-context examples with components of the question text, options, and the correct answer text. This style of few-shot learning enables the GPT-3 model to answer specific questions without parameter updates. Different from standard prompting, we build GPT-3 via chain-of-thought (CoT) prompting, as shown in Figure 5. To be specific, for each test problem $t$ , we map the prompt instruction $I : \\{ I _ { i } \\} _ { n } , I _ { t }$ into a textual format where $\\{ I _ { i } \\} _ { n }$ refers to the instruction set of $n$ -shot in-context examples from the training set, while $I _ { t }$ denotes the test instruction. Instead of the way where the explanation comes before the answer [54], we feed the instruction $I$ into the encoderdecoder model GPT-3 to generate the answer $a$ followed by the lecture lect and explanation exp: $M : \\{ I _ { i } \\} _ { n } , I _ { t } \\to a , l e c t , e x p$ . ",
|
| 529 |
+
"bbox": [
|
| 530 |
+
174,
|
| 531 |
+
772,
|
| 532 |
+
825,
|
| 533 |
+
911
|
| 534 |
+
],
|
| 535 |
+
"page_idx": 5
|
| 536 |
+
},
|
| 537 |
+
{
|
| 538 |
+
"type": "image",
|
| 539 |
+
"img_path": "images/876d55776abed6ad3892ae62fd2151924a63d53a1b4beb077052566aaa63f7d8.jpg",
|
| 540 |
+
"image_caption": [
|
| 541 |
+
"Figure 5: Prompt instruction encoding for the test example $t$ in GPT-3 (CoT). The prompt above consists of the instruction $\\{ I _ { i } \\} _ { 1 }$ for the 1-shot training example and $I _ { t }$ for the test example. "
|
| 542 |
+
],
|
| 543 |
+
"image_footnote": [],
|
| 544 |
+
"bbox": [
|
| 545 |
+
183,
|
| 546 |
+
85,
|
| 547 |
+
815,
|
| 548 |
+
207
|
| 549 |
+
],
|
| 550 |
+
"page_idx": 6
|
| 551 |
+
},
|
| 552 |
+
{
|
| 553 |
+
"type": "text",
|
| 554 |
+
"text": "",
|
| 555 |
+
"bbox": [
|
| 556 |
+
176,
|
| 557 |
+
262,
|
| 558 |
+
821,
|
| 559 |
+
291
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 6
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "text",
|
| 565 |
+
"text": "5 Experiments ",
|
| 566 |
+
"text_level": 1,
|
| 567 |
+
"bbox": [
|
| 568 |
+
174,
|
| 569 |
+
309,
|
| 570 |
+
312,
|
| 571 |
+
327
|
| 572 |
+
],
|
| 573 |
+
"page_idx": 6
|
| 574 |
+
},
|
| 575 |
+
{
|
| 576 |
+
"type": "text",
|
| 577 |
+
"text": "5.1 Experimental Setup ",
|
| 578 |
+
"text_level": 1,
|
| 579 |
+
"bbox": [
|
| 580 |
+
174,
|
| 581 |
+
340,
|
| 582 |
+
351,
|
| 583 |
+
356
|
| 584 |
+
],
|
| 585 |
+
"page_idx": 6
|
| 586 |
+
},
|
| 587 |
+
{
|
| 588 |
+
"type": "text",
|
| 589 |
+
"text": "Evaluation metrics. The heuristics and VQA baselines treat our SCIENCEQA task as a multi-class classification problem with multiple options and are evaluated with the accuracy metrics. UnifiedQA and GPT-3 treat SCIENCEQA as a text generation problem. So the most similar option is selected as the final prediction to evaluate the question answering accuracy. The generated lectures and explanations are evaluated by automatic metrics [44, 28, 49] and human scores by annotators. ",
|
| 590 |
+
"bbox": [
|
| 591 |
+
174,
|
| 592 |
+
366,
|
| 593 |
+
825,
|
| 594 |
+
435
|
| 595 |
+
],
|
| 596 |
+
"page_idx": 6
|
| 597 |
+
},
|
| 598 |
+
{
|
| 599 |
+
"type": "text",
|
| 600 |
+
"text": "Implementation details. The VQA baselines are trained for a maximum number of 50 epochs with a learning rate of $5 e { - 5 }$ . We fine-tune the UnifiedQA for $5 0 k$ iterations and evaluate every $1 k$ iteration. The training process is stopped following the early stopping strategy with a patience period of three evaluations. For GPT-3, we use the text-davinci-002 engine, which is the most capable model version suggested in the official documentation. More details can be found in Appendix B.1. ",
|
| 601 |
+
"bbox": [
|
| 602 |
+
174,
|
| 603 |
+
440,
|
| 604 |
+
825,
|
| 605 |
+
511
|
| 606 |
+
],
|
| 607 |
+
"page_idx": 6
|
| 608 |
+
},
|
| 609 |
+
{
|
| 610 |
+
"type": "text",
|
| 611 |
+
"text": "5.2 Results for Question Answering ",
|
| 612 |
+
"text_level": 1,
|
| 613 |
+
"bbox": [
|
| 614 |
+
174,
|
| 615 |
+
526,
|
| 616 |
+
434,
|
| 617 |
+
542
|
| 618 |
+
],
|
| 619 |
+
"page_idx": 6
|
| 620 |
+
},
|
| 621 |
+
{
|
| 622 |
+
"type": "text",
|
| 623 |
+
"text": "Table 3 demonstrates the empirical results for Science Question Answering. ",
|
| 624 |
+
"bbox": [
|
| 625 |
+
174,
|
| 626 |
+
551,
|
| 627 |
+
669,
|
| 628 |
+
566
|
| 629 |
+
],
|
| 630 |
+
"page_idx": 6
|
| 631 |
+
},
|
| 632 |
+
{
|
| 633 |
+
"type": "text",
|
| 634 |
+
"text": "VQA baselines. We feed the VQA baseline models with the input of QCM format to predict answers A. Out of all the VQA models we benchmarked, VisualBERT [26, 27] performs the best on average $( 6 1 . 8 7 \\% )$ . Interestingly, Patch-TRM [35] beats VisualBERT in natural science (NAT) and language science (LAN), and it also performs better in higher-grade questions $6 7 . 5 0 \\%$ v.s. $5 9 . 9 2 \\%$ ). However, in the subject of social science (SOC), VisualBERT outperforms Patch-TRM by a large margin $( + 2 2 . 3 9 \\% )$ . Such drastic changes in performance might imply that current VQA models are not generalized to process the challenging questions in SCIENCEQA. ",
|
| 635 |
+
"bbox": [
|
| 636 |
+
173,
|
| 637 |
+
573,
|
| 638 |
+
825,
|
| 639 |
+
670
|
| 640 |
+
],
|
| 641 |
+
"page_idx": 6
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"type": "text",
|
| 645 |
+
"text": "Language models. We evaluate whether large-scale pretraining on text can help language models learn scientific knowledge and thus perform better on the SCIENCEQA task. For this purpose, we have tried two of the state-of-the-art pre-trained language models: UnifiedQA and GPT-3. ",
|
| 646 |
+
"bbox": [
|
| 647 |
+
174,
|
| 648 |
+
676,
|
| 649 |
+
821,
|
| 650 |
+
718
|
| 651 |
+
],
|
| 652 |
+
"page_idx": 6
|
| 653 |
+
},
|
| 654 |
+
{
|
| 655 |
+
"type": "text",
|
| 656 |
+
"text": "(i) UnifiedQA. The results show that without any supervised fine-tuning (zero-shot), UnifiedQA cannot beat any VQA baseline model, while the pretraining does help the model obtain some scientific knowledge to outperform the random baseline. When fine-tuned with the answer labels in SCIENCEQA, UnifiedQABASE reports an accuracy of $7 0 . 1 2 \\%$ on average. By further teaching the model to generate the answer along with lecture and explanation, the developed language model with chain-of-thought (UnifiedQABASE (CoT)) brings additional improvements of $+ 3 . 2 1 \\%$ $\\mathrm { ( Q C M \\to A E }$ ) and $+ 3 . 9 9 \\%$ ( $\\mathrm { Q C M } $ ALE). These results show that generating the chain of thought along with the answer benefits the reasoning ability of language models. ",
|
| 657 |
+
"bbox": [
|
| 658 |
+
173,
|
| 659 |
+
724,
|
| 660 |
+
825,
|
| 661 |
+
835
|
| 662 |
+
],
|
| 663 |
+
"page_idx": 6
|
| 664 |
+
},
|
| 665 |
+
{
|
| 666 |
+
"type": "text",
|
| 667 |
+
"text": "(ii) GPT-3. The positive effect of pretraining is also proved by the surprisingly good results from GPT-3 in the same zero-shot setting as UnifiedQA. Without any fine-tuning, GPT-3 already reaches almost the best performance we can get. Interestingly, prompting the GPT-3 with two training examples with only answers results in a negligible difference. However, if we prompt GPT-3 with chain-of-thought prompting (QCM ALE), we obtain the state-of-the-art result so far $( 7 5 . 1 7 \\% )$ . ",
|
| 668 |
+
"bbox": [
|
| 669 |
+
174,
|
| 670 |
+
842,
|
| 671 |
+
823,
|
| 672 |
+
911
|
| 673 |
+
],
|
| 674 |
+
"page_idx": 6
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"type": "table",
|
| 678 |
+
"img_path": "images/35381b6302ad773eb1d16e25132af9cd51f567db318d2d94288e53036c04cb2c.jpg",
|
| 679 |
+
"table_caption": [
|
| 680 |
+
"Figure 6: One example of the predicted answer along with the chain of thought from GPT-3 (CoT). "
|
| 681 |
+
],
|
| 682 |
+
"table_footnote": [
|
| 683 |
+
"Table 3: Evaluation of baselines over different classes in accuracy $( \\% )$ . Model names: $\\mathrm { Q } =$ question, $\\mathbf { M } =$ multiple options, $\\mathbf { C } =$ context, $\\mathrm { C } _ { T } =$ text context, $\\mathrm { C } _ { I } =$ image context, $\\mathbf { C o T = }$ chain of thought. Format names: $\\mathbf { A } =$ answer, $\\mathbf { A } \\mathbf { E } =$ answer with explanation, $\\mathbf { A L E = }$ answer with lecture and explanation. Question classes: $\\mathbf { N A T } =$ natural science, ${ \\bf S O C = }$ social science, $\\mathrm { L A N } =$ language science, $\\mathrm { T X T = }$ text context, $\\mathbf { I M G } =$ image context, ${ \\mathrm { N O } } =$ no context, G1- $\\cdot 6 =$ grades 1-6, G7- $\\cdot 1 2 =$ grades 7-12. Segments 1: Random chance; Segment 2: Ablation studies on top of Top-Down; Segment 3: VQA baselines; Segment 4: UnifiedQA baselines and UnifiedQA with CoT; Segment 5: GPT-3 baselines and GPT-3 with CoT; Segment 6: Average human performance. "
|
| 684 |
+
],
|
| 685 |
+
"table_body": "<table><tr><td>Model</td><td>Learning</td><td>Format</td><td>NAT</td><td>SOC</td><td>LAN</td><td>TXT</td><td>IMG</td><td>NO</td><td>G1-6</td><td>G7-12</td><td>Avg</td></tr><tr><td>Random chance</td><td>-</td><td>M→A</td><td>40.28</td><td>46.13</td><td>29.25</td><td>47.45</td><td>40.08</td><td>33.66</td><td>39.35</td><td>40.67</td><td>39.83</td></tr><tr><td>Qonly [1]</td><td>train set</td><td>Q→A</td><td>41.34</td><td>27.22</td><td>47.00</td><td>41.79</td><td>35.15</td><td>44.60</td><td>39.28</td><td>40.87</td><td>39.85</td></tr><tr><td>C1 only[1]</td><td>train set</td><td>C1→A</td><td>41.34</td><td>29.25</td><td>45.45</td><td>42.33</td><td>36.09</td><td>42.93</td><td>39.21</td><td>41.07</td><td>39.87</td></tr><tr><td>Q+Monly [1]</td><td>train set</td><td>QM→A</td><td>52.66</td><td>51.86</td><td>60.18</td><td>55.57</td><td>50.37</td><td>57.42</td><td>52.53</td><td>57.88</td><td>54.44</td></tr><tr><td>Q+Cr+Monly [1]</td><td>train set</td><td>QCrM→A</td><td>57.28</td><td>49.04</td><td>61.36</td><td>60.46</td><td>52.80</td><td>58.82</td><td>54.44</td><td>60.51</td><td>56.61</td></tr><tr><td>Q+C1+Monly [1]</td><td>train set</td><td>QCiM→A</td><td>58.97</td><td>53.77</td><td>60.45</td><td>62.85</td><td>54.49</td><td>57.63</td><td>56.72</td><td>61.04</td><td>58.26</td></tr><tr><td>MCAN [55]</td><td>train set</td><td>QCM→A</td><td>56.08</td><td>46.23</td><td>58.09</td><td>59.43</td><td>51.17</td><td>55.40</td><td>51.65</td><td>59.72</td><td>54.54</td></tr><tr><td>Top-Down [1]</td><td>train set</td><td>QCM→A</td><td>59.50</td><td>54.33</td><td>61.82</td><td>62.90</td><td>54.88</td><td>59.79</td><td>57.27</td><td>62.16</td><td>59.02</td></tr><tr><td>BAN [21]</td><td>train set</td><td>QCM→A</td><td>60.88</td><td>46.57</td><td>66.64</td><td>62.61</td><td>52.60</td><td>65.51</td><td>56.83</td><td>63.94</td><td>59.37</td></tr><tr><td>DFAF [9]</td><td>train set</td><td>QCM→A</td><td>64.03</td><td>48.82</td><td>63.55</td><td>65.88</td><td>54.49</td><td>64.11</td><td>57.12</td><td>67.17</td><td>60.72</td></tr><tr><td>ViLT [22]</td><td>train set</td><td>QCM→A</td><td>60.48</td><td>63.89</td><td>60.27</td><td>63.20</td><td>61.38</td><td>57.00</td><td>60.72</td><td>61.90</td><td>61.14</td></tr><tr><td>Patch-TRM[35]</td><td>train set</td><td>QCM→A</td><td>65.19</td><td>46.79</td><td>65.55</td><td>66.96</td><td>55.28</td><td>64.95</td><td>58.04</td><td>67.50</td><td>61.42</td></tr><tr><td>VisualBERT [26,27]</td><td>train set</td><td>QCM→A</td><td>59.33</td><td>69.18</td><td>61.18</td><td>62.71</td><td>62.17</td><td>58.54</td><td>62.96</td><td>59.92</td><td>61.87</td></tr><tr><td>UnifiedQAsMALL [48]</td><td>zero-shot</td><td>QCM→A</td><td>47.78</td><td>40.49</td><td>46.00</td><td>50.24</td><td>44.12</td><td>44.39</td><td>45.56</td><td>46.21</td><td>45.79</td></tr><tr><td>UnifiedQABASE [48]</td><td>zero-shot</td><td>QCM→A</td><td>50.13</td><td>44.54</td><td>48.18</td><td>53.08</td><td>48.09</td><td>46.69</td><td>47.58</td><td>50.03</td><td>48.46</td></tr><tr><td>UnifiedQAsMALL [48]</td><td>train set</td><td>QCM→A</td><td>53.77</td><td>58.04</td><td>61.09</td><td>52.10</td><td>51.51</td><td>61.46</td><td>58.22</td><td>53.59</td><td>56.57</td></tr><tr><td>UnifiedQABASE [48]</td><td>train set</td><td>QCM→A</td><td>68.16</td><td>69.18</td><td>74.91</td><td>63.78</td><td>61.38</td><td>77.84</td><td>72.98</td><td>65.00</td><td>70.12</td></tr><tr><td>UnifiedQABASE (CoT)</td><td>train set</td><td>QCM→AE</td><td>70.60</td><td>74.02</td><td>78.36</td><td>65.69</td><td>64.80</td><td>81.53</td><td>75.48</td><td>69.48</td><td>73.333.21↑</td></tr><tr><td>UnifiedQABASE (CoT)</td><td>train set</td><td>QCM→ALE</td><td>71.00</td><td>76.04</td><td>78.91</td><td>66.42</td><td>66.53</td><td>81.81</td><td>77.06</td><td>68.82</td><td>74.113.99↑</td></tr><tr><td>GPT-3 [4]</td><td>zero-shot</td><td>QCM→A</td><td>75.04</td><td>66.59</td><td>78.00</td><td>74.24</td><td>65.74</td><td>79.58</td><td>76.36</td><td>69.87</td><td>74.04</td></tr><tr><td>GPT-3 [4]</td><td>2-shot</td><td>QCM→A</td><td>74.64</td><td>69.74</td><td>76.00</td><td>74.44</td><td>67.28</td><td>77.42</td><td>76.80</td><td>68.89</td><td>73.97</td></tr><tr><td>GPT-3 (CoT)</td><td>2-shot</td><td>QCM→AE</td><td>76.60</td><td>65.92</td><td>77.55</td><td>75.51</td><td>66.09</td><td>79.58</td><td>78.49</td><td>67.63</td><td>74.610.64↑</td></tr><tr><td>GPT-3 (CoT)</td><td>2-shot</td><td>QCM→ALE</td><td>75.44</td><td>70.87</td><td>78.09</td><td>74.68</td><td>67.43</td><td>79.93</td><td>78.23</td><td>69.68</td><td>75.171.20↑</td></tr><tr><td>Human</td><td>-</td><td>QCM→A</td><td>90.23</td><td>84.97</td><td>87.48</td><td>89.60</td><td>87.50</td><td>88.10</td><td>91.59</td><td>82.42</td><td>88.40</td></tr></table>",
|
| 686 |
+
"bbox": [
|
| 687 |
+
176,
|
| 688 |
+
90,
|
| 689 |
+
823,
|
| 690 |
+
438
|
| 691 |
+
],
|
| 692 |
+
"page_idx": 7
|
| 693 |
+
},
|
| 694 |
+
{
|
| 695 |
+
"type": "text",
|
| 696 |
+
"text": "Question: What do these two changes have in common?\\n \nbleaching clothes\\n \na piece of apple turning brown\\n \nOptions: (A) Both are only physical changes. (B) Both are caused by cooling. (C) Both are chemical changes. (D) Both are caused by heating. $\\backslash \\mathsf { n }$ \nAnswer: The answer is C. BECAUSE: A physical change is a change to a substance that does not change what the substance is. A physical change only changes the substance’s physical properties, such as its size, shape, or color.\\n \nA chemical change is a change to a substance that changes what the substance is. A chemical change changes the substance’s chemical properties, such as its chemical structure. $\\backslash \\mathsf { n }$ \nThe changes that you listed are both chemical changes.\\n \nWhen you bleach clothes, the chemical structure of the fabric changes. This change makes the fabric less likely to hold color. $\\backslash \\mathsf { n }$ When a piece of fruit turns brown, the chemical structure of the fruit changes. This change makes the fruit taste different. ",
|
| 697 |
+
"bbox": [
|
| 698 |
+
174,
|
| 699 |
+
579,
|
| 700 |
+
820,
|
| 701 |
+
718
|
| 702 |
+
],
|
| 703 |
+
"page_idx": 7
|
| 704 |
+
},
|
| 705 |
+
{
|
| 706 |
+
"type": "text",
|
| 707 |
+
"text": "Human performance. Humans outperform all benchmarks consistently across question classes, context types, and grades, e.g., a $2 0 . 0 7 \\%$ gap for questions with the image context (IMG) between humans and our best performing model. The gap is to be filled by future research on multimodal reasoning for scientific question answering. ",
|
| 708 |
+
"bbox": [
|
| 709 |
+
174,
|
| 710 |
+
770,
|
| 711 |
+
825,
|
| 712 |
+
825
|
| 713 |
+
],
|
| 714 |
+
"page_idx": 7
|
| 715 |
+
},
|
| 716 |
+
{
|
| 717 |
+
"type": "text",
|
| 718 |
+
"text": "5.3 Results for Generated Explanations ",
|
| 719 |
+
"text_level": 1,
|
| 720 |
+
"bbox": [
|
| 721 |
+
176,
|
| 722 |
+
843,
|
| 723 |
+
460,
|
| 724 |
+
858
|
| 725 |
+
],
|
| 726 |
+
"page_idx": 7
|
| 727 |
+
},
|
| 728 |
+
{
|
| 729 |
+
"type": "text",
|
| 730 |
+
"text": "One prediction example of GPT-3 (CoT) is visualized in Figure 6. We can see that GPT-3 (CoT) predicts the correct answer and generates a reasonable lecture and explanation to mimic the human thought process. We further report automatic metrics (BLEU-1/4 [44], ROUGE-L [44], and (sentence) ",
|
| 731 |
+
"bbox": [
|
| 732 |
+
174,
|
| 733 |
+
868,
|
| 734 |
+
825,
|
| 735 |
+
911
|
| 736 |
+
],
|
| 737 |
+
"page_idx": 7
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"type": "text",
|
| 741 |
+
"text": "Similarity [49] to evaluate the generated lectures and explanations, as shown in Table 4. The Similarity metric computes the cosine-similarity of semantic embeddings between two sentences based on the Sentence-BERT network [49]. The results show that UnifiedQABASE (CoT) generates the most similar explanations to the given ones. However, it’s commonly agreed that automatic evaluation of generated texts only provides a partial view and has to be complemented by a human study. By asking annotators to rate the relevance, correctness, and completeness of generated explanations, we find that the explanations generated by GPT-3 (CoT) conform best to human judgment. ",
|
| 742 |
+
"bbox": [
|
| 743 |
+
173,
|
| 744 |
+
90,
|
| 745 |
+
825,
|
| 746 |
+
189
|
| 747 |
+
],
|
| 748 |
+
"page_idx": 8
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"type": "table",
|
| 752 |
+
"img_path": "images/a1e8109ed6a1f8e188b772fc4288e3d89f5a6be9a3a9746ecffaa4e1f73b1cf7.jpg",
|
| 753 |
+
"table_caption": [],
|
| 754 |
+
"table_footnote": [
|
| 755 |
+
"Table 4: Automatic metrics (BLEU-1/4, ROUGE-L, Similarity) and human evaluation of generated explanations. Note that a gold explanation refers to one that is relevant, correct, and complete. "
|
| 756 |
+
],
|
| 757 |
+
"table_body": "<table><tr><td>Model</td><td>Format</td><td>BLEU-1</td><td>BLEU-4 ROUGE-L</td><td></td><td>Similarity</td><td>Relevant Correct Complete</td><td>Gold</td></tr><tr><td>UnifiedQABASE ：(CoT)</td><td>QCM→ALE</td><td>0.397 0.370</td><td>0.714</td><td>0.811</td><td>80.4%</td><td>76.6% 76.1%</td><td>56.9%</td></tr><tr><td>GPT-3 (CoT)</td><td>QCM→AE</td><td>0.234 0.048</td><td>0.351</td><td>0.561</td><td>76.9%</td><td>73.0% 70.5%</td><td>52.5%</td></tr><tr><td>GPT-3 (CoT)</td><td>QCM→ALE</td><td>0.192 0.052</td><td>0.323</td><td>0.595</td><td>88.5%</td><td>78.8% 84.5%</td><td>65.2%</td></tr></table>",
|
| 758 |
+
"bbox": [
|
| 759 |
+
176,
|
| 760 |
+
200,
|
| 761 |
+
823,
|
| 762 |
+
258
|
| 763 |
+
],
|
| 764 |
+
"page_idx": 8
|
| 765 |
+
},
|
| 766 |
+
{
|
| 767 |
+
"type": "text",
|
| 768 |
+
"text": "5.4 Analysis ",
|
| 769 |
+
"text_level": 1,
|
| 770 |
+
"bbox": [
|
| 771 |
+
173,
|
| 772 |
+
321,
|
| 773 |
+
272,
|
| 774 |
+
337
|
| 775 |
+
],
|
| 776 |
+
"page_idx": 8
|
| 777 |
+
},
|
| 778 |
+
{
|
| 779 |
+
"type": "text",
|
| 780 |
+
"text": "Blind studies. Blind studies are conducted on top of the modification of the full model, Top-Down [1]. The results achieved in blind studies of Q only and $\\mathrm { C } _ { I }$ only are close to random chance, showing that the SCIENCEQA dataset is robust and reliable in distribution. The performance drops in $\\mathbf { Q } { + } \\mathbf { M }$ only, $\\mathrm { Q + C } _ { T } { + } \\mathrm { M }$ only, and $\\mathrm { Q + C } _ { I } + \\mathrm { M }$ only indicate that all input components provide critical information for answering SCIENCEQA questions. ",
|
| 781 |
+
"bbox": [
|
| 782 |
+
173,
|
| 783 |
+
347,
|
| 784 |
+
826,
|
| 785 |
+
416
|
| 786 |
+
],
|
| 787 |
+
"page_idx": 8
|
| 788 |
+
},
|
| 789 |
+
{
|
| 790 |
+
"type": "text",
|
| 791 |
+
"text": "Prompt types. We study the effect of prompt types and visualize the comparison in Figure 7 (a). It shows that prompting the GPT-3 model with both lectures and explanations $( \\mathrm { Q C M } { } \\mathrm { A L E }$ ) results in the highest accuracy on average and the smallest variance. In contrast, prompting with only explanations $( \\mathrm { Q C M } { } \\mathrm { A E }$ ) gives the largest variance, resulting in a less stable model. ",
|
| 792 |
+
"bbox": [
|
| 793 |
+
173,
|
| 794 |
+
422,
|
| 795 |
+
825,
|
| 796 |
+
479
|
| 797 |
+
],
|
| 798 |
+
"page_idx": 8
|
| 799 |
+
},
|
| 800 |
+
{
|
| 801 |
+
"type": "image",
|
| 802 |
+
"img_path": "images/a2766859334855e5040f180d65c4fbbe5a8e73d5405ce43a36b1af8c81c655d0.jpg",
|
| 803 |
+
"image_caption": [
|
| 804 |
+
"(a) Acc. v.s. different prompts with 4-shot examples. ",
|
| 805 |
+
"(b) Acc. v.s. different # of training examples. ",
|
| 806 |
+
"Figure 7: Accuracy of GPT-3 (CoT) cross different prompt types (a) and # of training examples (b). "
|
| 807 |
+
],
|
| 808 |
+
"image_footnote": [],
|
| 809 |
+
"bbox": [
|
| 810 |
+
194,
|
| 811 |
+
489,
|
| 812 |
+
777,
|
| 813 |
+
636
|
| 814 |
+
],
|
| 815 |
+
"page_idx": 8
|
| 816 |
+
},
|
| 817 |
+
{
|
| 818 |
+
"type": "text",
|
| 819 |
+
"text": "Number of in-context examples. In Figure 7 (b), we further investigate how different numbers of training examples encoded in prompts can affect the prediction accuracy. The $\\mathrm { Q C M } { } \\mathrm { A L E }$ prompt type outperforms or performs comparably to the $\\mathrm { Q C M } { } \\mathbf { A }$ type with all numbers of examples. And we observe the peak performance of $\\mathrm { Q C M } .$ ALE with 2 training examples being prompted. After that, the accuracy goes down as more training examples are added to the model. ",
|
| 820 |
+
"bbox": [
|
| 821 |
+
173,
|
| 822 |
+
690,
|
| 823 |
+
825,
|
| 824 |
+
761
|
| 825 |
+
],
|
| 826 |
+
"page_idx": 8
|
| 827 |
+
},
|
| 828 |
+
{
|
| 829 |
+
"type": "text",
|
| 830 |
+
"text": "Dynamic sampling. In Table 5, instead of random sampling, we try to dynamically select the in-context examples to prompt with the same class as the test sample. However, slight differences in prediction accuracy are observed when comparing them to simple random sampling. ",
|
| 831 |
+
"bbox": [
|
| 832 |
+
174,
|
| 833 |
+
766,
|
| 834 |
+
485,
|
| 835 |
+
849
|
| 836 |
+
],
|
| 837 |
+
"page_idx": 8
|
| 838 |
+
},
|
| 839 |
+
{
|
| 840 |
+
"type": "table",
|
| 841 |
+
"img_path": "images/4ea89df5ad1d5f41b92768802c1bdc54c10414d61c160994d86444fc0a61e426.jpg",
|
| 842 |
+
"table_caption": [],
|
| 843 |
+
"table_footnote": [
|
| 844 |
+
"Table 5: Dynamic sampling for GPT-3 (CoT). "
|
| 845 |
+
],
|
| 846 |
+
"table_body": "<table><tr><td>Prompt type</td><td>Sampling</td><td>Acc.(%)</td></tr><tr><td>QCM→ALE</td><td>Dynamic (same topic)</td><td>75.15</td></tr><tr><td>QCM→ALE</td><td>Dynamic (same category)</td><td>74.58</td></tr><tr><td>QCM→ALE</td><td>Dynamic (same skill)</td><td>75.10</td></tr></table>",
|
| 847 |
+
"bbox": [
|
| 848 |
+
508,
|
| 849 |
+
770,
|
| 850 |
+
810,
|
| 851 |
+
830
|
| 852 |
+
],
|
| 853 |
+
"page_idx": 8
|
| 854 |
+
},
|
| 855 |
+
{
|
| 856 |
+
"type": "text",
|
| 857 |
+
"text": "Upper bound. We search the upper bound of the GPT-3 accuracy by feeding the gold lecture and explanation in the test prompt. As reported in Table 6, $\\mathrm { Q C M E ^ { * } { } A }$ outperforms the $\\mathrm { Q C M } { } \\mathrm { A L E }$ baseline by $1 8 . 8 6 \\%$ and $\\mathrm { Q C M L E ^ { * } { } A }$ outperforms $\\mathrm { Q C M } { } \\mathrm { A L E }$ by $1 8 . 9 6 \\%$ , indicating a potential improvement direction by generating correct explanations before answering science questions. ",
|
| 858 |
+
"bbox": [
|
| 859 |
+
174,
|
| 860 |
+
856,
|
| 861 |
+
825,
|
| 862 |
+
911
|
| 863 |
+
],
|
| 864 |
+
"page_idx": 8
|
| 865 |
+
},
|
| 866 |
+
{
|
| 867 |
+
"type": "table",
|
| 868 |
+
"img_path": "images/ce9cc89bc4fa4cf6cdc15b91d1d692813f171a4772a414b2d5a9c2fa216acda9.jpg",
|
| 869 |
+
"table_caption": [
|
| 870 |
+
"Table 6: Upper bound of GPT-3 (CoT). "
|
| 871 |
+
],
|
| 872 |
+
"table_footnote": [],
|
| 873 |
+
"table_body": "<table><tr><td>Prompt type</td><td>Sampling</td><td>Acc.(%)</td></tr><tr><td>QCML*→A</td><td>Random</td><td>73.59</td></tr><tr><td>QCML*→AE</td><td>Random</td><td>74.32</td></tr><tr><td>QCME*→A</td><td>Random</td><td>94.0318.86↑</td></tr><tr><td>QCMLE*→A</td><td>Random</td><td>94.1318.96↑</td></tr><tr><td>QCM→ALE</td><td>Random</td><td>75.17</td></tr></table>",
|
| 874 |
+
"bbox": [
|
| 875 |
+
173,
|
| 876 |
+
88,
|
| 877 |
+
436,
|
| 878 |
+
175
|
| 879 |
+
],
|
| 880 |
+
"page_idx": 9
|
| 881 |
+
},
|
| 882 |
+
{
|
| 883 |
+
"type": "table",
|
| 884 |
+
"img_path": "images/5579132218d25612644b2788242633b1175a3f37ea5beda71faff478b479112b.jpg",
|
| 885 |
+
"table_caption": [],
|
| 886 |
+
"table_footnote": [
|
| 887 |
+
"Table 7: Different positions of L/E for GPT-3 (CoT). "
|
| 888 |
+
],
|
| 889 |
+
"table_body": "<table><tr><td>Prompt type</td><td>Sampling</td><td>Acc.(%)</td></tr><tr><td>QCM→LA</td><td>Random</td><td>60.6</td></tr><tr><td>QCM→EA</td><td>Random</td><td>56.0</td></tr><tr><td>QCM→LEA</td><td>Random</td><td>55.4</td></tr><tr><td>QCM→ELA</td><td>Random</td><td>51.5</td></tr><tr><td>QCM→ALE</td><td>Random</td><td>73.6</td></tr></table>",
|
| 890 |
+
"bbox": [
|
| 891 |
+
506,
|
| 892 |
+
88,
|
| 893 |
+
751,
|
| 894 |
+
176
|
| 895 |
+
],
|
| 896 |
+
"page_idx": 9
|
| 897 |
+
},
|
| 898 |
+
{
|
| 899 |
+
"type": "text",
|
| 900 |
+
"text": "Positions of lectures and explanations. We study the performance of GPT-3 (CoT) in terms of different positions of lectures and explanations on 1,000 test examples. The results are shown in Table 7. There could be huge accuracy decreases if GPT-3 (CoT) predicts lectures and explanations before answers. It is mainly because if GPT-3 (CoT) is formulated to generate the long lecture and explanation first, there is a greater chance that it will stop generating the prediction early or use up the maximum token limits before obtaining the required answer. ",
|
| 901 |
+
"bbox": [
|
| 902 |
+
173,
|
| 903 |
+
218,
|
| 904 |
+
825,
|
| 905 |
+
301
|
| 906 |
+
],
|
| 907 |
+
"page_idx": 9
|
| 908 |
+
},
|
| 909 |
+
{
|
| 910 |
+
"type": "text",
|
| 911 |
+
"text": "CoT learns with fewer data. To study if the chain of thought helps language models learn more efficiently, we report the accuracies of UnifiedQA and UnifiedQA (CoT) fine-tuned on different sizes of the training set in Figure 8. UnifiedQA (CoT) benefits language models by learning the coherent reasoning path when answering questions, resulting in similar accuracy with fewer training examples. ",
|
| 912 |
+
"bbox": [
|
| 913 |
+
174,
|
| 914 |
+
308,
|
| 915 |
+
562,
|
| 916 |
+
405
|
| 917 |
+
],
|
| 918 |
+
"page_idx": 9
|
| 919 |
+
},
|
| 920 |
+
{
|
| 921 |
+
"type": "text",
|
| 922 |
+
"text": "Error analysis. GPT-3 via chain-of-thought prompting obtains promising results but still fails to answer a wide range of challenging questions in SCIENCEQA. See examples of failure cases in Appendix B.4. The failure cases can be classified into two types: (a) the model fails to understand the multimodal inputs and lacks domain-specific knowledge to arrive at the correct answer; (b) the model generates the wrong chain of thought with irrelevant, incorrect, or incomplete information. ",
|
| 923 |
+
"bbox": [
|
| 924 |
+
174,
|
| 925 |
+
411,
|
| 926 |
+
562,
|
| 927 |
+
481
|
| 928 |
+
],
|
| 929 |
+
"page_idx": 9
|
| 930 |
+
},
|
| 931 |
+
{
|
| 932 |
+
"type": "image",
|
| 933 |
+
"img_path": "images/937c2ce21b0f22e8bf103dd8666bbdfba8582936e413aad8da0c9300e4b3adb9.jpg",
|
| 934 |
+
"image_caption": [
|
| 935 |
+
"Figure 8: UnifiedQA (CoT) learns efficiently with fewer training examples. "
|
| 936 |
+
],
|
| 937 |
+
"image_footnote": [],
|
| 938 |
+
"bbox": [
|
| 939 |
+
575,
|
| 940 |
+
309,
|
| 941 |
+
823,
|
| 942 |
+
440
|
| 943 |
+
],
|
| 944 |
+
"page_idx": 9
|
| 945 |
+
},
|
| 946 |
+
{
|
| 947 |
+
"type": "text",
|
| 948 |
+
"text": "",
|
| 949 |
+
"bbox": [
|
| 950 |
+
176,
|
| 951 |
+
481,
|
| 952 |
+
823,
|
| 953 |
+
508
|
| 954 |
+
],
|
| 955 |
+
"page_idx": 9
|
| 956 |
+
},
|
| 957 |
+
{
|
| 958 |
+
"type": "text",
|
| 959 |
+
"text": "6 Discussion and Conclusion ",
|
| 960 |
+
"text_level": 1,
|
| 961 |
+
"bbox": [
|
| 962 |
+
176,
|
| 963 |
+
522,
|
| 964 |
+
428,
|
| 965 |
+
539
|
| 966 |
+
],
|
| 967 |
+
"page_idx": 9
|
| 968 |
+
},
|
| 969 |
+
{
|
| 970 |
+
"type": "text",
|
| 971 |
+
"text": "In this paper, we propose SCIENCEQA, a dataset that features 21,208 multi-option questions with multimodal contexts from the science curriculum. To the best of our knowledge, SCIENCEQA is the first large-scale multimodal science dataset where most questions are annotated with corresponding lectures and explanations. We establish various baselines, including recent VQA models and large language models on SCIENCEQA. We further study if language models can generate reasonable explanations and then benefit the reasoning ability. Experiments show that UnifiedQA with the chain of thought can achieve an improvement of $3 . 9 9 \\%$ and few-shot GPT-3 via chain-of-thought (CoT) prompting can obtain a satisfactory accuracy of $7 5 . 1 7 \\%$ on SCIENCEQA. $6 5 . 2 \\%$ of the generated explanations from GPT-3 (CoT) meet the gold standard by human evaluations. ",
|
| 972 |
+
"bbox": [
|
| 973 |
+
174,
|
| 974 |
+
546,
|
| 975 |
+
825,
|
| 976 |
+
671
|
| 977 |
+
],
|
| 978 |
+
"page_idx": 9
|
| 979 |
+
},
|
| 980 |
+
{
|
| 981 |
+
"type": "text",
|
| 982 |
+
"text": "7 Acknowledgment ",
|
| 983 |
+
"text_level": 1,
|
| 984 |
+
"bbox": [
|
| 985 |
+
176,
|
| 986 |
+
691,
|
| 987 |
+
349,
|
| 988 |
+
709
|
| 989 |
+
],
|
| 990 |
+
"page_idx": 9
|
| 991 |
+
},
|
| 992 |
+
{
|
| 993 |
+
"type": "text",
|
| 994 |
+
"text": "We would like to thank the anonymous reviewers for their valuable comments and suggestions. We would also like to thank Xiaodan Liang for insightful discussions on dataset collection. We thank our colleagues at The Allen Institute of AI (AI2), Jiasen Lu and Jungo Kasai for helpful discussions. The work does not relate to Liang Qiu’s position at Amazon Alexa. ",
|
| 995 |
+
"bbox": [
|
| 996 |
+
174,
|
| 997 |
+
724,
|
| 998 |
+
825,
|
| 999 |
+
780
|
| 1000 |
+
],
|
| 1001 |
+
"page_idx": 9
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"type": "text",
|
| 1005 |
+
"text": "References ",
|
| 1006 |
+
"text_level": 1,
|
| 1007 |
+
"bbox": [
|
| 1008 |
+
174,
|
| 1009 |
+
799,
|
| 1010 |
+
266,
|
| 1011 |
+
815
|
| 1012 |
+
],
|
| 1013 |
+
"page_idx": 9
|
| 1014 |
+
},
|
| 1015 |
+
{
|
| 1016 |
+
"type": "text",
|
| 1017 |
+
"text": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. \n[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision (CVPR), pages 2425–2433, 2015. ",
|
| 1018 |
+
"bbox": [
|
| 1019 |
+
178,
|
| 1020 |
+
823,
|
| 1021 |
+
825,
|
| 1022 |
+
911
|
| 1023 |
+
],
|
| 1024 |
+
"page_idx": 9
|
| 1025 |
+
},
|
| 1026 |
+
{
|
| 1027 |
+
"type": "text",
|
| 1028 |
+
"text": "[3] Jonathan Bragg, Arman Cohan, Kyle Lo, and Iz Beltagy. Flex: Unifying evaluation for few-shot nlp. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. ",
|
| 1029 |
+
"bbox": [
|
| 1030 |
+
176,
|
| 1031 |
+
92,
|
| 1032 |
+
825,
|
| 1033 |
+
118
|
| 1034 |
+
],
|
| 1035 |
+
"page_idx": 10
|
| 1036 |
+
},
|
| 1037 |
+
{
|
| 1038 |
+
"type": "text",
|
| 1039 |
+
"text": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877–1901, 2020. ",
|
| 1040 |
+
"bbox": [
|
| 1041 |
+
178,
|
| 1042 |
+
126,
|
| 1043 |
+
823,
|
| 1044 |
+
166
|
| 1045 |
+
],
|
| 1046 |
+
"page_idx": 10
|
| 1047 |
+
},
|
| 1048 |
+
{
|
| 1049 |
+
"type": "text",
|
| 1050 |
+
"text": "[5] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. ",
|
| 1051 |
+
"bbox": [
|
| 1052 |
+
178,
|
| 1053 |
+
174,
|
| 1054 |
+
823,
|
| 1055 |
+
213
|
| 1056 |
+
],
|
| 1057 |
+
"page_idx": 10
|
| 1058 |
+
},
|
| 1059 |
+
{
|
| 1060 |
+
"type": "text",
|
| 1061 |
+
"text": "[6] Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. Explaining answers with entailment trees. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. ",
|
| 1062 |
+
"bbox": [
|
| 1063 |
+
176,
|
| 1064 |
+
220,
|
| 1065 |
+
823,
|
| 1066 |
+
261
|
| 1067 |
+
],
|
| 1068 |
+
"page_idx": 10
|
| 1069 |
+
},
|
| 1070 |
+
{
|
| 1071 |
+
"type": "text",
|
| 1072 |
+
"text": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations (ICLR), 2021. ",
|
| 1073 |
+
"bbox": [
|
| 1074 |
+
178,
|
| 1075 |
+
267,
|
| 1076 |
+
823,
|
| 1077 |
+
319
|
| 1078 |
+
],
|
| 1079 |
+
"page_idx": 10
|
| 1080 |
+
},
|
| 1081 |
+
{
|
| 1082 |
+
"type": "text",
|
| 1083 |
+
"text": "[8] Feng Gao, Qing Ping, Govind Thattai, Aishwarya Reganti, Ying Nian Wu, and Prem Natarajan. Transformretrieve-generate: Natural language-centric outside-knowledge visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5067–5077, June 2022. ",
|
| 1084 |
+
"bbox": [
|
| 1085 |
+
179,
|
| 1086 |
+
328,
|
| 1087 |
+
826,
|
| 1088 |
+
380
|
| 1089 |
+
],
|
| 1090 |
+
"page_idx": 10
|
| 1091 |
+
},
|
| 1092 |
+
{
|
| 1093 |
+
"type": "text",
|
| 1094 |
+
"text": "[9] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6639–6648, 2019. ",
|
| 1095 |
+
"bbox": [
|
| 1096 |
+
178,
|
| 1097 |
+
387,
|
| 1098 |
+
825,
|
| 1099 |
+
428
|
| 1100 |
+
],
|
| 1101 |
+
"page_idx": 10
|
| 1102 |
+
},
|
| 1103 |
+
{
|
| 1104 |
+
"type": "text",
|
| 1105 |
+
"text": "[10] Peng Gao, Hongsheng Li, Shuang Li, Pan Lu, Yikang Li, Steven CH Hoi, and Xiaogang Wang. Questionguided hybrid convolution for visual question answering. In The European Conference on Computer Vision (ECCV), pages 469–485, 2018. ",
|
| 1106 |
+
"bbox": [
|
| 1107 |
+
174,
|
| 1108 |
+
434,
|
| 1109 |
+
823,
|
| 1110 |
+
474
|
| 1111 |
+
],
|
| 1112 |
+
"page_idx": 10
|
| 1113 |
+
},
|
| 1114 |
+
{
|
| 1115 |
+
"type": "text",
|
| 1116 |
+
"text": "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. ",
|
| 1117 |
+
"bbox": [
|
| 1118 |
+
173,
|
| 1119 |
+
482,
|
| 1120 |
+
823,
|
| 1121 |
+
522
|
| 1122 |
+
],
|
| 1123 |
+
"page_idx": 10
|
| 1124 |
+
},
|
| 1125 |
+
{
|
| 1126 |
+
"type": "text",
|
| 1127 |
+
"text": "[12] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6700–6709, 2019. ",
|
| 1128 |
+
"bbox": [
|
| 1129 |
+
176,
|
| 1130 |
+
529,
|
| 1131 |
+
821,
|
| 1132 |
+
568
|
| 1133 |
+
],
|
| 1134 |
+
"page_idx": 10
|
| 1135 |
+
},
|
| 1136 |
+
{
|
| 1137 |
+
"type": "text",
|
| 1138 |
+
"text": "[13] Peter A Jansen, Elizabeth Wainwright, Steven Marmorstein, and Clayton T Morrison. Worldtree: A corpus of explanation graphs for elementary science questions supporting multi-hop inference. arXiv preprint arXiv:1802.03052, 2018. ",
|
| 1139 |
+
"bbox": [
|
| 1140 |
+
173,
|
| 1141 |
+
577,
|
| 1142 |
+
823,
|
| 1143 |
+
616
|
| 1144 |
+
],
|
| 1145 |
+
"page_idx": 10
|
| 1146 |
+
},
|
| 1147 |
+
{
|
| 1148 |
+
"type": "text",
|
| 1149 |
+
"text": "[14] Harsh Jhamtani and Peter Clark. Learning to explain: Datasets and models for identifying valid reasoning chains in multihop question-answering. arXiv preprint arXiv:2010.03274, 2020. ",
|
| 1150 |
+
"bbox": [
|
| 1151 |
+
171,
|
| 1152 |
+
623,
|
| 1153 |
+
821,
|
| 1154 |
+
650
|
| 1155 |
+
],
|
| 1156 |
+
"page_idx": 10
|
| 1157 |
+
},
|
| 1158 |
+
{
|
| 1159 |
+
"type": "text",
|
| 1160 |
+
"text": "[15] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2901–2910, 2017. ",
|
| 1161 |
+
"bbox": [
|
| 1162 |
+
173,
|
| 1163 |
+
659,
|
| 1164 |
+
823,
|
| 1165 |
+
709
|
| 1166 |
+
],
|
| 1167 |
+
"page_idx": 10
|
| 1168 |
+
},
|
| 1169 |
+
{
|
| 1170 |
+
"type": "text",
|
| 1171 |
+
"text": "[16] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5648–5656, 2018. ",
|
| 1172 |
+
"bbox": [
|
| 1173 |
+
171,
|
| 1174 |
+
718,
|
| 1175 |
+
823,
|
| 1176 |
+
757
|
| 1177 |
+
],
|
| 1178 |
+
"page_idx": 10
|
| 1179 |
+
},
|
| 1180 |
+
{
|
| 1181 |
+
"type": "text",
|
| 1182 |
+
"text": "[17] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Min Joon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. ",
|
| 1183 |
+
"bbox": [
|
| 1184 |
+
173,
|
| 1185 |
+
765,
|
| 1186 |
+
823,
|
| 1187 |
+
804
|
| 1188 |
+
],
|
| 1189 |
+
"page_idx": 10
|
| 1190 |
+
},
|
| 1191 |
+
{
|
| 1192 |
+
"type": "text",
|
| 1193 |
+
"text": "[18] Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4999–5007, 2017. ",
|
| 1194 |
+
"bbox": [
|
| 1195 |
+
171,
|
| 1196 |
+
813,
|
| 1197 |
+
825,
|
| 1198 |
+
864
|
| 1199 |
+
],
|
| 1200 |
+
"page_idx": 10
|
| 1201 |
+
},
|
| 1202 |
+
{
|
| 1203 |
+
"type": "text",
|
| 1204 |
+
"text": "[19] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings of the Association for Computational Linguistics (EMNLP), pages 1896–1907, 2020. ",
|
| 1205 |
+
"bbox": [
|
| 1206 |
+
171,
|
| 1207 |
+
872,
|
| 1208 |
+
826,
|
| 1209 |
+
911
|
| 1210 |
+
],
|
| 1211 |
+
"page_idx": 10
|
| 1212 |
+
},
|
| 1213 |
+
{
|
| 1214 |
+
"type": "text",
|
| 1215 |
+
"text": "[20] Tushar Khot, Peter Clark, Michal Guerquin, Peter Alexander Jansen, and Ashish Sabharwal. Qasc: A dataset for question answering via sentence composition. ArXiv, abs/1910.11473, 2020. \n[21] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 1571–1581, 2018. \n[22] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 5583–5594, 2021. \n[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision (IJCV), pages 32–73, 2017. \n[24] Jayant Krishnamurthy, Oyvind Tafjord, and Aniruddha Kembhavi. Semantic parsing to probabilistic programs for situated question answering. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 160–170, 2016. \n[25] Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329, 2022. \n[26] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. \n[27] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5265–5275, 2020. \n[28] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81, 2004. \n[29] Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? arXiv preprint arXiv:2101.06804, 2021. \n[30] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems (NeurIPS), pages 13–23, 2019. \n[31] Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. Intergps: Interpretable geometry problem solving with formal language and symbolic reasoning. In The 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. \n[32] Pan Lu, Lei Ji, Wei Zhang, Nan Duan, Ming Zhou, and Jianyong Wang. R-vqa: learning visual relation facts with semantic attention for visual question answering. In The ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (SIGKDD), pages 1880–1889, 2018. \n[33] Pan Lu, Hongsheng Li, Wei Zhang, Jianyong Wang, and Xiaogang Wang. Co-attending free-form regions and detections with multi-modal multiplicative feature embedding for visual question answering. In The AAAI Conference on Artificial Intelligence (AAAI), 2018. \n[34] Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022. \n[35] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. \n[36] Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786, 2021. \n[37] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2018. \n[38] Swaroop Mishra, Matthew Finlayson, Pan Lu, Leonard Tang, Sean Welleck, Chitta Baral, Tanmay Rajpurohit, Oyvind Tafjord, Ashish Sabharwal, Peter Clark, and Ashwin Kalyan. Lila: A unified benchmark for mathematical reasoning. In The 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2022. \n[39] Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Reframing instructional prompts to gptk’s language. ACL Findings, 2021. \n[40] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. The 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. \n[41] Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan. Wt5?! training text-to-text models to explain their predictions. arXiv preprint arXiv:2004.14546, 2020. \n[42] Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models. arXiv preprint arXiv:2112.00114, 2021. \n[43] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155, 2022. \n[44] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (ACL), pages 311–318, 2002. \n[45] Mihir Parmar, Swaroop Mishra, Mirali Purohit, Man Luo, Murad Mohammad, and Chitta Baral. InBoXBART: Get instructions into biomedical multi-task learning. In Findings of the Association for Computational Linguistics: NAACL 2022, pages 112–128, Seattle, United States, July 2022. Association for Computational Linguistics. \n[46] Ethan Perez, Douwe Kiela, and Kyunghyun Cho. True few-shot learning with language models. Advances in Neural Information Processing Systems (NeurIPS), 34, 2021. \n[47] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. \n[48] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21:1–67, 2020. \n[49] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 11 2019. \n[50] Mrinmaya Sachan, Kumar Dubey, and Eric Xing. From textbooks to knowledge: A case study in harvesting axiomatic knowledge from textbooks to solve geometry problems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 773–784, 2017. \n[51] Mrinmaya Sachan and Eric Xing. Learning to solve geometry problems from natural language demonstrations in textbooks. In Proceedings of the 6th Joint Conference on Lexical and Computational Semantics (\\* SEM 2017), pages 251–261, 2017. \n[52] Shailaja Keyur Sampat, Yezhou Yang, and Chitta Baral. Visuo-lingustic question answering (vlqa) challenge. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings (EMNLP), pages 4606–4616, 2020. \n[53] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. The International Conference on Learning Representations (ICLR), 2021. \n[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022. \n[55] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6281–6290, 2019. \n[56] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and Yang: Balancing and answering binary visual questions. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. \n[57] Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning (ICML), pages 12697–12706. PMLR, 2021. \n[58] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. ",
|
| 1216 |
+
"bbox": [
|
| 1217 |
+
171,
|
| 1218 |
+
50,
|
| 1219 |
+
828,
|
| 1220 |
+
919
|
| 1221 |
+
],
|
| 1222 |
+
"page_idx": 11
|
| 1223 |
+
},
|
| 1224 |
+
{
|
| 1225 |
+
"type": "text",
|
| 1226 |
+
"text": "",
|
| 1227 |
+
"bbox": [
|
| 1228 |
+
169,
|
| 1229 |
+
54,
|
| 1230 |
+
828,
|
| 1231 |
+
916
|
| 1232 |
+
],
|
| 1233 |
+
"page_idx": 12
|
| 1234 |
+
},
|
| 1235 |
+
{
|
| 1236 |
+
"type": "text",
|
| 1237 |
+
"text": "",
|
| 1238 |
+
"bbox": [
|
| 1239 |
+
173,
|
| 1240 |
+
92,
|
| 1241 |
+
826,
|
| 1242 |
+
213
|
| 1243 |
+
],
|
| 1244 |
+
"page_idx": 13
|
| 1245 |
+
},
|
| 1246 |
+
{
|
| 1247 |
+
"type": "text",
|
| 1248 |
+
"text": "Checklist ",
|
| 1249 |
+
"text_level": 1,
|
| 1250 |
+
"bbox": [
|
| 1251 |
+
174,
|
| 1252 |
+
238,
|
| 1253 |
+
254,
|
| 1254 |
+
253
|
| 1255 |
+
],
|
| 1256 |
+
"page_idx": 13
|
| 1257 |
+
},
|
| 1258 |
+
{
|
| 1259 |
+
"type": "text",
|
| 1260 |
+
"text": "1. For all authors... ",
|
| 1261 |
+
"bbox": [
|
| 1262 |
+
214,
|
| 1263 |
+
263,
|
| 1264 |
+
339,
|
| 1265 |
+
279
|
| 1266 |
+
],
|
| 1267 |
+
"page_idx": 13
|
| 1268 |
+
},
|
| 1269 |
+
{
|
| 1270 |
+
"type": "text",
|
| 1271 |
+
"text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] Yes, we did the error analysis in Section 5.4 and discussed the limitations of the work in Appendix B.4. \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] We discussed the broader impacts in Appendix B.5. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
|
| 1272 |
+
"bbox": [
|
| 1273 |
+
238,
|
| 1274 |
+
284,
|
| 1275 |
+
825,
|
| 1276 |
+
401
|
| 1277 |
+
],
|
| 1278 |
+
"page_idx": 13
|
| 1279 |
+
},
|
| 1280 |
+
{
|
| 1281 |
+
"type": "text",
|
| 1282 |
+
"text": "2. If you are including theoretical results... ",
|
| 1283 |
+
"bbox": [
|
| 1284 |
+
214,
|
| 1285 |
+
405,
|
| 1286 |
+
493,
|
| 1287 |
+
420
|
| 1288 |
+
],
|
| 1289 |
+
"page_idx": 13
|
| 1290 |
+
},
|
| 1291 |
+
{
|
| 1292 |
+
"type": "text",
|
| 1293 |
+
"text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ",
|
| 1294 |
+
"bbox": [
|
| 1295 |
+
238,
|
| 1296 |
+
422,
|
| 1297 |
+
738,
|
| 1298 |
+
454
|
| 1299 |
+
],
|
| 1300 |
+
"page_idx": 13
|
| 1301 |
+
},
|
| 1302 |
+
{
|
| 1303 |
+
"type": "text",
|
| 1304 |
+
"text": "3. If you ran experiments... ",
|
| 1305 |
+
"bbox": [
|
| 1306 |
+
214,
|
| 1307 |
+
458,
|
| 1308 |
+
393,
|
| 1309 |
+
472
|
| 1310 |
+
],
|
| 1311 |
+
"page_idx": 13
|
| 1312 |
+
},
|
| 1313 |
+
{
|
| 1314 |
+
"type": "text",
|
| 1315 |
+
"text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We included 100 data examples and the data visualizer tool in the supplemental material. The whole dataset and code will be available at https://scienceqa.github.io. \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section 5.1 and Appendix B.1 for experimental details. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] We reported the error bars for GPT-3 (CoT) experiments in Figure 7, where each experiment was repeated four times. \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] We discussed compute resources in Appendix B.1. ",
|
| 1316 |
+
"bbox": [
|
| 1317 |
+
238,
|
| 1318 |
+
477,
|
| 1319 |
+
825,
|
| 1320 |
+
648
|
| 1321 |
+
],
|
| 1322 |
+
"page_idx": 13
|
| 1323 |
+
},
|
| 1324 |
+
{
|
| 1325 |
+
"type": "text",
|
| 1326 |
+
"text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
|
| 1327 |
+
"bbox": [
|
| 1328 |
+
217,
|
| 1329 |
+
654,
|
| 1330 |
+
823,
|
| 1331 |
+
667
|
| 1332 |
+
],
|
| 1333 |
+
"page_idx": 13
|
| 1334 |
+
},
|
| 1335 |
+
{
|
| 1336 |
+
"type": "text",
|
| 1337 |
+
"text": "(a) If your work uses existing assets, did you cite the creators? [Yes] We collected the SCIENCEQA dataset from https://www.ixl.com/. The copyright belongs to IXL. \n(b) Did you mention the license of the assets? [Yes] SCIENCEQA is under the CC BY-NCSA 4.0 license and is used for non-commercial research purposes. \n(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We included data examples and a visualizer tool in the supplemental material. The dataset will be available at https://scienceqa.github.io. \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] The collected data does not contain personally identifiable information or offensive content. ",
|
| 1338 |
+
"bbox": [
|
| 1339 |
+
238,
|
| 1340 |
+
671,
|
| 1341 |
+
825,
|
| 1342 |
+
847
|
| 1343 |
+
],
|
| 1344 |
+
"page_idx": 13
|
| 1345 |
+
},
|
| 1346 |
+
{
|
| 1347 |
+
"type": "text",
|
| 1348 |
+
"text": "5. If you used crowdsourcing or conducted research with human subjects... ",
|
| 1349 |
+
"bbox": [
|
| 1350 |
+
214,
|
| 1351 |
+
852,
|
| 1352 |
+
704,
|
| 1353 |
+
866
|
| 1354 |
+
],
|
| 1355 |
+
"page_idx": 13
|
| 1356 |
+
},
|
| 1357 |
+
{
|
| 1358 |
+
"type": "text",
|
| 1359 |
+
"text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes] We included screenshots of the instructions in Appendix B.2 and B.3. ",
|
| 1360 |
+
"bbox": [
|
| 1361 |
+
240,
|
| 1362 |
+
869,
|
| 1363 |
+
823,
|
| 1364 |
+
910
|
| 1365 |
+
],
|
| 1366 |
+
"page_idx": 13
|
| 1367 |
+
},
|
| 1368 |
+
{
|
| 1369 |
+
"type": "text",
|
| 1370 |
+
"text": "(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [Yes] We included the monetary compensation details in Appendix B.2 and B.3. ",
|
| 1371 |
+
"bbox": [
|
| 1372 |
+
238,
|
| 1373 |
+
92,
|
| 1374 |
+
825,
|
| 1375 |
+
164
|
| 1376 |
+
],
|
| 1377 |
+
"page_idx": 14
|
| 1378 |
+
}
|
| 1379 |
+
]
|
parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/HjwK-Tc_Bc/HjwK-Tc_Bc_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/IJNDyqdRF0m/IJNDyqdRF0m.md
ADDED
|
@@ -0,0 +1,429 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Decomposing NeRF for Editing via Feature Field Distillation
|
| 2 |
+
|
| 3 |
+
Sosuke Kobayashi Preferred Networks, Inc. sosk@preferred.jp
|
| 4 |
+
|
| 5 |
+
Eiichi Matsumoto Preferred Networks, Inc. matsumoto@preferred.jp
|
| 6 |
+
|
| 7 |
+
Vincent Sitzmann Massachusetts Institute of Technology sitzmann@mit.edu
|
| 8 |
+
|
| 9 |
+
pfnet-research.github.io/distilled-feature-fields/
|
| 10 |
+
|
| 11 |
+
# Abstract
|
| 12 |
+
|
| 13 |
+
Emerging neural radiance fields (NeRF) are a promising scene representation for computer graphics, enabling high-quality 3D reconstruction and novel view synthesis from image observations. However, editing a scene represented by a NeRF is challenging, as the underlying connectionist representations such as MLPs or voxel grids are not object-centric or compositional. In particular, it has been difficult to selectively edit specific regions or objects. In this work, we tackle the problem of semantic scene decomposition of NeRFs to enable query-based local editing of the represented 3D scenes. We propose to distill the knowledge of off-the-shelf, supervised and self-supervised 2D image feature extractors such as CLIP-LSeg or DINO into a 3D feature field optimized in parallel to the radiance field. Given a user-specified query of various modalities such as text, an image patch, or a point-and-click selection, 3D feature fields semantically decompose 3D space without the need for re-training and enable us to semantically select and edit regions in the radiance field. Our experiments validate that the distilled feature fields can transfer recent progress in 2D vision and language foundation models to 3D scene representations, enabling convincing 3D segmentation and selective editing of emerging neural graphics representations.
|
| 14 |
+
|
| 15 |
+
# 1 Introduction
|
| 16 |
+
|
| 17 |
+
Emerging neural implicit representations or neural fields have been shown to be a promising approach for representing a variety of signals [82, 53, 65, 106, 56]. In particular, they play an important role in 3D scene reconstruction and novel view synthesis from a limited number of context images. Neural radiance fields (NeRF) [56] enabled the recovery of a continuous volume density and radiance field from a limited number of observations, producing high-quality images from arbitrary views via volume rendering with promising applications in computer graphics. However, editing a scene reconstructed by NeRF is non-obvious because the scene is not object-centric and is implicitly encoded in the weights of a connectionist representation such as an MLP [56] or a voxelgrid [23]. Although we can transform the scene in input or output space or via optimization-based editing [37, 97], this does not enable selective object-centric or semantic, local edits, such as moving a single object. Prior work has addressed this challenge via coordinate-level, semantic decompositions which allow to selectively move, deform, paint, or optimize parts of a NeRF, but relies on costly annotation of instance segmentations and training of instance-specific networks [104]. While this can be alleviated with pre-trained segmentation models [25, 41], such models require pre-defined closed label sets and domains (e.g., traffic scenes), limiting decomposition and editing. Local editing of NeRFs ideally requires an efficient, open-set method for coordinate-level decomposition.
|
| 18 |
+
|
| 19 |
+
In this work, we present distilled feature fields (DFFs), a novel approach to query-based scene decomposition for local, interactive editing of NeRFs. We focus on 3D neural feature fields, which map every 3D coordinate to a semantic feature descriptor of that coordinate. Conditioned on a user query such as a text or image patch, this 3D feature field can compute a decomposition of a scene without re-training. We train a scene-specific DFF via teacher-student distillation [34], using supervision from feature encoders pre-trained on the image domain. Unlike the domain of 3D scenes, the image domain boasts massive high-quality datasets and abundant prior work on self-supervised and supervised training of effective feature extraction models. Notably, recently proposed transformer-based models [96, 22] have demonstrated impressive capabilities across various vision- and text-based tasks (e.g., CLIP [69], LSeg [44], DINO [12]). Such feature spaces capture the semantic properties of regions and make it possible to correspond and segment them well by text, image queries, or clustering. We employ these models as teacher networks and distill them into 3D feature fields via volume rendering. The trained feature field enables us to semantically select and edit specific regions in 3D NeRF scenes and render multi-view consistent images from the locally edited scenes.
|
| 20 |
+
|
| 21 |
+
In extensive experiments, we investigate the applications of neural feature fields with two different pre-trained teacher networks, (1) LSeg [44], a CLIP-inspired language-driven semantic segmentation network, and (2) DINO [12, 3], a self-supervised network aware of various object boundaries and correspondences. LSeg and DINO features allow us to select 3D regions by a simple text query or an image patch, respectively. We first quantitatively demonstrate that LSeg-based DFFs with label queries can have high 3D segmentation performance compared with an existing point-cloud based 3D segmentation baseline trained on ScanNet [20], a supervised point-cloud dataset. We then demonstrate a variety of 3D appearance and geometry edits across real-world NeRF scenes with no annotations of segmentation; and show that we may edit regions with a single query of text, image, pixel, or cluster choice.
|
| 22 |
+
|
| 23 |
+
# 2 Related Work
|
| 24 |
+
|
| 25 |
+
Neural Implicit Representations. Neural implicit representations or neural fields have recently advanced neural processing for 3D data and multi-view 2D images [82, 53, 65, 106, 56]. For a review of this emerging space we point the reader to the reports by Kato et al. [39], Tewari et al. [90], and Xie et al. [102]. In particular, a neural radiance field (NeRF) can be fitted to a set of posed 2D images and maps a 3D point coordinate and a view direction to RGB color and density. When observations are limited, NeRF often overfits and fails to synthesize novel views with correct geometry and appearance. Pre-trained vision models have been used for regularizing NeRF via flows [62], multi-view consistency [35], perceptual loss [110], or depth estimation [100, 77]. Some pre-trained models operate not only in the visual world but also in other modalities such as language. The recently proposed CLIP model [69] has demonstrated impressive performance in image-and-text alignment, with strong generalization to various textual and visual concepts. Wang et al. [97], Jain et al. [36], and Poole et al. [68] use CLIP or Imagen [79] to edit or generate a single-object NeRF with a text prompt query by optimizing the NeRF parameters to generate images matched with the text. While such methods are promising, they do not enable accurate selective editing of specific scene regions. For example, the prompt “yellow flowers” may affect unintended scene regions, such as the leaves of a plant. Our proposed decomposition method leverages pre-trained foundation models to enable selective editing of real-world NeRF scenes. Neural descriptor fields [80] use intermediate features that emerge in a 3D occupancy field network [53] for efficiently teaching robots object grasping. Instead of a pre-trained object-centric 3D model, we use 2D vision models as teacher networks via distillation, exploiting recent progress in pre-trained foundation models [7].
|
| 26 |
+
|
| 27 |
+
Geometric Decomposition of Neural Scene Representations Kohli et al. [40] and Zhi et al. [112] show that neural implicit representations can be combined with the supervision of semantic labels. Yang et al. [104] demonstrate that given view-consistent ground-truth instance segmentation masks during training, NeRF can be trained to represent each object as different volumes, although such an annotation is expensive in practice. Concurrently, Benaim et al. [6] also experiment with the different parametarization. Conditional [49, 37, 21, 63] and generative models [60, 61, 31] enable a degree of category-specific decomposition (e.g., human bodyparts) and editing on constrained domains with large datasets. Regular structures such as voxelgrids or octrees [13, 48, 14, 88, 94, 89, 43, 81, 107, 59, 60] or unsupervised decomposition [73, 85, 109, 83] enable editability via manipulation of localized parameters. However, the decomposition is limited due to the inflexibly structured boundaries or strong assumptions about scenes; self-supervised object-centric learning is a difficult task. Other studies also explored reconstruction with more structured hybrid representations via pipelines specialized to a domain (e.g., traffic scene) [64, 25, 41] or situation (e.g., each object data is independently accessible) [28, 27, 105]. Note that this line of work defines and constrains domains or the types of segmentation during or before training and thus limits the degrees of freedom for editable scenes and objects. In contrast, our method can decompose scene-specific NeRFs into arbitrary semantic units via text and image queries, enabling versatile scene edits without retraining. A concurrent paper by Tschernezki et al. [93] also explores the same training framework and, in particular, investigates how fused features are improved from 2D teacher networks. It also complementarily shows the results with other teacher models (MoCo-v3 [17] and DeiT [91]), dimension reduction via PCA, and NeuralDiff [92]-based neural fields. Other concurrent studies explore decomposition through training scene-specific segmentation field [113] or 3DCNN [76] supervised by click or scribble annotations. Lastly, in a different but related task, video editing, Kasten et al. [38] use foreground-background decomposition and atlas representation for time-consistent, local editing; Loeschcke et al. [51] and Bar-Tal et al. [5] further use CLIP for editing.
|
| 28 |
+
|
| 29 |
+
Zero-shot Semantic Segmentation. Zero-shot semantic segmentation is a challenging task [24, 2, 10] where a model has to predict semantic labels of pixels in images without a-priori information of the categories. A typical solution is to use vision-and-language cross-modal encoders. They are trained to encode images (pixels) and text labels into the same semantic space and perform zero-shot prediction based on the similarity or alignments of the two inputs. Recent development of image encoder architectures [96, 22, 71] and large-scale training [69, 12] have improved the ability and generalization of vision models, including zero-shot models [44, 52, 99, 103, 114, 72]. On the other hand, ongoing studies on zero-shot perception in 3D still suffer from the lack of effective, efficient, and high-resolution architectures and large-scale annotated datasets [54, 33, 29, 101, 78, 30]. Our method is a new approach to perform zero-shot semantic segmentation on scene-specific 3D fields by exploiting progress in the image domain without semantic 3D supervision. We note that the goal of this paper is not to achieve state-of-the-art performance on 3D semantic segmentation tasks. Instead, our goal is the decomposition of neural scene representations for editing, which requires smooth segmentation results on continuous 3D space rather than segmentation of discrete point clouds or voxelgrids.
|
| 30 |
+
|
| 31 |
+
# 3 Preliminaries
|
| 32 |
+
|
| 33 |
+
# 3.1 Neural Radiance Fields (NeRF)
|
| 34 |
+
|
| 35 |
+
NeRF [56] uses MLPs to output density $\sigma$ and color c given a point coordinate $\mathbf { x } = ( x , y , z )$ in a 3D scene. This simple scene representation can be rendered and optimized via volume rendering. Given a pixel’s camera ray $\mathbf { r } ( t ) = \mathbf { o } + t \mathbf { d }$ , depth $t$ with bounds $[ t _ { \mathrm { n e a r } } , t _ { \mathrm { f a r } } ]$ , camera position $\mathbf { o }$ , and its view direction $\mathbf { d }$ , NeRF calculates the color of a ray using quadrature of $K$ sampled points $\{ \mathbf { x } _ { k } \} _ { k = 1 } ^ { K }$ with depths $\{ t _ { k } \} _ { k = 1 } ^ { K }$ as
|
| 36 |
+
|
| 37 |
+
$$
|
| 38 |
+
\hat { \mathbf { C } } ( \mathbf { r } ) = \sum _ { k = 1 } ^ { K } \hat { T } ( t _ { k } ) \alpha \left( \sigma ( \mathbf { x } _ { k } ) \delta _ { k } \right) \mathbf { c } ( \mathbf { x } _ { k } , \mathbf { d } ) , \quad \hat { T } ( t _ { k } ) = \exp \left( - \sum _ { k ^ { \prime } = 1 } ^ { k - 1 } \sigma ( \mathbf { x } _ { k ^ { \prime } } ) \delta _ { k ^ { \prime } } \right) ,
|
| 39 |
+
$$
|
| 40 |
+
|
| 41 |
+
where $\alpha \left( x \right) = 1 - \exp ( - x )$ , and $\delta _ { k } = t _ { k + 1 } - t _ { k }$ is the distance between adjacent point samples. NeRFs are optimized solely on a dataset of images and their camera poses by minimizing a rerendering loss.
|
| 42 |
+
|
| 43 |
+
# 3.2 Pre-trained Models and Zero-shot Segmentation of Images
|
| 44 |
+
|
| 45 |
+
Most semantic segmentation models pre-define a closed set of labels, and cannot flexibly change the segmentation categories or boundaries without supervised training. In contrast, zero-shot semantic segmentation predicts target regions given open-set queries. Li et al. [44] proposes LSeg, a model to feature field f and the pretrained text encoder f . Specifically, probability of a label l of a point x in 2 2Rperform zero-shot semantic segmentation by aligning pixel-level features and a text query feature. 182 the 3D space, p(l x), are predicted by dot product of the 3D feature f (x) and text label feature fq(l) = Lp + Lf , Lp = C (r) C(r) 2 , Lf = F(r) fimg(I, r) 1 , (4)4LSeg employs an image feature encoder with the DPT architecture [71] and a CLIP-based text label 183 followed by softmax: 5stency. In addition, importantly for user-friendly interactive editing, w52R 2Rfeature encoder [69], trained via large-scale language-image contrastive learning. The probability of a text label $l$ 4given a pixel $r$ in an image $I$ , $\mathbf { p } ( l | I , r )$ T 4, is then calculated via dot product of pixel-level image feature ${ \bf f } _ { \mathrm { i m g } } ( I , r )$ p(l|x) = Pand queried text feature ${ \bf f } _ { \mathrm { q } } ( l )$ q T . followed by a softmax:
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
as distillation from 2D teacher network to 3D student 4122 k k+1 k ˆ 214 breaks 3D consistency. In addition, importantly f214 breaks 3D consistency. In addition, importantly for usemize f through SGD on minimizing the difference between rendered features F (r) andLp = X Cˆ (r) C(r) , Lf = X Fˆ(r) fimg(I, r) , (4)iginal NeRF [46] for the training objective and the volume rendering strategy. Inps://huggingface.co/sentence-transformers/clip-ViT-B-32-multilingual-v1Figure 1: Left: A Distilled Feature Field (DFF) maps a coordinate x and a viewing direction d to ll this mode 124 work via ther2Rs Fˆ (r) anddensity $\sigma$ eural Perceptual Fields (NePeRF). els and ground-truth pixels of real images. nconsistent supervision with noise could harm reconstruction quality of geometry, although the lume rendering trick. We call this model distilled feature field (DFF).r2Re teacher’s outputs f (I, r). For volume rendering, we use two 4184 , color c, and feature f . It is trained by minimizing the difference between rendered features sformers/clip-ViT-B-32-multilingual-v1178 effect seems negligible in preliminary experiments. d at any 3D point without limiting resolution, so naturally used tog41 It is an interesting direction to introduce view de1 It is an interesting direction to introduce view dependehe original NeRF [46] for the training objective and the volume rendering strategy. Inume rendering with coarse-and-fine hierarchical sampling as well as the original185 and feature loss Lf , in total, L:and features as predicted by a pre-trained image feature encoder, as well as the rendered color and 3t is an interesting direction to introduce view dependency to the segmentation for discriminating view-dependent query like referring expressions (e.g., “the chdependent query like referring expressions (e.g., “the chair lef the photometric loss, we add a new objective for minimizing the difference betweenˆ 4 pLf , in total, L: 2ground-truth pixel color. Right: At test time, we may decompose and edit 3D space via selecting and 4 179 4.2 Query-based Decomposition and Editingdent query like referring expressions (e.g., “the chair left to the table” 2020, Liu et img or volume rendering with coarse-and-fine hierarchical sampling as well as the originalX ˆ 2 X ˆ L = Lp + Lf , Lp manipulating different 3D regions with a variety of queries.
|
| 49 |
+
|
| 50 |
+
$$
|
| 51 |
+
{ \bf p } ( l | I , r ) = \frac { \exp ( { \bf f } _ { \mathrm { i m g } } ( I , r ) { \bf f } _ { \mathrm { q } } ( l ) ^ { \mathrm { T } } ) } { \sum _ { l ^ { \prime } \in \mathcal { L } } \exp ( { \bf f } _ { \mathrm { i m g } } ( I , r ) { \bf f } _ { \mathrm { q } } ( l ^ { \prime } ) ^ { \mathrm { T } } ) } ,
|
| 52 |
+
$$
|
| 53 |
+
|
| 54 |
+
186 where $\mathcal { L }$ is a set of possible labels. If negative labels are not available, we may use other scores like and the query2 188 , so it and view synthesis with it are 3D consistent as well as the original NeRF. Unlikethresholded cosine similarity to directly compute the probability of a label. During training, LSeg 189 the proposed method, editing optimizes only the image encoder ${ \bf f } _ { \mathrm { i m g } } \bar { ( } I , r )$ nthesized images by image-based postprocessing breaksby minimizing cross-entropy on supervised semantic 190 3D consistency. In addition, we can chsegmentation datasets. The text encoder ${ \bf f } _ { \mathrm { q } } ( l )$ he segmentation by changing only the query withoutis obtained from a pre-trained CLIP model [69]. 191 retraining, which cannot be realized by existing methods using closed-set semantic segmentation [ZhiRecently, pre-trained CLIP has been leveraged as the backbone for a variety of tasks and has been 192 et al., 2021a] or instance segmentation annotation [Yang et al., 2021], but important for user-friendlyextended with additional modules sharing the same latent space. For example, Reimers and Gurevych [74, 75] trains a multi-lingual (more than $5 0 +$ languages) text encoder, which enables CLIP and 2 It is an interesting direction to introduce view dependency to the segmentation for discriminating view-CLIP-inspired variants to use non-English queries like Japanese. We similarly use the latent space of dependent query like referring expressions (e.g., “the chair left to the table”), but left for future work.a pre-trained CLIP for LSeg via distillation, enabling the decomposition of NeRFs with both English and non-English queries. Segmentation can further be performed with other modalities such as image, patch or pixel query features $\mathbf { f } _ { \mathrm { q } }$ 5using a similar dot-product similarity formulation as in Eq. 2. Notably, DINO [12], a self-supervised vision model, solves video instance segmentation and tracking by calculating similarity among features in adjacent frames. Amir et al. [3] also demonstrate that DINO features work well on co-segmentation and point correspondence by similarity and clustering. In our experiments, we use these two publicly available models, LSeg and DINO, to obtain features of images and texts for 3D decomposition.
|
| 55 |
+
|
| 56 |
+
# 4 Distilled Feature Fields
|
| 57 |
+
|
| 58 |
+
# 4.1 Distilling Foundation Modules into 3D Feature Fields via Volume Rendering
|
| 59 |
+
|
| 60 |
+
NeRF learns a neural field to compute the density and view-dependent color, $\sigma ( \mathbf { x } )$ and $\mathbf { c } ( \mathbf { x } , \mathbf { d } )$ . We may extend NeRF by adding decoders for other quantities of interest. For example, SemanticNeRF [112] adds a branch outputting a probability distribution of closed-set semantic labels, trained with supervision via images with ground-truth semantic labels. This enables the prediction of pairs of RGB and semantic segmentation masks from novel views, useful for data augmentation. However, because ground-truth annotation is costly, the method is inefficient as a means of scene editing [104]. For specific domains like traffic scenes [25, 41], we may instead train a closed-set segmentation model and use its prediction for training object-aware neural fields. However, this approach is possible only if the types of objects are limited and the domain-specific supervised dataset is available; limiting the application of scene editing in terms of domain and flexibility of decomposition.
|
| 61 |
+
|
| 62 |
+
We build on top of these ideas and perform 3D zero-shot segmentation of NeRFs using open-set text labels or other feature queries. Instead of a branch performing closed-set classification, we propose to add a feature branch outputting a feature vector itself. This branch models a 3D feature field describing the semantics of each spatial point. We supervise the feature field by a pretrained pixel-level image encoder $\mathbf { f } _ { \mathrm { i m g } }$ as a teacher network. Given a 3D coordinate $\mathbf { x }$ , the feature field outputs a feature vector $\mathbf { f } \left( \mathbf { x } \right)$ in addition to density $\sigma ( \mathbf { x } )$ and color $\mathbf { c } ( \mathbf { x } , \mathbf { d } )$ , as shown in Fig. 1. Volume rendering of the feature field is similarly performed via
|
| 63 |
+
|
| 64 |
+
$$
|
| 65 |
+
\hat { \mathbf { F } } ( \mathbf { r } ) = \sum _ { k = 1 } ^ { K } \hat { T } ( t _ { k } ) \alpha ( \sigma ( \mathbf { x } _ { k } ) \delta _ { k } ) \mathbf { f } ( \mathbf { x } _ { k } ) \ .
|
| 66 |
+
$$
|
| 67 |
+
|
| 68 |
+
We can optimize f by minimizing the difference between rendered features $\hat { \mathbf { F } } ( \mathbf { r } )$ and the teacher’s features ${ \bf f } _ { \mathrm { i m g } } ( I , r )$ . Effectively, we are distilling [34] the 2D teacher network into our 3D student network via differentiable rendering, and thus dub this model a distilled feature field (DFF). We add a feature objective $\mathcal { L } _ { f }$ penalizing the difference between rendered features $\hat { \mathbf { F } } ( \mathbf { r } )$ and the teacher’s outputs ${ \bf f } _ { \mathrm { i m g } } ( I , r )$ to the photometric loss of the original NeRF. We use two networks for volume rendering with coarse-and-fine hierarchical sampling. We thus minimize the sum of photometric loss $L _ { p }$ and feature loss $L _ { f }$ , in total, $L$ :
|
| 69 |
+
|
| 70 |
+
$$
|
| 71 |
+
L = L _ { p } + \lambda L _ { f } , L _ { p } = \sum _ { \mathbf { r } \in \mathbb { R } } \left\| \hat { \mathbf { C } } ( \mathbf { r } ) - \mathbf { C } ( r ) \right\| _ { 2 } ^ { 2 } , L _ { f } = \sum _ { \mathbf { r } \in \mathbb { R } } \left\| \hat { \mathbf { F } } ( \mathbf { r } ) - \mathbf { f } _ { \mathrm { i m g } } ( I , r ) \right\| _ { 1 } ,
|
| 72 |
+
$$
|
| 73 |
+
|
| 74 |
+
where $\mathcal { R }$ are sampled rays, $\mathbf { C } ( r )$ is the ground truth pixel color of ray $r$ , $\lambda$ is the weight of the feature loss and is set to 0.04 to balance the losses [112]. We apply stop-gradient to density in rendering of features $\hat { \mathbf { F } } ( \mathbf { r } )$ in Equation 3 as the teacher’s features ${ \bf f } _ { \mathrm { i m g } } ( I , r )$ are not fully multi-view consistent, which could harm the quality of reconstructed geometry.
|
| 75 |
+
|
| 76 |
+
# 4.2 Query-based Decomposition and Editing
|
| 77 |
+
|
| 78 |
+
A trained DFF model can perform 3D zero-shot segmentation by its feature field $\mathbf { f }$ and a query encoder $\mathbf { f } _ { \mathrm { q } }$ . Probability of a label $l$ of a point $\mathbf { x }$ in the 3D space, ${ \bf p } ( l | { \bf x } )$ , is calculated by dot product of the 3D feature $\mathbf { f } \left( \mathbf { x } \right)$ and text label feature ${ \bf f } _ { \mathrm { q } } ( l )$ followed by a softmax:
|
| 79 |
+
|
| 80 |
+
$$
|
| 81 |
+
\mathbf { p } ( l | \mathbf { x } ) = \frac { \exp ( \mathbf { f } ( \mathbf { x } ) \mathbf { f } _ { \mathrm { q } } ( l ) ^ { \mathrm { T } } ) } { \sum _ { l ^ { \prime } \in \mathcal { L } } \exp ( \mathbf { f } ( \mathbf { x } ) \mathbf { f } _ { \mathrm { q } } ( l ^ { \prime } ) ^ { \mathrm { T } } ) } \mathrm { ~ . ~ }
|
| 82 |
+
$$
|
| 83 |
+
|
| 84 |
+
This query-based segmentation field is at the core of the proposed method. It can be calculated at any 3D point without limiting resolution, naturally used in tandem with a radiance field and volume rendering. Note that the segmentation depends on only the 3D coordinate and the query1. As the original NeRF, it is thus multi-view consistent. In addition and important for interactive editing, we can change the segmentation via queries without re-training, which cannot be realized by closed-set methods using semantic [112] or instance segmentation annotation [104]. We may now use this query-conditional segmentation to identify a specific 3D region for editing. Various edits can be generalized to the merging of two NeRF scenes $\sigma _ { 1 } ( \mathbf { x } ) , \mathbf { c } _ { 1 } ( \mathbf { x } , \mathbf { d } )$ and $\sigma _ { 2 } ( \mathbf { x } ) , \mathbf { c } _ { 2 } ( \mathbf { x } , \mathbf { d } )$ , where we use the segmentation field $\mathbf { p }$ for blending. In the experiments section, we simply modify Eq. 1 as a blend of two scenes based on the ratio of $\alpha$ :
|
| 85 |
+
|
| 86 |
+
$$
|
| 87 |
+
\begin{array} { l } { { \displaystyle { \hat { \bf C } } ( { \bf r } ) = \sum _ { k = 1 } ^ { K } { \hat { T } } ( t _ { k } ) \left( \alpha ( \sigma _ { 1 } ( { \bf x } _ { k } ) \delta _ { k } ) { \bf c } _ { 1 } ( { \bf x } _ { k } , { \bf d } ) \rho _ { k } + \alpha ( \sigma _ { 2 } ( { \bf x } _ { k } ) \delta _ { k } ) { \bf c } _ { 2 } ( { \bf x } _ { k } , { \bf d } ) ( 1 - \rho _ { k } ) \right) } , } \\ { { \displaystyle ~ , ~ \rho _ { k } = \frac { \alpha \left( \sigma _ { 1 } ( { \bf x } _ { k } ) \delta _ { k } \right) } { \alpha \left( \sigma _ { 1 } ( { \bf x } _ { k } ) \delta _ { k } \right) + \alpha \left( \sigma _ { 2 } ( { \bf x } _ { k } ) \delta _ { k } \right) } , ~ { \hat { T } } ( t _ { k } ) = \prod _ { k ^ { \prime } = 1 } ^ { k - 1 } \alpha ( \sigma _ { 1 } ( { \bf x } _ { k ^ { \prime } } ) \delta _ { k ^ { \prime } } ) + \alpha ( \sigma _ { 2 } ( { \bf x } _ { k ^ { \prime } } ) \delta _ { k ^ { \prime } } ) } . } \end{array}
|
| 88 |
+
$$
|
| 89 |
+
|
| 90 |
+
For example, if we want to apply a geometric transformation $\mathbf { g }$ to a region of a query $l$ in a NeRF scene $( \sigma , \mathbf { c } )$ , we can render the transformed scene via Eqs. 6 and 7 by setting $\alpha ( \sigma _ { 1 } ( { \bf x } _ { k } ) \delta _ { k } ) =$ $( 1 - \mathbf { p } ( l | \mathbf { x } _ { k } ) ) \alpha ( \sigma ( \mathbf { x } _ { k } ) \delta _ { k } )$ , $\boldsymbol { \alpha } ( \sigma _ { 2 } ( \mathbf { x } _ { k } ) \boldsymbol { \delta } _ { k } ) = \mathbf { p } ( l | \mathbf { g } ^ { - 1 } ( \mathbf { x } _ { k } ) \big ) \boldsymbol { \alpha } ( \sigma ( \mathbf { g } ^ { - 1 } ( \mathbf { x } _ { k } \big ) ) \boldsymbol { \delta } _ { k } ) ,$ $\mathbf { c } _ { 1 } ( \mathbf { x } _ { k } , \mathbf { d } ) = ( 1 -$ $\mathbf { p } ( l | \mathbf { x } _ { k } ) ) \mathbf { c } ( \mathbf { x } _ { k } , \mathbf { d } )$ , and $\mathbf { c } _ { 2 } ( \mathbf { x } _ { k } , \mathbf { d } ) = \mathbf { p } ( l | \mathbf { g } ^ { - 1 } ( \mathbf { x } _ { k } ) ) \mathbf { c } ( \mathbf { g } ^ { - 1 } ( \mathbf { x } _ { k } ) , \mathbf { g } ^ { - 1 } ( \mathbf { d } ) )$ . More details of editing for colorization, translation, and deletion are shown in Appendix B. We can combine this with more complex edits, including optimization-based methods like CLIPNeRF [97]. While CLIPNeRF itself cannot selectively edit specific regions in multi-object scenes, our decomposition method enables it to update only desired objects without breaking unintended areas.
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
Figure 2: Comparison of predictions by coarse and fine MLPs.
|
| 94 |
+
|
| 95 |
+
Table 1: Performance of 3D semantic segmentation on Replica dataset. DFF outperforms a supervised point-cloud segmentation model MinkowskiNet42.
|
| 96 |
+
|
| 97 |
+
<table><tr><td></td><td>mIoU</td><td>accuracy</td></tr><tr><td>Supervised 3DCNN</td><td>0.475</td><td>0.758</td></tr><tr><td>DFF (Coarse)</td><td>0.589</td><td>0.855</td></tr><tr><td>DFF (Fine)</td><td>0.583</td><td>0.855</td></tr></table>
|
| 98 |
+
|
| 99 |
+
# 5 Experiments
|
| 100 |
+
|
| 101 |
+
We first conduct a quantitative evaluation of the decomposition achieved by DFF. We demonstrate that DFF enables 3D semantic segmentation in a benchmark dataset using scanned point clouds with human-annotated semantic segmentation labels. We then investigate the capabilities of DFF for editing and subsequent novel-view synthesis on real-world datasets. We use two teacher networks, LSeg [44] and DINO [12], which are pre-trained and publicly available. Each training image is encoded by the image encoders of the networks and used as target feature maps, ${ \bf f } _ { \mathrm { i m g } } ( I , r )$ , defined in Equation 4. Because the feature maps are of reduced sizes due to the limitation of the networks, we first resize them to the original image size. The implementation and settings of NeRF, unless otherwise stated, follow Zhi et al. [112]. During the training of 200K iterations, the loss $L$ in Equation 4 is minimized by Adam with a linearly decaying learning rate (5e-4 to 8e-5). During training, Gaussian noise for density is also applied. The number of coarse and fine samplings is 64 and 128, respectively. The MLP of the neural radiance field consists of eight ReLU layers with 256 dimensions, followed by a linear layer for density, three layers for color, and three layers for feature, as shown in Fig. 1. Positional encoding of length 10 is used for the input coordinate and its skip connection, and that of length 4 is for viewing direction. If an independent MLP is prepared for the feature field, it consists of four layers (with a skip connection at the third layer if the positional encoding is used). The size of a training image is $3 2 0 \times 2 4 0$ for the Replica dataset and $1 0 0 8 \times 7 5 6$ for the other datasets. The batchsize of training rays is 1024 for Replica and 2048 for the others. During finetuning of feature fields or radiance fields, Gaussian noise is removed, and the learning rate is set to 1e-4. See appendix A and C for further training details.
|
| 102 |
+
|
| 103 |
+
# 5.1 3D Semantic Segmentation
|
| 104 |
+
|
| 105 |
+
We construct a 3D semantic segmentation benchmark from four scenes in the Replica dataset [86] with data split and posed images provided by [112]. See appendix D for further details of the dataset. We train DFF to reconstruct each scene with radiance and feature fields from training images and evaluate the quality of novel view synthesis and 3D segmentation of the annotated point clouds. We use LSeg as a teacher network. The LSeg text encoder encodes each label, and the probability of each point is calculated by Equation $\cdot$ . Note that the training uses only the photometric and feature losses (Equation 4) and does not access any supervision via semantic labels.
|
| 106 |
+
|
| 107 |
+
Semantic Segmentation Results. First, we show evaluation metrics of 3D semantic segmentation, mean intersection-over-union (mIoU) and accuracy in Table 1. For comparison, we also experiment with a sparse 3D convolution-based segmentation model, MinkowskiNet42 [18] taking a colored point cloud as input. It has a standard state-of-the-art architecture for point cloud segmentation and is trained on the ScanNet dataset [20], the largest annotated training dataset of 3D semantic segmentation3. Results demonstrate that DFF, taught by $\mathrm { L S e g }$ , achieves promising performance, even better than the supervised model. This indicates that DFF succeeds at distilling 3D semantic segmentation from the 2D teacher network.
|
| 108 |
+
|
| 109 |
+
Table 2: Performance of novel view synthesis on Replica dataset. PSNR, SSIM, and LPIPS are metrics of image synthesis. $\delta { < } 1 . 2 5$ and absrel are metrics of geometry (depth estimation).
|
| 110 |
+
|
| 111 |
+
<table><tr><td></td><td>PSNR↑</td><td>SSIM↑</td><td>LPIPS↓</td><td>δ<1.25↑</td><td>absrel</td></tr><tr><td>basic NeRF</td><td>32.87</td><td>0.934</td><td>0.148</td><td>0.993</td><td>0.018</td></tr><tr><td>DFF</td><td>32.85</td><td>0.932</td><td>0.150</td><td>0.993</td><td>0.017</td></tr><tr><td>DFF (overweighting 入)</td><td>32.68</td><td>0.927</td><td>0.162</td><td>0.993</td><td>0.018</td></tr></table>
|
| 112 |
+
|
| 113 |
+
Impact of Sampling on Semantic Segmentation. NeRF employs two MLPs for hierarchical sampling, where the coarse MLP performs volume rendering with fewer points (64) using stratified sampling, and the fine MLP works with importance sampling (192 in total). So, we have two sampling options to train a feature field. Although fine sampling is critical for training accurate radiance fields, segmentation is of significantly lower spatial frequency than texture. We thus analyze the impact of coarse and fine training in Fig. 2. As expected, the coarse model produces smooth segmentations, while the fine version introduces high-frequency artifacts. This smoothness property is important for natural editable novel view synthesis and is discussed again later.
|
| 114 |
+
|
| 115 |
+
Compatibility with View Synthesis. We also check and compare the quality of novel view synthesis with NeRF, which does not learn feature fields. Because the feature branch partially shares the layers with the radiance field (as shown in Fig. 1), learning feature fields could possibly harm the radiance field. Despite this concern, as shown in Tab. 2, the performance of view synthesis is not degraded. Thus, we can train and use the branch-based DFF with small computational and parameter overhead compared to the original NeRF. If we excessively increased the weight of the feature loss, $\lambda \times 1 0$ , it hurt view synthesis while not improving segmentation performance further. We further confirm that training independent, light-weight feature-field MLP, instead of a branch of the radiance-field MLP, achieves semantic segmentation results competitive with the branch-based approach (see appendix Tab. 3 for the result of all variants)4. This option is useful especially when we want to introduce DFF decomposition into arbitrary 3D scene representations, including off-the-shelf NeRF models, dynamic NeRFs [26, 66, 45], or meshes, without re-training of the radiance field.
|
| 116 |
+
|
| 117 |
+
# 5.2 Editable Novel View Synthesis
|
| 118 |
+
|
| 119 |
+
In the previous section, we quantitatively validated the ability of DFF to perform semantic decomposition. We now discuss the capability for editable view synthesis on real-world scenes, including the LLFF dataset [55] and our own dataset. Our method can be used even for LLFF scenes based on normalized device coordinates. Please see the supplemental web page for further results, including videos. In addition to $\mathrm { L S e g }$ using a text query, we also experiment with self-supervised DINO [12] as another teacher network to enable query-based decomposition using image patch queries. Here, we use thresholded cosine similarity to directly compute the probability of a query instead of softmax with negative queries in Eq. 5 and set $\mathbf { p } = 1$ if the similarity exceeds the threshold 5, and $\mathbf { p } = 0$ otherwise for hard decomposition. We first train NeRFs without a feature branch for each scene for 200K iterations $( L _ { p } )$ and then finetune them with a feature branch via distillation for 5K iterations $( L _ { p } + \lambda L _ { f } )$ since we found that the feature loss converged significantly faster than the photometric loss and short training was thus sufficient. We use coarse sampling for training feature branches and use it for edited rendering with fine sampling. See appendix A for the details.
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
Figure 3: Appearance edits of specific objects via different query modalities: an image patch or text.
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
Figure 4: Extraction and deletion of specific objects via different query modalities, an image patch or text. The edited views are 3D consistent, unlike an image inpainting baseline [87]
|
| 126 |
+
|
| 127 |
+
Appearance Editing, Deletion, Extraction. We show qualitative evaluations of novel view synthesis in Fig. 3 and Fig. 4. Specific 3D regions in these scenes are identified and locally edited via decomposition depending on various query modalities. In these experiments, we use a text query for LSeg-DFF as in Section 5.1 and use an image patch query for DINO-DFF. Because DINO features capture the similarity and correspondences of regions well thanks to self-supervised learning [12, 3], image patch queries help select all semantically similar areas at once. The patch feature is then calculated by averaging the features of all pixels in the patch.
|
| 128 |
+
|
| 129 |
+
In Fig. 3, we demonstrate that the DFF enables convincing selective appearance edits. Because our focus is region selection via decomposition, we use simple color transformation for clarity here (e.g., flip RGB to BGR, blend colors). One might think that the MLP of a radiance field by the original NeRF also has hidden layers, and their features could possibly be used for decomposition. We confirm that the naive usage of NeRF features is not robust to decomposition, as shown in Fig. 5, especially in a complex multi-object scene. We use the 8th hidden layer of the fine radiance field network (i.e., the layer just before branching in Fig. 1) 6. NeRF features cannot clearly decompose even objects with simple shapes and colors. The region selections are leaked to other parts with similar colors, geometry, or positions while they do not entirely cover the targets. For example, floor selection is leaked to walls, a table, bins, or ceilings. Chair selection is leaked to irrelevant black parts like television, cables, lighting equipment, or shadows. This indicates that the feature space of the original NeRF does not learn semantic similarity well and is entangled with unpredictable and more low-level factors like color or spatial adjacency.
|
| 130 |
+
|
| 131 |
+
In Fig. 4, we demonstrate that the DFF also works well on deletion or extraction of objects, using two patch queries (query- $\textcircled{1}$ for leaves and ground, query- $\textcircled{2}$ for flowers) and a text query- $\textcircled{3}$ “flower”. For comparison with a baseline editing method, we show the results by a state-of-the-art image inpainting model, LaMa [87]. Because the model requires masks for inpainting regions, we manually annotate the views for evaluation. As shown in the figure, the image inpainting model cannot generate clear and realistic images, and the different views are inconsistent. On the other hand, DFF produces multi-view consistent plausible results, especially succeeding at extracting foreground objects. Although the performance on deleting foreground objects is high, a remaining shortcoming is the existence of floating artifacts and blurred volumes in the far distance behind the deleted object.
|
| 132 |
+
|
| 133 |
+
Priors for Smooth Decomposition We can organize the challenges of editable NeRFs into several categories: surface decomposition, volume decomposition, lighting decomposition, and estimation of less or never observed parts. If we edit appearances only, it practically requires decomposing regions only near the surface of objects, i.e., surface decomposition, because the color of a ray is determined mostly in a condensed interval around the surface. On the other hand, geometric transformations often require a higher level of decomposition. As shown in the deletion examples, geometric transformation may move or remove some surfaces and expose the space behind them. This forces models to render unknown regions less or never observed due to occlusions, including even the inside of objects. Thus, it is desirable to decompose volumes smoothly while synthesizing their inside and back7. Although these include the same challenges as novel view synthesis tackles, editability further highlights their importance.
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
Figure 5: Appearance edits of specific objects, compared with decomposition using features of a NeRF hidden layer. For reference, we also show PCA-based visualizations of the features.
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
Figure 6: Comparison of predictions by a branch-based feature field MLP and independent MLP with no positional encoding, each of which is trained with coarse and fine sampling.
|
| 140 |
+
|
| 141 |
+
Apart from lighting decomposition discussed in prior work [8, 9, 111], we further investigate the new challenge of smooth volume decomposition by experimenting with different DFF setups. As discussed in Section 5.1, DFF has two sampling options to train feature fields. The coarse training may introduce smoothness regularization and help cohesive decomposition and smoother in-painting of unobserved regions. Another reasonable smoothness regularizer is to eliminate the high-frequency positional encoding (PE). We thus train an independent MLP network for a feature field without PE. We compare four combinations of renderings in Fig. 6. To better understand their behavior, we use the DINO-DFF, show k-means clusters of the rendered feature map, and delete the head of the Triceratops by a query choosing its corresponding clusters. As expected, coarsely trained models and no-PE models succeed in smoother volume decomposition, and this combination can minimize high-frequency floating artifacts. A side effect is the lack of high-frequency representation power, which sometimes deletes disparate background regions and misses to represent features of complex structures (e.g., see the cluster visualization of the thin frames of the window). Towards the best of both worlds, developing proper priors or inductive biases is an important direction for future work [70]. Otherwise, surface-aware representations like IDR [106, 98] could avoid problems with floating artifacts. Note that not all geometric edits suffer from these problems. For example, it is often less problematic to move objects closer to the camera, enlarge them, or warp them to other scenes, as shown in Fig. 7.
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
Figure 7: Editing with warping, deformation, shift, and rotation.
|
| 145 |
+
|
| 146 |
+
Localizing Optimization-based Editing. Finally, we show a combination with an optimizationbased editing method. CLIPNeRF [97] optimizes the parameters of a radiance field so that its rendered images match with a text prompt via CLIP. While it is mainly designed for a single-object scene of specific categories, it is possible to apply to other real-world NeRFs. However, because it cannot control the scope of editing, a prompt like “white flower” may change the color of unintentional targets like leaves. Our DFF-based decomposition can upgrade such an optimization-based method to render a scene via the composition of a CLIP-optimized NeRF scene and the original NeRF scene. We show the results in Fig. 88. Although the naive CLIPNeRF edits unintentional parts, our method helps it to locally edit intentional parts only. In addition to switching rendering, we can also use the decomposition for controlling training signals during backpropagation. The additional experiment is shown in Appendix F. These extensions broaden the application of CLIPNeRF or other optimization-based editing methods to complex scenes.
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
Figure 8: Comparison of appearance editing by CLIPNeRF and our extension.
|
| 150 |
+
|
| 151 |
+
# 6 Discussion, Limitations, and Conclusions
|
| 152 |
+
|
| 153 |
+
In this work, we propose distilled feature field (DFF), a novel method of NeRF scene decomposition for selective editing. We present quantitative evaluations of segmentation and extensive qualitative evaluations of editable novel view synthesis. In addition to these promising results, DFF-based models will benefit from future improvements to self-supervised 2D foundation models. We also clarify future directions on editable view synthesis through our experiments, especially for smoothness priors and estimation of unobserved regions. Furthermore, while this work focuses on editable view synthesis, it is also intriguing to transfer DFF to other applications, including 3D registration of text queries [16, 1, 47, 4] or robot teaching [32, 80].
|
| 154 |
+
|
| 155 |
+
The limitations of the DFF framework are two-fold. The first one is the upper bduround of the performance due to distillation. The student model of distillation cannot largely outperform the teacher model9. If the resolution of teacher encoders is low, the corresponding DFFs also becomes coarse-grained. If the LSeg cannot understand a text query, the LSeg-DFF also cannot. Secondly, the DFF uses volume rendering depending on the 3D reconstruction by NeRF. A NeRF model is sometimes optimized to geometrically wrong solutions (e.g., floaters). Such geometry errors of the radiance fields would make supervision to DFFs noisy.
|
| 156 |
+
|
| 157 |
+
As a possible negative societal impact, one might use our method for making realistic but fake content by editing NeRFs as desired. Automatic fake detection methods may help in preventing such misuse. NeRFs are further computation-intense, leading to high electricity usage. Recent work on efficient NeRFs [23, 57, 15] may alleviate this concern.
|
| 158 |
+
|
| 159 |
+
# Acknowledgements
|
| 160 |
+
|
| 161 |
+
We thank Tsukasa Takagi, Toshiki Nakanishi, Hiroharu Kato, and Masaaki Fukuda for their helpful feedback.
|
| 162 |
+
|
| 163 |
+
# References
|
| 164 |
+
|
| 165 |
+
[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In 16th European Conference on Computer Vision (ECCV), 2020.
|
| 166 |
+
[2] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38 (7):1425–1438, 2016.
|
| 167 |
+
|
| 168 |
+
[3] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021.
|
| 169 |
+
|
| 170 |
+
[4] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
|
| 171 |
+
|
| 172 |
+
[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Proceedings of the European Conference on Computer Vision (ECCV), 2022.
|
| 173 |
+
|
| 174 |
+
[6] Sagie Benaim, Frederik Warburg, Peter Ebert Christensen, and Serge Belongie. Volumetric disentanglement for 3d scene manipulation. arXiv, 2022.
|
| 175 |
+
|
| 176 |
+
[7] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri S. Chatterji, Annie S. Chen, Kathleen Creel, Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li FeiFei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah D. Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark S. Krass, Ranjay Krishna, Rohith Kuditipudi, and et al. On the opportunities and risks of foundation models. arXiv, 2021. URL https://arxiv.org/abs/2108.07258.
|
| 177 |
+
|
| 178 |
+
[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P. A. Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL http://arxiv. org/abs/2012.03918v4.
|
| 179 |
+
|
| 180 |
+
[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik Lensch. Neural-PIL: Neural pre-integrated lighting for reflectance decomposition. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. URL https://openreview.net/forum?id=fATZNtA1-V0.
|
| 181 |
+
|
| 182 |
+
[10] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché- Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32, 2019. URL https://proceedings.neurips.cc/paper/2019/file/ 0266e33d3f546cb5436a10798e657d97-Paper.pdf.
|
| 183 |
+
|
| 184 |
+
[11] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Coco-stuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209–1218, 2018.
|
| 185 |
+
|
| 186 |
+
[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650–9660, October 2021.
|
| 187 |
+
|
| 188 |
+
[13] Rohan Chabra, Jan Eric Lenssen, Eddy Ilg, Tanner Schmidt, Julian Straub, Steven Lovegrove, and Richard Newcombe. Deep local shapes: Learning local sdf priors for detailed 3d reconstruction. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. URL http://arxiv.org/abs/2003.10983v3.
|
| 189 |
+
|
| 190 |
+
[14] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL http://arxiv.org/abs/2103.15595v2.
|
| 191 |
+
|
| 192 |
+
[15] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), 2022.
|
| 193 |
+
|
| 194 |
+
[16] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. 16th European Conference on Computer Vision (ECCV), 2020.
|
| 195 |
+
|
| 196 |
+
[17] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9640–9649, 2021.
|
| 197 |
+
|
| 198 |
+
[18] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3075–3084, 2019.
|
| 199 |
+
|
| 200 |
+
[19] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016.
|
| 201 |
+
|
| 202 |
+
[20] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017.
|
| 203 |
+
|
| 204 |
+
[21] Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. Nasa: Neural articulated shape approximation. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. URL http://arxiv.org/abs/1912.03207v4.
|
| 205 |
+
|
| 206 |
+
[22] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. URL https:/ /openreview.net/forum?id $\cdot ^ { = }$ YicbFdNTTy.
|
| 207 |
+
|
| 208 |
+
[23] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022.
|
| 209 |
+
|
| 210 |
+
[24] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc' Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc., 2013. URL https://proceedings.neurips.cc/paper/2013/file/ 7cce53cf90577442771720a370c3c723-Paper.pdf.
|
| 211 |
+
|
| 212 |
+
[25] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. arxiv, 2022. URL https://arxiv.org/abs/2203.15224.
|
| 213 |
+
|
| 214 |
+
[26] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. arXiv preprint arXiv:2105.06468, 2021. URL http://arxiv. org/abs/2105.06468v1.
|
| 215 |
+
|
| 216 |
+
[27] Jonathan Granskog, Till N Schnabel, Fabrice Rousselle, and Jan Novák. Neural scene graph rendering. ACM Transactions on Graphics (TOG), 40(4):1–11, 2021.
|
| 217 |
+
|
| 218 |
+
[28] Michelle Guo, Alireza Fathi, Jiajun Wu, and Thomas Funkhouser. Object-centric neural scene rendering. arXiv preprint arXiv:2012.08503, 2020. URL http://arxiv.org/abs/2012. 08503v1.
|
| 219 |
+
|
| 220 |
+
[29] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338–4364, 2021. doi: 10.1109/TPAMI.2020.3005434.
|
| 221 |
+
|
| 222 |
+
[30] Huy Ha and Shuran Song. Semantic abstraction: Open-world 3D scene understanding from 2D vision-language models. In Proceedings of the 2022 Conference on Robot Learning, 2022.
|
| 223 |
+
|
| 224 |
+
[31] Zekun Hao, Arun Mallya, Serge Belongie, and Ming-Yu Liu. Gancraft: Unsupervised 3d neural rendering of minecraft worlds. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL http://arxiv.org/abs/2104.07659v1.
|
| 225 |
+
|
| 226 |
+
[32] Jun Hatori, Yuta Kikuchi, Sosuke Kobayashi, Kuniyuki Takahashi, Yuta Tsuboi, Yuya Unno, Wilson Ko, and Jethro Tan. Interactively picking real-world objects with unconstrained spoken language instructions. In Proceedings of International Conference on Robotics and Automation, 2018.
|
| 227 |
+
|
| 228 |
+
[33] Yong He, Hongshan Yu, Xiaoyan Liu, Zhengeng Yang, Wei Sun, Yaonan Wang, Qiang Fu, Yanmei Zou, and Ajmal Mian. Deep learning based 3d segmentation: A survey. arxiv, 2021. URL https://arxiv.org/abs/2103.05423.
|
| 229 |
+
|
| 230 |
+
[34] Geoffrey Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. URL http://arxiv.org/abs/1503.02531.
|
| 231 |
+
|
| 232 |
+
[35] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885–5894, October 2021.
|
| 233 |
+
|
| 234 |
+
[36] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. URL https://arxiv.org/ abs/2112.01455.
|
| 235 |
+
|
| 236 |
+
[37] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL http://arxiv.org/abs/2109.01750v1.
|
| 237 |
+
|
| 238 |
+
[38] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 2021. URL http://arxiv.org/abs/ 2109.11418v1.
|
| 239 |
+
|
| 240 |
+
[39] Hiroharu Kato, Deniz Beker, Mihai Morariu, Takahiro Ando, Toru Matsuoka, Wadim Kehl, and Adrien Gaidon. Differentiable rendering: A survey. arxiv, 2020. URL https://arxiv. org/abs/2006.12057.
|
| 241 |
+
|
| 242 |
+
[40] Amit Kohli, Vincent Sitzmann, and Gordon Wetzstein. Semantic implicit neural scene representations with semi-supervised training. In International Conference on 3D Vision (3DV). IEEE, 2020. URL http://arxiv.org/abs/2003.12673v2.
|
| 243 |
+
|
| 244 |
+
[41] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation. In CVPR, 2022.
|
| 245 |
+
|
| 246 |
+
[42] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multi-domain semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879–2888, 2020.
|
| 247 |
+
|
| 248 |
+
[43] Verica Lazova, Vladimir Guzov, Kyle Olszewski, Sergey Tulyakov, and Gerard Pons-Moll. Control-nerf: Editable feature volumes for scene rendering and manipulation. arXiv preprint arXiv:2204.10850, 2022.
|
| 249 |
+
|
| 250 |
+
[44] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Languagedriven semantic segmentation. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=RriDjddCLN.
|
| 251 |
+
|
| 252 |
+
[45] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, and Zhaoyang Lv. Neural 3d video synthesis. arXiv preprint arXiv:2103.02597, 2021. URL http://arxiv.org/abs/ 2103.02597v1.
|
| 253 |
+
|
| 254 |
+
[46] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014.
|
| 255 |
+
|
| 256 |
+
[47] Haolin Liu, Anran Lin, Xiaoguang Han, Lei Yang, Yizhou Yu, and Shuguang Cui. Refer-it-inrgbd: A bottom-up approach for 3d visual grounding in rgbd images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6032–6041, 2021.
|
| 257 |
+
|
| 258 |
+
[48] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. URL http://arxiv.org/abs/2007.11571v2.
|
| 259 |
+
|
| 260 |
+
[49] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. arXiv preprint arXiv:2105.06466, 2021. URL http://arxiv.org/abs/2105.06466v2.
|
| 261 |
+
|
| 262 |
+
[50] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer V2: scaling up capacity and resolution. arxiv, abs/2111.09883, 2021. URL https://arxiv.org/abs/2111.09883.
|
| 263 |
+
|
| 264 |
+
[51] Sebastian Loeschcke, Serge Belongie, and Sagie Benaim. Text-driven stylization of video objects. arXiv, 2022. URL https://arxiv.org/abs/2206.12396.
|
| 265 |
+
|
| 266 |
+
[52] Timo Lüddecke and Alexander Ecker. Prompt-based multi-modal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. URL https://arxiv.org/abs/2112.10003.
|
| 267 |
+
|
| 268 |
+
[53] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. URL http://arxiv.org/abs/1812.03828v2.
|
| 269 |
+
|
| 270 |
+
[54] Björn Michele, Alexandre Boulch, Gilles Puy, Maxime Bucher, and Renaud Marlet. Generative zero-shot learning for semantic segmentation of 3d point clouds. In 2021 International Conference on 3D Vision (3DV), pages 992–1002, 2021. doi: 10.1109/3DV53792.2021.00107.
|
| 271 |
+
|
| 272 |
+
[55] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren $\mathrm { N g }$ , and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019.
|
| 273 |
+
|
| 274 |
+
[56] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. URL http: //arxiv.org/abs/2003.08934v2.
|
| 275 |
+
|
| 276 |
+
[57] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1–102:15, July 2022. doi: 10.1145/3528223.3530127. URL https://doi.org/10.1145/3528223. 3530127.
|
| 277 |
+
|
| 278 |
+
[58] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990–4999, 2017.
|
| 279 |
+
|
| 280 |
+
[59] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proc. ICCV, pages 7588–7597, 2019.
|
| 281 |
+
|
| 282 |
+
[60] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6767–6778, 2020. URL https://proceedings. neurips.cc/paper/2020/file/4b29fa4efe4fb7bc667c7b301b74d52d-Paper.pdf.
|
| 283 |
+
|
| 284 |
+
[61] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. URL http://arxiv.org/abs/2011.12100v2.
|
| 285 |
+
|
| 286 |
+
[62] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
|
| 287 |
+
|
| 288 |
+
[63] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL http://arxiv.org/abs/2104.03110v2.
|
| 289 |
+
|
| 290 |
+
[64] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2856–2865, June 2021.
|
| 291 |
+
|
| 292 |
+
[65] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. URL http://arxiv.org/abs/1901.05103v1.
|
| 293 |
+
|
| 294 |
+
[66] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. URL http://arxiv.org/abs/2106.13228v2.
|
| 295 |
+
|
| 296 |
+
[67] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.
|
| 297 |
+
|
| 298 |
+
[68] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. URL https://arxiv.org/abs/2209.14988.
|
| 299 |
+
|
| 300 |
+
[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. PMLR, 18–24 Jul 2021. URL https://proceedings.mlr.press/v139/radford21a. html.
|
| 301 |
+
|
| 302 |
+
[70] Sameera Ramasinghe, Lachlan E. MacDonald, and Simon Lucey. On regularizing coordinatemlps. arxiv, 2022.
|
| 303 |
+
|
| 304 |
+
[71] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL https://arxiv.org/abs/2103.13413.
|
| 305 |
+
|
| 306 |
+
[72] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
|
| 307 |
+
|
| 308 |
+
[73] Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, and Andrea Tagliasacchi. Derf: Decomposed radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. URL http://arxiv.org/abs/2011. 12490v1.
|
| 309 |
+
|
| 310 |
+
[74] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural
|
| 311 |
+
|
| 312 |
+
Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992, November 2019. doi: 10.18653/v1/D19-1410. URL https://aclanthology.org/D19-1410.
|
| 313 |
+
|
| 314 |
+
[75] Nils Reimers and Iryna Gurevych. Making monolingual sentence embeddings multilingual using knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4512–4525, November 2020. doi: 10.18653/ v1/2020.emnlp-main.365. URL https://aclanthology.org/2020.emnlp-main.365.
|
| 315 |
+
|
| 316 |
+
[76] Zhongzheng Ren, Aseem Agarwala†, Bryan Russell†, Alexander G. Schwing†, and Oliver Wang†. Neural volumetric object selection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. († alphabetic ordering).
|
| 317 |
+
|
| 318 |
+
[77] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022.
|
| 319 |
+
|
| 320 |
+
[78] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022.
|
| 321 |
+
|
| 322 |
+
[79] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic textto-image diffusion models with deep language understanding. arXiv, 2022. URL https: //arxiv.org/abs/2205.11487.
|
| 323 |
+
|
| 324 |
+
[80] Anthony Simeonov, Yilun Du, Andrea Tagliasacchi, Joshua B. Tenenbaum, Alberto Rodriguez, Pulkit Agrawal, and Vincent Sitzmann. Neural descriptor fields: Se(3)-equivariant object representations for manipulation. arXiv preprint arXiv:2112.05124, 2021.
|
| 325 |
+
|
| 326 |
+
[81] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhöfer. Deepvoxels: Learning persistent 3d feature embeddings. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2019.
|
| 327 |
+
|
| 328 |
+
[82] Vincent Sitzmann, Michael Zollhofer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2019. URL http: //arxiv.org/abs/1906.01618v2.
|
| 329 |
+
|
| 330 |
+
[83] Cameron Smith, Hong-Xing Yu, Sergey Zakharov, Fredo Durand, Joshua B. Tenenbaum, Jiajun Wu, and Vincent Sitzmann. Unsupervised discovery and composition of object light fields. arXiv, 2022.
|
| 331 |
+
|
| 332 |
+
[84] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567–576, 2015.
|
| 333 |
+
|
| 334 |
+
[85] Karl Stelzner, Kristian Kersting, and Adam R. Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. arxiv, 2021. URL http://arxiv.org/abs/2104. 01148.
|
| 335 |
+
|
| 336 |
+
[86] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arxiv, 2019.
|
| 337 |
+
|
| 338 |
+
[87] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2149–2159, January 2022. URL https://arxiv.org/abs/2109.07161.
|
| 339 |
+
|
| 340 |
+
[88] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. URL http://arxiv.org/abs/2101.10994v1.
|
| 341 |
+
|
| 342 |
+
[89] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. arXiv, 2022.
|
| 343 |
+
|
| 344 |
+
[90] A. Tewari, O. Fried, J. Thies, V. Sitzmann, S. Lombardi, Z. Xu, T. Simon, M. Nießner, E. Tretschk, L. Liu, B. Mildenhall, P. Srinivasan, R. Pandey, S. Orts-Escolano, S. Fanello, M. Guo, G. Wetzstein, J.-Y. Zhu, C. Theobalt, M. Agrawala, D. B Goldman, and M. Zollhöfer. Advances in neural rendering. In ACM SIGGRAPH 2021 Courses, SIGGRAPH ’21, 2021. doi: 10.1145/3450508.3464573. URL https://arxiv.org/abs/2111.05849.
|
| 345 |
+
|
| 346 |
+
[91] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the International Conference on Machine Learning (ICML), pages 10347– 10357, 2021.
|
| 347 |
+
|
| 348 |
+
[92] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. Neuraldiff: Segmenting 3d objects that move in egocentric videos. arXiv preprint arXiv:2110.09936, 2021.
|
| 349 |
+
|
| 350 |
+
[93] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3D distillation of self-supervised 2D image representations. In Proceedings of the International Conference on 3D Vision (3DV), 2022. URL https://arxiv.org/pdf/2209.03494.pdf.
|
| 351 |
+
|
| 352 |
+
[94] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. arxiv, 2021. URL https://arxiv.org/ abs/2112.10703.
|
| 353 |
+
|
| 354 |
+
[95] Girish Varma, Anbumani Subramanian, Anoop Namboodiri, Manmohan Chandraker, and CV Jawahar. Idd: A dataset for exploring problems of autonomous navigation in unconstrained environments. In 2019 IEEE Winter Conference on Applications of Computer Vision, pages 1743–1751. IEEE, 2019.
|
| 355 |
+
|
| 356 |
+
[96] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper/2017/file/ 3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf.
|
| 357 |
+
|
| 358 |
+
[97] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Textand-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
|
| 359 |
+
|
| 360 |
+
[98] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI). International Joint Conferences on Artificial Intelligence Organization, 2021. URL http: //arxiv.org/abs/2106.10689v1.
|
| 361 |
+
|
| 362 |
+
[99] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. URL https://arxiv.org/abs/2111.15174.
|
| 363 |
+
|
| 364 |
+
[100] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. Sep 2021. URL http://arxiv.org/abs/2109.01129v2.
|
| 365 |
+
|
| 366 |
+
[101] Xiaoshi Wu, Hadar Averbuch-Elor, Jin Sun, and Noah Snavely. Towers of babel: Combining images, language, and 3d geometry for learning multimodal vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 428–437, October 2021.
|
| 367 |
+
|
| 368 |
+
[102] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond, 2021. URL https://neuralfields.cs.brown.edu/.
|
| 369 |
+
|
| 370 |
+
[103] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for zero-shot semantic segmentation with pre-trained vision-language model. arxiv, 2021. URL https://arxiv.org/abs/2112.14757.
|
| 371 |
+
|
| 372 |
+
[104] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editable scene rendering. In International Conference on Computer Vision (ICCV), October 2021. URL https://arxiv.org/abs/2109.01847.
|
| 373 |
+
|
| 374 |
+
[105] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: Amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Trans. Graph., 41(4):101:1–101:10, July 2022. doi: 10.1145/3528223.3530163. URL https://doi.org/10.1145/3528223. 3530163.
|
| 375 |
+
|
| 376 |
+
[106] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Ronen Basri, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In Advances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., 2020. URL http://arxiv.org/abs/2003.09852v3.
|
| 377 |
+
|
| 378 |
+
[107] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. Proc. ICCV, 2021.
|
| 379 |
+
|
| 380 |
+
[108] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636–2645, 2020.
|
| 381 |
+
|
| 382 |
+
[109] Hong-Xing Yu, Leonidas Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. In International Conference on Learning Representations, 2022. URL https:// openreview.net/forum?id $=$ rwE8SshAlxw.
|
| 383 |
+
|
| 384 |
+
[110] Jason Zhang, Gengshan Yang, Shubham Tulsiani, and Deva Ramanan. Ners: Neural reflectance surfaces for sparse-view 3d reconstruction in the wild. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 29835–29847. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper/2021/file/ f95ec3de395b4bce25b39ef6138da871-Paper.pdf.
|
| 385 |
+
|
| 386 |
+
[111] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. arXiv preprint arXiv:2106.01970, 2021. URL http://arxiv.org/abs/2106. 01970v1.
|
| 387 |
+
|
| 388 |
+
[112] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. URL http://arxiv.org/abs/ 2103.15875v2.
|
| 389 |
+
|
| 390 |
+
[113] Shuaifeng Zhi, Edgar Sucar, Andre Mouton, Iain Haughton, Tristan Laidlow, and Andrew J. Davison. iLabel: Interactive neural scene labelling. arXiv, 2021.
|
| 391 |
+
|
| 392 |
+
[114] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. URL https://arxiv.org/abs/2112.09106.
|
| 393 |
+
|
| 394 |
+
[115] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302–321, 2019.
|
| 395 |
+
|
| 396 |
+
# Checklist
|
| 397 |
+
|
| 398 |
+
1. For all authors...
|
| 399 |
+
|
| 400 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] The claims are empirically demonstrated in Section 5, and described in the whole paper.
|
| 401 |
+
(b) Did you describe the limitations of your work? [Yes] See mainly Section 5, and 6.
|
| 402 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] See Section 6.
|
| 403 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 404 |
+
|
| 405 |
+
2. If you are including theoretical results...
|
| 406 |
+
|
| 407 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] Mathematical formulations of the models are written in Section 3 and 4.
|
| 408 |
+
(b) Did you include complete proofs of all theoretical results? [N/A] Mathematical formulations of the models are written in Section 3 and 4.
|
| 409 |
+
|
| 410 |
+
3. If you ran experiments...
|
| 411 |
+
|
| 412 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] The complete code for the reproduction of all the experimental results is not publicly available. Because the code is a modification from a public code by Zhi et al. [112], reproduction is also easier than from scratch. We will make our scene dataset publicly available.
|
| 413 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] Some of them are further described in the supplementary material.
|
| 414 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No] Error bars are not reported because it would be computationally expensive and results are expected to be stable. Note that most existing studies on NeRF have not reported the bars too.
|
| 415 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No] It is difficult to completely track and sum the total amount of computing in the experiments. Instead, we reported the setup of the main experiments.
|
| 416 |
+
|
| 417 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 418 |
+
|
| 419 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes] Codebase and datasets are appropriately cited, mainly in Section 5.
|
| 420 |
+
(b) Did you mention the license of the assets? [No] We refer the readers to the original source instead of mentioning them in this paper.
|
| 421 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We include generated video by models in the supplemental material.
|
| 422 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] No data about people.
|
| 423 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] No such data.
|
| 424 |
+
|
| 425 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 426 |
+
|
| 427 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] No such data or experiment.
|
| 428 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] No such data or experiment.
|
| 429 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] No such data or experiment.
|
parse/dev/IJNDyqdRF0m/IJNDyqdRF0m_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/IJNDyqdRF0m/IJNDyqdRF0m_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/IJNDyqdRF0m/IJNDyqdRF0m_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/LSKlp_aceOC/LSKlp_aceOC.md
ADDED
|
@@ -0,0 +1,271 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Merging Models with Fisher-Weighted Averaging
|
| 2 |
+
|
| 3 |
+
Michael Matena Colin Raffel Department of Computer Science University of North Carolina at Chapel Hill {mmatena,craffel}@cs.unc.edu
|
| 4 |
+
|
| 5 |
+
# Abstract
|
| 6 |
+
|
| 7 |
+
Averaging the parameters of models that have the same architecture and initialization can provide a means of combining their respective capabilities. In this paper, we take the perspective that this “merging” operation can be seen as choosing parameters that approximately maximize the joint likelihood of the posteriors of the models’ parameters. Computing a simple average of the models’ parameters therefore corresponds to making an isotropic Gaussian approximation to their posteriors. We develop an alternative merging procedure based on the Laplace approximation where we approximate each model’s posterior as a Gaussian distribution whose precision matrix corresponds to its Fisher information. We first show that our “Fisher merging” technique provides a performance boost in settings where simple parameter averaging is currently used – specifically, robust fine-tuning and model ensembling. Then, we compare merging to standard gradient-based transfer learning and demonstrate that merging enables a fundamentally different method for transferring capabilities across models. Specifically, we show that Fisher merging is competitive with gradient-based transfer learning approaches (while being significantly cheaper) in intermediate-task training and domain-adaptive pre-training. We also show that our merging procedure makes it possible to combine models in previously unexplored ways. We release our code to facilitate future research into methods for merging models.1
|
| 8 |
+
|
| 9 |
+
# 1 Introduction
|
| 10 |
+
|
| 11 |
+
How should we transfer knowledge and capabilities across trained models? One popular approach is transfer learning [44], which fine-tunes a pre-trained model on a target task through additional gradient-based training. The preparatory step of pre-training the model on a data-rich task ideally instills useful “knowledge” into the network’s parameters, which allows the model to learn more rapidly and effectively when fine-tuned on a downstream task of interest. Transfer learning has therefore become a particularly important and omnipresent tool across many fields, including natural language processing [57, 13, 9, 52, 53, 46] and computer vision [43, 24, 68]. Recently, it has been shown that training on an “intermediate” task between pre-training and fine-tuning can further boost performance through additional transfer of capabilities from the intermediate task [47, 60, 51, 48]. Alternatively, continued self-supervised training on unlabeled domain-specialized data can serve as a form of domain adaptation [19].
|
| 12 |
+
|
| 13 |
+
All of the aforementioned transfer learning methods transfer knowledge by using a trained network to initialize another network followed by iterative gradient descent. While demonstrably powerful, several drawbacks arise from this: First, improvements to ancestor models cannot be passed down to descendants; instead, we must restart the whole process from the improved ancestor model, throwing away our previous work. For example, if we fine-tune a pre-trained model on a downstream task, but then the pre-trained model is improved through additional training, we must re-fine-tune the new model on our downstream task if we want to confer benefits from this additional pre-training. Furthermore, if we gain access to a checkpoint that has been fine-tuned on a useful intermediate task, we must again throw away our previous work and fine-tune from the intermediate task checkpoint. Existing methods for transfer learning also have the disadvantage of only being able to transfer information from a single model. While it may be possible to train on multiple intermediate tasks sequentially, one quickly either runs into a combinatorial explosion of saved checkpoints or faces the issue of “catastrophic forgetting” in continual learning [28]. In addition to slowing down experimentation by preventing reuse of work, these drawbacks impose limitations on the types of transfer that can occur.
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
Figure 1: Merging patterns considered in this work. Left: Merging many fine-tuned models as a form of ensembling. Center, top: “Robust fine-tuning” [66] , where a fine-tuned model is merged with the pre-trained model to improve performance on the original pre-training task. Center, bottom: Merging a fine-tuned model with a “donor” task, analogous to intermediate-task transfer learning [47, 51]. Right: Merging an intermediate-task trained model with a donor model.
|
| 17 |
+
|
| 18 |
+
A less common way of transferring capabilities across models is to simply average their parameters. This procedure, which we call “merging”, is generally only feasible when the models being averaged share a common architecture and initialization. Merging is the core component of the FedAvg algorithm used in Federated Learning [39], where updates to a shared model computed by individual workers that are training on different datasets are combined by simpling averaging the updates. Recently, Wortsman et al. [66] demonstrated that merging can be used to improve robustness to domain shift in fine-tuned models by averaging the parameters of the original pre-trained model with the fine-tuned parameters. Merging is also a common way of performing ensembling [49, 67], where the parameters of individual models trained on the same dataset are averaged to create a single performant model.
|
| 19 |
+
|
| 20 |
+
In this work, we view model merging as approximately maximizing the joint likelihood of the models’ posterior distribution over parameters. Since gradient-based maximum likelihood training only provides a point estimate of the posterior, some approximation of the posterior distribution is required. When an isotropic Gaussian distribution is used to approximate the posterior (with identity precision matrix and mean set to the model’s parameter values), we show that maximizing the joint likelihood across models is equivalent to simply averaging their parameters. We therefore refer to merging models by averaging parameters as isotropic merging. The view of merging as maximizing the joint likelihood of model posteriors suggests that using a better estimate of the posterior distribution may yield improved merging results. This leads us to introduce Fisher merging, which leverages the Laplace approximation by using the diagonal of each model’s Fisher information as the precision matrix for that model’s posterior.
|
| 21 |
+
|
| 22 |
+
Empirically, we demonstrate that merging models with Fisher merging outperforms isotropic merging in a variety of settings. We first focus on the existing applications of model ensembling [49, 67] and improving fine-tuned model robustness [66]. Then, we demonstrate for the first time that merging is a viable alternative to traditional gradient-based transfer learning. Specifically, we compare merging to intermediate-task transfer learning [47, 51] and domain-adaptive pre-training [19], finding that merging can achieve comparable performance at significantly lower cost. Additionally, we show that merging can provide an additional boost to models created via traditional intermediate-task training. This provides a concrete example of transfer that is fast and easy with merging but onerous or impossible to do with existing methods. Diagrams of the merging patterns we consider in this work are shown in fig. 1.
|
| 23 |
+
|
| 24 |
+
The rest of our paper is structured as follows: In section 2, we provide necessary background and detail our Fisher merging procedure. Section 3 provides experimental results on model ensembling, robust fine-tuning, intermediate-task training, and domain adaptation. We explore related works in section 4 and provide conclusions and thoughts on future work in section 5.
|
| 25 |
+
|
| 26 |
+
# 2 Weighted Parameter Averaging for Model Merging
|
| 27 |
+
|
| 28 |
+
Our focus is on procedures for model merging, i.e. averaging the parameters of models that share an architecture and initialization. In this section, we first frame the common practice of averaging together model parameters as approximately maximizing the joint likelihood of model posteriors. Specifically, we show that parameter averaging corresponds to using an isotropic Gaussian as the approximate posterior for each model. We then introduce Fisher merging, which uses the model’s diagonal Fisher information matrix as the precision matrix of the Gaussian approximate posterior. Fisher merging can be implemented by setting each merged parameter value to a weighted average of the corresponding parameter values from the original models, with the weighting for each parameter determined by its Fisher information. In addition, we add model-level weightings as additional hyperparameters to set the relative importance of each model.
|
| 29 |
+
|
| 30 |
+
# 2.1 Isotropic merging
|
| 31 |
+
|
| 32 |
+
Consider the problem setting where we have $M$ trained neural networks with parameters $\theta _ { 1 } , \dots , \theta _ { M }$ and our goal is to create a single neural network with parameters $\theta$ that, loosely speaking, inherits the capabilities of the $M$ trained neural networks. Assume that all of these neural networks share a common architecture and had the same set of initial parameter values before being trained. Merging attacks this problem by finding the parameters $\theta$ that maximize the joint likelihood of the posterior distributions of the $M$ models. Unfortunately, typical neural network training procedures do not provide access to a posterior distribution, which necessitates approximation. If the posterior of each model is approximated via an isotropic Gaussian with mean set to the model’s parameters, the optimization problem can be written as $\begin{array} { r } { \theta ^ { * } = \operatorname * { a r g m a x } _ { \theta } \sum _ { i } \log p ( \theta | \theta _ { i } , I ) } \end{array}$ where $p ( \theta | \theta _ { i } , I )$ is the probability distribution of the aforementioned approximate isotropic Gaussian posterior distribution used for model $i$ and $I$ is the identity matrix. This optimization problem has a closed-form solution given by $\begin{array} { r } { \theta ^ { * } = \frac { 1 } { M } \sum _ { i } \theta _ { i } } \end{array}$ , i.e. an average of the model parameters. Such an averaging procedure has been used in past work aiming to combine model capabilities, e.g. in federated learning [39], model ensembling [49, 67], and robust fine-tuning [66].
|
| 33 |
+
|
| 34 |
+
# 2.2 Per-model weights
|
| 35 |
+
|
| 36 |
+
In this work, we additionally introduce model-specific scalar hyperparameters $\lambda _ { i } , i \in \{ 1 , \dots , M \}$ into the model merging framework described above. Specifically, we change the optimization problem to $\begin{array} { r } { \theta ^ { * } = \mathrm { a r g m a x } _ { \theta } \sum _ { i } \bar { \lambda } _ { i } \log p ( \underline { { \theta } } | \theta _ { i } , I ) } \end{array}$ where $\begin{array} { r } { \lambda _ { i } \geq 0 , \bar { \sum _ { i } } \lambda _ { i } = \bar { 1 } } \end{array}$ . In the case of isotropic merging, this changes the solution to $\theta ^ { * } = \textstyle \sum _ { i } \lambda _ { i } \theta _ { i }$ , These hyperparameters provide control over the importance assigned to each of the models that are being merged. For example, when using merging to perform ensembling we might expect each model to be equally important and therefore set $\lambda _ { i } = 1 / M$ for all $i$ On the other hand, when mimicking the setup of intermediate-task training where the capabilities of a “donor” model are used to improve performance of a recipient model, we might weigh the recipient model more highly. Wortsman et al. [66] introduce a similar hyperparameter $\alpha$ when averaging the parameters of two models and report results for varying values of $\alpha$ .
|
| 37 |
+
|
| 38 |
+
# 2.3 Laplace Approximation
|
| 39 |
+
|
| 40 |
+
Framing merging as approximate maximization of the joint posterior likelihood reveals that simple parameter averaging is implicitly using an isotropic Gaussian posterior approximation. Such an approximation may be overly simplistic and lead to degraded performance. To explore improved merging procedures, we consider improved methods for creating an approximate posterior from a point estimate. Specifically, we use the Laplace approximation to the posterior, which corresponds to a second-order Taylor expansion of the log density around a mode [36, 10]. This leads to a Gaussian approximation $\mathcal { N } ( \theta , H ^ { - 1 } )$ of the posterior, where $H$ is the Hessian matrix and $\theta$ are the model’s trained parameter values. More precisely, we assume that the parameter values $\theta$ of a trained neural network are a local maximum of the posterior. It can then be shown that the precision matrix of the Laplace approximation is given by the Fisher information matrix of the network at $\theta$ .
|
| 41 |
+
|
| 42 |
+
The Fisher information matrix $F _ { \theta }$ [16, 3] of a neural network $p _ { \theta } ( y | x )$ trained to predict an output $y$ from input data $x$ is a $\left| \theta \right| \times \left| \theta \right|$ positive semidefinite matrix given by the formula
|
| 43 |
+
|
| 44 |
+
$$
|
| 45 |
+
F _ { \theta } = \mathbb { E } _ { x } \left[ \underset { y \sim p _ { \theta } ( y | x ) } { \mathbb { E } } \nabla _ { \theta } \log p _ { \theta } ( y | x ) \nabla _ { \theta } \log p _ { \theta } ( y | x ) ^ { T } \right] .
|
| 46 |
+
$$
|
| 47 |
+
|
| 48 |
+
It can be shown that the Fisher information matrix coincides with the Hessian $H$ at modes of the distribution [45], explaining its use in the Laplace approximation. The Fisher information matrix $F _ { \theta }$ can also be used to relate changes in the model parameters to changes in the model output by noting that $\begin{array} { r } { \mathbb { E } _ { x } \left[ D _ { \mathrm { K L } } ( p _ { \theta } ( y | x ) | | p _ { \theta + \delta } ( y | x ) ) \right] \approx \frac { 1 } { 2 } \delta ^ { T } F _ { \theta } \delta } \end{array}$ as $\delta 0$ , where $D _ { \mathrm { K L } }$ denotes the KL-divergence [45].
|
| 49 |
+
|
| 50 |
+
As the full Fisher matrix takes $O ( | \theta | ^ { 2 } )$ memory to store, it quickly becomes impractical for all but the smallest models. We are thus forced to use an approximation to the full Fisher in practice. In this paper, we follow the common practice of using the diagonal of the Fisher matrix [28]. While other methods (e.g. [1]) exist for estimating the Fisher, we leave their exploration for future work. In our experiments, we estimated the diagonal of the Fisher matrix via
|
| 51 |
+
|
| 52 |
+
$$
|
| 53 |
+
\hat { F } _ { \theta } = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \underset { y \sim p _ { \theta } ( y | x _ { i } ) } { \mathbb { E } } ( \nabla _ { \theta } \log p _ { \theta } ( y | x _ { i } ) ) ^ { 2 } ,
|
| 54 |
+
$$
|
| 55 |
+
|
| 56 |
+
where $x _ { 1 } , \ldots , x _ { N }$ are drawn i.i.d. from the dataset that was used to train the model. The expectation over $y$ can be estimated via sampling from $p _ { \theta } ( y | x _ { i } )$ or computed exactly when the number of classes is small. We note that computing the Fisher requires $N$ per-example gradients, which can be straightforwardly computed for neural networks using backpropagation. This makes computing the diagonal Fisher have roughly the same computational cost as training on $N$ examples.
|
| 57 |
+
|
| 58 |
+
# 2.4 Fisher Merging
|
| 59 |
+
|
| 60 |
+
Having noted that the Laplace approximation provides a tractable way to obtain a better approximation to the posterior, we now use it to create an improved merging procedure that we call Fisher merging. Letting $F _ { 1 } , \ldots , F _ { M }$ correspond to the diagonal approximate Fisher matrices, we construct $p ( \boldsymbol { \theta } | \boldsymbol { \theta } _ { i } , F _ { i } )$ as a Gaussian-distributed posterior over the parameters of the merged model with mean $\theta _ { i }$ and precision $F _ { i }$ . To obtain the merged model, we find a single set of parameters that is given a high probability under all posteriors. Formally, we have
|
| 61 |
+
|
| 62 |
+
$$
|
| 63 |
+
\theta ^ { * } = \operatorname { a r g m a x } _ { \theta } \sum _ { i = 1 } ^ { M } \lambda _ { i } \log p ( \theta | \theta _ { i } , F _ { i } ) ,
|
| 64 |
+
$$
|
| 65 |
+
|
| 66 |
+
which has the closed-form solution
|
| 67 |
+
|
| 68 |
+
$$
|
| 69 |
+
\theta ^ { * ( j ) } = \frac { \sum _ { i = 1 } ^ { M } \lambda _ { i } F _ { i } ^ { ( j ) } \theta _ { i } ^ { ( j ) } } { \sum _ { i = 1 } ^ { M } \lambda _ { i } F _ { i } ^ { ( j ) } } ,
|
| 70 |
+
$$
|
| 71 |
+
|
| 72 |
+
where $j = 1 , \ldots , | \theta |$ . Intuitively, we can think of Fisher merging as computing a weighted average of the parameter values in each model where the weighting is done according to each parameter’s Fisher information. Since the Fisher information is a local property of a single parameter value, Fisher merging might be less performant when applied to models whose parameters are far apart in parameter space. We therefore limit our focus to models that were trained from the same initialization.
|
| 73 |
+
|
| 74 |
+
Numerical Issues. Note that (4) can run into numerical issues when the Fisher is close to zero across all models for a given parameter. In practice, we choose a privileged “target model” in all of our experiments and “default” to the parameter’s value in the target model in these cases. An alternative would be to take an average weighted only by the merging coefficients (i.e., pretend the Fisher is the same across all models). In practice, the choice of a “default” value for these parameters had little impact on performance (likely because a small Fisher value implies that changing the parameter has a minute effect on the model’s outputs and is therefore relatively unimportant to the model’s behavior).
|
| 75 |
+
|
| 76 |
+
Unmergeable Parameters. In many cases, we have some parameters from each model that do not appear in all of the models we are merging. For example, this includes having task-specific classification heads on top of a common body architecture. We handle this by only applying the merging procedure (3) to the shared body parameters and keeping the task-specific heads unchanged. Although this may lead to a distribution shift in the classification head inputs, we found it to work well in practice for the datasets and tasks we consider.
|
| 77 |
+
|
| 78 |
+
# 3 Experiments
|
| 79 |
+
|
| 80 |
+
Our first experimental goal is to validate that our use of an improved estimate of the posterior yields improved merging performance. To test this hypothesis, we apply Fisher merging to two settings where isotropic merging has already proven successful: Model ensembling [49, 67] and robust fine-tuning [66]. Then, we demonstrate that Fisher merging provides a cheap and effective alternative to traditional transfer learning pipelines by validating its performance in intermediate-task transfer learning [47, 51] and domain-adaptive pre-training [19]. Finally, we demonstrate that merging opens up new paths of transferring capabilities across models by demonstrating a boost in performance when merging an intermediate task-trained model with different donor models.
|
| 81 |
+
|
| 82 |
+
# 3.1 Ensembling
|
| 83 |
+
|
| 84 |
+
An existing application of isotropic merging is for ensembling, i.e. combining models trained on the same dataset to obtain better predictions. Ensembling is most commonly performed by averaging the predictions of the individual models. This form of ensembling requires computing the output of all $M$ models in the ensemble, thereby increasing the computational cost by a factor of $M$ compared to computing the output for a single model. A cheaper alternative is to average the parameters of the models themselves. This approach is diagrammed in fig. 1, left. Such an approach is used in the classical method of Polyak averaging [49], where parameter values from the final $M$ iterations of training are averaged. More recently, Wortsman et al. [67] introduced the “Model Soup” approach where fine-tuned models with different hyperparameter settings are averaged to improve performance. To the best of our knowledge, all parameter-averaging ensemble methods have used isotropic merging, i.e. an unweighted average.
|
| 85 |
+
|
| 86 |
+
To test whether Fisher merging provides a boost over isotropic merging when averaging parameters for ensembling, we consider ensembling fine-tuned checkpoints derived from the same pre-trained model. Specifically, we consider the BERT-Base model [13] fine-tuned on the RTE [8], MRPC [14], and SST-2 [59] datasets. For each dataset, we use five fine-tuned checkpoints downloaded from the Hugging Face model hub.2 These checkpoints were fine-tuned with a variety of hyperparameter settings that were not chosen by us, so our experimental setting most closely matches the “Model Soup” approach [67]. A list of the checkpoints used is available in appendix A. Since we do not anticipate that any member of the ensemble should be given a larger weight, we set $\lambda _ { i } = 1 / 5$ for all models.
|
| 87 |
+
|
| 88 |
+
Our results are shown in fig. 2. We report validation set scores for Fisher merging, isotropic merging, and prediction ensembling (specifically, averaging the output probabilties of all models). Fisher merging significantly outperforms isotropic merging in all cases and attains comparable performance to prediction ensembling. Notably, performing inference after merging is $M \times$ cheaper than prediction ensembling, suggesting that merging can provide a cheaper alternative to standard ensembling procedures.
|
| 89 |
+
|
| 90 |
+
# 3.2 Robust Fine-Tuning
|
| 91 |
+
|
| 92 |
+
Recently, Wortsman et al. [66] found that while fine-tuning a pre-trained vision model tends to improve performance on the downstream task, it also tends to decreases accuracy on the original pretraining task. They therefore propose a “robust fine-tuning” procedure called WiSE-FT that computes a weighted average of the original pre-trained parameters and the fine-tuned parameters. Different weighting values produce different trade-offs between pre-training and fine-tuning task performance. In some cases, robust fine-tuning can even improve performance on the original pre-training task without sacrificing performance on the downstream fine-tuning task relative to traditional fine-tuning.
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
Figure 2: Validation set accuracy for ensembles of five fine-tuned BERT models using different ensembling methods on the RTE, MRPC, and SST-2 datasets. Fisher merging produces a single model that performs comparably to output ensembling while being $5 \times$ cheaper.
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
Figure 3: IID (ImageNet) and average outof-domain (OOD) accuracy across five OOD datasets when using the WiSE-FT procedure [66] with either Fisher or isotropic merging. Dark to light color indicates increasing $\lambda _ { 1 }$ from 0 to 1.
|
| 99 |
+
|
| 100 |
+
This procedure implicitly uses isotropic merging and therefore provides another natural testbed for determining whether Fisher merging provides a boost in performance. A schematic of robust fine-tuning is shown in fig. 1, center top.
|
| 101 |
+
|
| 102 |
+
We use the codebase and experimental setup of Wortsman et al. [66] exactly, simply replacing isotropic merging with Fisher merging. For full details of this setup, we refer to Wortsman et al. [66]. As a short summary, we apply WiSE-FT to the ImageNet [11, 58] pre-trained ViT-B/16 model [15] on five out-of-domain (OOD) datasets: ImageNet-A [21], ImageNet-R [20], ImageNet Sketch [62], ImageNet V2 [56], and ObjectNet [4]. Following Wortsman et al. [66], we measure IID (ImageNet) and OOD performance when averaging together the original pre-trained model parameters and parameters from models fine-tuned on each of the OOD datasets, varying $\lambda _ { 1 }$ (the averaging weight for the pre-trained model, called $\alpha$ by Wortsman et al. [66]) from 0 to 1 in 0.1-step increments (with $\lambda _ { 2 } = 1 - \lambda _ { 1 }$ correspondingly decreasing from 1 to 0). To determine whether Fisher merging confers a boost in performance, we compare parameter averaging using either isotropic or Fisher merging.
|
| 103 |
+
|
| 104 |
+
We plot the IID (ImageNet) accuracy against the average accuracy on the five OOD datasets for varying values of $\lambda _ { 1 }$ in fig. 3, with plots for individual OOD datasets in fig. 7 (appendix). Fisher merging produces a significantly better trade-off between IID and OOD accuracy. In particular, Fisher merging seems to general improve IID accuracy compared to isotropic merging. For example, for the value of $\lambda _ { 1 }$ producing the best average OOD accuracy, Fisher merging produces about $1 \%$ higher IID accuracy than isotropic merging.
|
| 105 |
+
|
| 106 |
+
# 3.3 Intermediate-task training
|
| 107 |
+
|
| 108 |
+
Having established that Fisher merging produces better results than isotropic merging in settings where merging has been attempted before, we now explore the use of merging as an alternative to a gradient-based transfer learning procedure. Specifically, we explore intermediate-task training [47, 51], where a model is fine-tuned on an intermediate “donor” task before being trained on the target task of interest. To the best of our knowledge, no prior work has considered parameter averaging as a way of performing intermediate-task transfer learning. For the most part, intermediate-task training has mainly been considered in the NLP domain; as such, we limit our experiments to the BERT [13] and RoBERTa [33] pre-trained language models. To enable comparison to past work, we mostly explored merging pairs of models but we are interested in exploring merging more than two models in future work. As in section 3.1, we made use of fine-tuned BERT and RoBERTa checkpoints from the Hugging Face repository [65].
|
| 109 |
+
|
| 110 |
+
Following previous work [47, 51], we first ran experiments using BERT-base on the GLUE benchmark [61]. The GLUE benchmark consists of the sentence acceptability task CoLA [64], the sentiment detection task SST-2 [59], the paraphrase detection tasks MRPC and QQP [14, 23], the sentence similarity task STS-B [7], and the natural language inference (NLI) tasks MNLI, QNLI, RTE, and
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
Figure 4: Validation set accuracy on RTE when performing intermediate-task training with datasets from GLUE as the donor task. Dashed line denotes RTE accuracy without intermediatetask training.
|
| 114 |
+
|
| 115 |
+
|
| 116 |
+
Figure 5: Validation accuracy on RTE after first fine-tuning on MNLI, then fine-tuning on RTE, and finally Fisher merging with various donor task models. Dashed line denotes RTE accuracy after MNLI intermediate-task training.
|
| 117 |
+
|
| 118 |
+
WNLI [6, 54, 8, 31]. All of the GLUE tasks are classification tasks except for STS-B, which is a regression task with a score ranging from 0 to 5. To simplify computation of the Fisher, we turn STS-B into a classification task by partitioning the continuous label into 25 equally-sized buckets [53]. Following common practice, we do not run experiments on WNLI due to the tricks required to get a good score [12, 29]. See Wang et al. [61] for more details on these tasks and their associated metrics. We detail how we obtained fine-tuned checkpoints on these tasks in appendix C. We computed a diagonal Fisher approximation for each checkpoint using up to 4096 examples from the corresponding train set. Since it is not clear a priori what weighting coefficients $\lambda _ { i }$ to use in this setting, we chose $\lambda _ { i }$ by a grid search with 50 points, using the score on the first 2048 validation examples as the selection metric. We compare Fisher merging to isotropic merging as well as a standard gradient-based intermediate-task fine-tuning baseline [47]. A diagram of intermediate-task merging is shown in fig. 1, center bottom.
|
| 119 |
+
|
| 120 |
+
In initial experiments (reported in tables A1 to A3), we performed intermediate-task training for possible pair of datasets from the GLUE benchmark. Congruent with past work [47, 51, 60], we found that intermediate-task training provided the most notable performance boost when the RTE dataset was the target. We therefore focus on RTE results in the main text. Figure 4 shows the results of intermediate-task training of BERT-base with RTE as the target task and the other GLUE datasets as donor tasks, using Fisher merging, isotropic merging, or standard gradient-based training. Notably, performing gradient-based intermediate-task training hurts on some datasets, whereas merging always helps. Fisher merging gets comparable or better performance than isotropic merging with the largest gap observed when using MNLI as the intermediate task. On the other hand, merging performs worse than standard gradient-based training when using MNLI as the donor task.
|
| 121 |
+
|
| 122 |
+
Exploring new paths for transfer Given this performance gap, we were interested to see whether merging could provide an additional boost on top of gradient-based intermediate-task training. We therefore performed Fisher merging on a BERT-base model that was first fine-tuned on MNLI and then fine-tuned on RTE. A diagram of this setup is shown in fig. 1, right. This procedure does not have a direct analog in traditional gradient-based, and as we will show later, performing multi-stage gradient-based intermediate-task training generally harms results.
|
| 123 |
+
|
| 124 |
+
We consider Fisher merging the intermediate-task trained RTE model with all GLUE tasks and show the results in fig. 5. Fisher merging provides a boost over gradient-based intermediate-task training for all tasks. Interestingly, a boost is still conferred when merging with an MNLI-trained model, suggesting that merging provides a complementary path for transferring capabilities across models.
|
| 125 |
+
|
| 126 |
+
Scaling to RoBERTa-large Seeing that merging can provide a boost on top of intermediate-task training, we explored whether this boost could still be obtained for a stronger model than BERTbase. We therefore applied the same procedure to a RoBERTa-large RTE model that had been fine-tuned from an MNLI intermediate checkpoint. Our donor models were the original RoBERTalarge checkpoint (i.e., not fine-tuned on MNLI) fine-tuned on MRPC, RTE, STS-B, and SST-2. We additionally ran a sequential gradient-based fine-tuning baseline where we started with the MNLI checkpoint, fine-tuned on the donor task, and then fine-tuned on the target task.
|
| 127 |
+
|
| 128 |
+
The results are shown in fig. 6. We find merging provides a boost in performance even on the more performant RoBERTa model. The largest boost of 2.2 points came from Fisher merging with another RTE checkpoint, which is reminiscent of using merging for ensembling. Notably, including an additional intermediate task in gradient-based training significantly harmed performance compared to performing intermediate-task training on MNLI alone. We hypothesize this is related to the phenomena of catastrophic forgetting [17], where the model’s capabilities on MNLI are forgotten as it is trained on the next intermediate task. Nevertheless, this illustrates model merging’s ability to sidestep the issue of catastrophic forgetting and enable exploration of novel transfer strategies.
|
| 129 |
+
|
| 130 |
+
Costs We had previously noted that our merging procedure could potentially be substantially more efficient than standard gradient-based fine-tuning. To measure this claim concretely, we computed the FLOPs required for fine-tuning and merging an RTE checkpoint based on the heuristics described in Kaplan et al. [26]. Fine-tuning BERT-base on RTE for 10 epochs would require about 5.5e14 FLOPs. Our merging procedures require computing the merged checkpoint (eq. (4)) and then evaluating it on the validation set with Fisher merging also requiring the estimation of the Fisher matrix (eq. (2)) beforehand. These steps require about 4.0e8, 2.0e12, and 9.1e13 FLOPs respectively, resulting in a roughly $6 \times$ lower total cost compared to fine-tuning for Fisher merging and $2 7 5 \times$ lower cost for isotropic merging. We note that the Fisher matrix only needs to be computed once and can be reused for subsequent merges, which amortizes the most expensive step in Fisher merging.
|
| 131 |
+
|
| 132 |
+
To explore methods for further reducing costs, we experimented with using fewer examples to estimate the Fisher. Specifically, we experimented with intermediate-task Fisher merging of BERT-base with MNLI as the donor task and RTE as the target task. The results are shown in table A4. While using the full training set to estimate the Fisher produced the best performance $( 7 3 . 4 \% )$ , using only 256 examples to estimate the Fisher only produced a mild degradation in accuracy $( 7 2 . 7 \% )$ and still outperformed the isotropic merging baseline. This suggests that computing the Fisher over fewer examples could further reduce computational costs without sacrificing a great deal of accuracy.
|
| 133 |
+
|
| 134 |
+
# 3.4 Domain Adaptation
|
| 135 |
+
|
| 136 |
+
We now turn our attention to the “domain-adaptive pre-training” (DAPT) approach for domain adaptation advocated by Gururangan et al. [19], which is methodologically similar to intermediatetask training. DAPT consists of additional pre-training of an original general-purpose pre-trained checkpoint on domain-specific unlabeled data. We explore the benefits of merging in an experimental setup similar to Gururangan et al. [19]. We focus on the biomedical (BIOMED) and computer science (CS) domains because they correspond to the classification tasks that saw the largest gains from domain-adaptive pre-training in [19]. Namely, we experimented with the CHEMPROT [30] relation classification task on the BIOMED domain. On the CS domain, we used the citation intent task of ACL-ARC [25] and the relation classification task of SCIERC [35]. Following Gururangan et al. [19], we report macro- $F _ { 1 }$ for ACL-ARC and SCIERC, and we report micro- $F _ { 1 }$ for CHEMPROT. We used RoBERTa-base [33] as our baseline model. Appendix D includes full details of the pre-training, fine-tuning, and merging procedures used.
|
| 137 |
+
|
| 138 |
+
We present our results in table 1. Merging provided the largest boost on ACL-ARC, and outperformed traditional fine-tuning in this setting. We only observed a minor improvement in performance on CHEMPROT and SCIERC. We note that our boosts from gradient-based fine-tuning were smaller than reported in [19], which was likely because we were only able to train on public data and we applied domain-adaptive pre-training for fewer steps. However, our results are consistent in the sense that ACL-ARC received the largest boost and CHEMPROT received the smallest boost.
|
| 139 |
+
|
| 140 |
+
# 4 Related Work
|
| 141 |
+
|
| 142 |
+
Like our work, elastic weight consolidation (EWC) [28] uses the Laplace approximation to the posterior over model parameters to create a regularizer to prevent catastrophic forgetting in the context of continual learning. While their framework supports the use of posteriors from multiple
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
Figure 6: Validation accuracy on RTE using the setup of fig. 5, but with RoBERTa-large instead of BERT-base. “Standard training” fine-tunes on MNLI, then the donor task, then RTE. Dashed line denotes MNLI intermediate-task training.
|
| 146 |
+
|
| 147 |
+
<table><tr><td>Method</td><td>ChemProt</td><td>ACL-ARC</td><td>SciERC</td></tr><tr><td>Unmerged</td><td>82.70.3</td><td>70.53.2</td><td>81.00.4</td></tr><tr><td>Fisher</td><td>83.10.4</td><td>73.21.7</td><td>81.30.5</td></tr><tr><td>Isotropic</td><td>82.80.4</td><td>72.52.3</td><td>81.70.5</td></tr><tr><td>Fine-tuned</td><td>82.50.1</td><td>71.53.0</td><td>81.61.0</td></tr></table>
|
| 148 |
+
|
| 149 |
+
Table 1: Domain adaptation results. “Unmerged” refers to checkpoints fine-tuned from RoBERTabase. “Fisher” and “Isotropic” refer to the result of merging those checkpoints with the domainadaptive pre-trained (DAPT) checkpoint. “Finetuned” refers to models fine-tuned from the DAPT checkpoint. Subscripts provide the standard deviation across five trials.
|
| 150 |
+
|
| 151 |
+
models as well, they restrict such models to be previous checkpoints of a continually trained model. EWC keeps the model from losing previously acquired knowledge while merging provides a means of directly adding new knowledge to a model.
|
| 152 |
+
|
| 153 |
+
Some other existing procedures such as distillation [22] and ensembling [42] can also be thought of as combining or transferring knowledge between neural networks. However, those methods represent knowledge solely through the output of models. The knowledge contained within the parameters of a network will necessarily be greater than the knowledge contained in its output [2]. Hence, methods that directly combine model parameters such as merging have the potential to be more powerful than those methods. Furthermore, our merging procedure has an efficient and closed-form solution (eq. (4)) while distillation requires iterative gradient descent-based training.
|
| 154 |
+
|
| 155 |
+
Isotropic checkpoint averaging is used by federated learning [39] and Polyak averaging [49]. However, the checkpoints merged by those methods can be thought of coming from the same training run of single model. We believe we are the first to demonstrate cross-task transfer coming from checkpoint averaging and to explore it in the context of transfer learning. However, adapting ideas from federated learning such as [32, 63] could provide a fruitful avenue for future model merging research.
|
| 156 |
+
|
| 157 |
+
Natural gradient descent refers to an optimization procedure that uses KL-divergence of model predictions as a distance measure rather than the Euclidean distance in parameter space employed by regular gradient descent [3]. It does this by performing gradient descent on a Riemannian manifold with the Fisher information matrix as its metric [45]. In practice, this amounts to using the Fisher as a preconditioner during gradient descent. Some work on natural gradient descent may prove relevant for model merging such as using Kronecker-factorized Fisher matrices as an alternative to the diagonal approximation employed in this paper [37, 18, 38]. More broadly, in the field of information geometry the Fisher information matrix plays the role of a metric on a Riemannian manifold [40]. This has led to explorations of model averaging using tools from information geometry, e.g. [5, 41, 50].
|
| 158 |
+
|
| 159 |
+
# 5 Conclusion
|
| 160 |
+
|
| 161 |
+
In this paper, we introduced Fisher merging, a way to combine the capabilities of different models by computing a weighted average of their parameters. Fisher merging is motivated by a novel perspective of parameter averaging as maximizing the joint likelihood of model posteriors. Through extensive experiments, we demonstrated that using the Fisher information as a weight on the contribution of each parameter outperforms using an unweighted average. Furthermore, we showed that Fisher merging attains comparable and sometimes better performance than traditional gradient-based transfer learning methods at significantly lower costs. Our experiments also demonstrated various merging strategies that would be onerous with traditional gradient-based training, which opens up new avenues for transferring capabilities across models. In future work, we plan to investigate different methods for approximating the Fisher information and model posteriors as well as more esoteric combinations of models.
|
| 162 |
+
|
| 163 |
+
# Acknowledgments and Disclosure of Funding
|
| 164 |
+
|
| 165 |
+
This work was supported by the NSF CAREER award under Grant No. 2145822.
|
| 166 |
+
|
| 167 |
+
# References
|
| 168 |
+
|
| 169 |
+
[1] Achille, A., Lam, M., Tewari, R., Ravichandran, A., Maji, S., Fowlkes, C. C., Soatto, S., and Perona, P. Task2vec: Task embedding for meta-learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6430–6439, 2019.
|
| 170 |
+
[2] Achille, A., Paolini, G., and Soatto, S. Where is the information in a deep neural network? arXiv preprint arXiv:1905.12213, 2019.
|
| 171 |
+
[3] Amari, S. Neural learning in structured parameter spaces-natural riemannian gradient. Advances in neural information processing systems, pp. 127–133, 1997.
|
| 172 |
+
[4] Barbu, A., Mayo, D., Alverio, J., Luo, W., Wang, C., Gutfreund, D., Tenenbaum, J., and Katz, B. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In Advances in Neural Information Processing Systems (NeurIPS), 2019.
|
| 173 |
+
[5] Bartels, L. M. Specification uncertainty and model averaging. American Journal of Political Science, 1997.
|
| 174 |
+
[6] Bowman, S. R., Angeli, G., Potts, C., and Manning, C. D. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326, 2015.
|
| 175 |
+
[7] Cer, D., Diab, M., Agirre, E., Lopez-Gazpio, I., and Specia, L. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055, 2017.
|
| 176 |
+
[8] Dagan, I., Glickman, O., and Magnini, B. The pascal recognising textual entailment challenge. In Machine Learning Challenges Workshop, pp. 177–190. Springer, 2005.
|
| 177 |
+
[9] Dai, A. M. and Le, Q. V. Semi-supervised sequence learning. In Advances in neural information processing systems, 2015.
|
| 178 |
+
[10] Daxberger, E., Kristiadi, A., Immer, A., Eschenhagen, R., Bauer, M., and Hennig, P. Laplace redux-effortless bayesian deep learning. Advances in Neural Information Processing Systems (NeurIPS), 2021.
|
| 179 |
+
[11] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. ImageNet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 2009.
|
| 180 |
+
[12] Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
|
| 181 |
+
[13] Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
|
| 182 |
+
[14] Dolan, W. B. and Brockett, C. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP2005), 2005.
|
| 183 |
+
[15] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.
|
| 184 |
+
[16] Fisher, R. A. On the mathematical foundations of theoretical statistics. Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character, 222(594-604):309–368, 1922.
|
| 185 |
+
[17] Goodfellow, I. J., Mirza, M., Xiao, D., Courville, A., and Bengio, Y. An empirical investigation of catastrophic forgetting in gradient-based neural networks. arXiv preprint arXiv:1312.6211, 2013.
|
| 186 |
+
[18] Grosse, R. and Martens, J. A kronecker-factored approximate fisher matrix for convolution layers. In International Conference on Machine Learning, pp. 573–582. PMLR, 2016.
|
| 187 |
+
[19] Gururangan, S., Marasovic, A., Swayamdipta, S., Lo, K., Beltagy, I., Downey, D., and Smith, ´ N. A. Don’t stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964, 2020.
|
| 188 |
+
[20] Hendrycks, D., Basart, S., Mu, N., Kadavath, S., Wang, F., Dorundo, E., Desai, R., Zhu, T., Parajuli, S., Guo, M., Song, D., Steinhardt, J., and Gilmer, J. The many faces of robustness: A critical analysis of out-of-distribution generalization. International Conference on Computer Vision (ICCV), 2021.
|
| 189 |
+
[21] Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., and Song, D. Natural adversarial examples. Conference on Computer Vision and Pattern Recognition (CVPR), 2021.
|
| 190 |
+
[22] Hinton, G., Vinyals, O., and Dean, J. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.
|
| 191 |
+
[23] Iyer, S., Dandekar, N., and Csernai, K. First quora dataset release: Question pairs, 2017. URL https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs.
|
| 192 |
+
[24] Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., and Darrell, T. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, 2014.
|
| 193 |
+
[25] Jurgens, D., Kumar, S., Hoover, R., McFarland, D., and Jurafsky, D. Measuring the evolution of a scientific field through citation frames. Transactions of the Association for Computational Linguistics, 6:391–406, 2018.
|
| 194 |
+
[26] Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020.
|
| 195 |
+
[27] Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
|
| 196 |
+
[28] Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521–3526, 2017.
|
| 197 |
+
[29] Kocijan, V., Cretu, A.-M., Camburu, O.-M., Yordanov, Y., and Lukasiewicz, T. A surprisingly robust trick for winograd schema challenge. arXiv preprint arXiv:1905.06290, 2019.
|
| 198 |
+
[30] Kringelum, J., Kjaerulff, S. K., Brunak, S., Lund, O., Oprea, T. I., and Taboureau, O. Chemprot3.0: a global chemical biology diseases mapping. Database, 2016, 2016.
|
| 199 |
+
[31] Levesque, H., Davis, E., and Morgenstern, L. The winograd schema challenge. In Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning, 2012.
|
| 200 |
+
[32] Liu, X., Masana, M., Herranz, L., Van de Weijer, J., Lopez, A. M., and Bagdanov, A. D. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In 2018 24th International Conference on Pattern Recognition (ICPR), pp. 2262–2268. IEEE, 2018.
|
| 201 |
+
[33] Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov, V. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019.
|
| 202 |
+
[34] Lo, K., Wang, L. L., Neumann, M., Kinney, R., and Weld, D. S. S2orc: The semantic scholar open research corpus. arXiv preprint arXiv:1911.02782, 2019.
|
| 203 |
+
[35] Luan, Y., He, L., Ostendorf, M., and Hajishirzi, H. Multi-task identification of entities, relations, and coreference for scientific knowledge graph construction. arXiv preprint arXiv:1808.09602, 2018.
|
| 204 |
+
[36] MacKay, D. J. A practical bayesian framework for backpropagation networks. Neural computation, 4(3):448–472, 1992.
|
| 205 |
+
[37] Martens, J. and Grosse, R. Optimizing neural networks with kronecker-factored approximate curvature. In International conference on machine learning, pp. 2408–2417. PMLR, 2015.
|
| 206 |
+
[38] Martens, J., Ba, J., and Johnson, M. Kronecker-factored curvature approximations for recurrent neural networks. In International Conference on Learning Representations, 2018.
|
| 207 |
+
[39] McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communicationefficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pp. 1273–1282. PMLR, 2017.
|
| 208 |
+
[40] Nielsen, F. An elementary introduction to information geometry. Entropy, 22(10):1100, 2020.
|
| 209 |
+
[41] Nielsen, F. and Nock, R. Sided and symmetrized bregman centroids. IEEE transactions on Information Theory, 55(6), 2009.
|
| 210 |
+
[42] Opitz, D. W. and Shavlik, J. W. Actively searching for an effective neural network ensemble. Connection Science, 8(3-4):337–354, 1996.
|
| 211 |
+
[43] Oquab, M., Bottou, L., Laptev, I., and Sivic, J. Learning and transferring mid-level image representations using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2014.
|
| 212 |
+
[44] Pan, S. J. and Yang, Q. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345–1359, 2009.
|
| 213 |
+
[45] Pascanu, R. and Bengio, Y. Revisiting natural gradient for deep networks. arXiv preprint arXiv:1301.3584, 2013.
|
| 214 |
+
[46] Peters, M. E., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., and Zettlemoyer, L. Deep contextualized word representations. arXiv preprint arXiv:1802.05365, 2018.
|
| 215 |
+
[47] Phang, J., Févry, T., and Bowman, S. R. Sentence encoders on stilts: Supplementary training on intermediate labeled-data tasks. arXiv preprint arXiv:1811.01088, 2018.
|
| 216 |
+
[48] Phang, J., Htut, P. M., Pruksachatkun, Y., Liu, H., Vania, C., Kann, K., Calixto, I., and Bowman, S. R. English intermediate-task training improves zero-shot cross-lingual transfer too. arXiv preprint arXiv:2005.13013, 2020.
|
| 217 |
+
[49] Polyak, B. T. and Juditsky, A. B. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 30(4):838–855, 1992.
|
| 218 |
+
[50] Posada, D. and Buckley, T. R. Model selection and model averaging in phylogenetics: advantages of akaike information criterion and bayesian approaches over likelihood ratio tests. Systematic biology, 53(5), 2004.
|
| 219 |
+
[51] Pruksachatkun, Y., Phang, J., Liu, H., Htut, P. M., Zhang, X., Pang, R. Y., Vania, C., Kann, K., and Bowman, S. R. Intermediate-task transfer learning with pretrained models for natural language understanding: When and why does it work? arXiv preprint arXiv:2005.00628, 2020.
|
| 220 |
+
[52] Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. Improving language understanding by generative pre-training, 2018.
|
| 221 |
+
[53] Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P. J. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019.
|
| 222 |
+
[54] Rajpurkar, P., Zhang, J., Lopyrev, K., and Liang, P. Squad: $1 0 0 { , } 0 0 0 { + }$ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250, 2016.
|
| 223 |
+
[55] Rajpurkar, P., Jia, R., and Liang, P. Know what you don’t know: Unanswerable questions for squad. arXiv preprint arXiv:1806.03822, 2018.
|
| 224 |
+
[56] Recht, B., Roelofs, R., Schmidt, L., and Shankar, V. Do ImageNet classifiers generalize to ImageNet? In International Conference on Machine Learning (ICML), 2019.
|
| 225 |
+
[57] Ruder, S., Peters, M. E., Swayamdipta, S., and Wolf, T. Transfer learning in natural language processing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Tutorials, pp. 15–18, 2019.
|
| 226 |
+
[58] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. ImageNet large scale visual recognition challenge. International journal of computer vision, 2015.
|
| 227 |
+
[59] Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A. Y., and Potts, C. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pp. 1631–1642, 2013.
|
| 228 |
+
[60] Vu, T., Wang, T., Munkhdalai, T., Sordoni, A., Trischler, A., Mattarella-Micke, A., Maji, S., and Iyyer, M. Exploring and predicting transferability across nlp tasks. arXiv preprint arXiv:2005.00770, 2020.
|
| 229 |
+
[61] Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., and Bowman, S. R. Glue: A multitask benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018.
|
| 230 |
+
[62] Wang, H., Ge, S., Lipton, Z., and Xing, E. P. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems (NeurIPS), 2019.
|
| 231 |
+
[63] Wang, H., Yurochkin, M., Sun, Y., Papailiopoulos, D., and Khazaeni, Y. Federated learning with matched averaging. arXiv preprint arXiv:2002.06440, 2020.
|
| 232 |
+
[64] Warstadt, A., Singh, A., and Bowman, S. R. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625–641, 2019.
|
| 233 |
+
[65] Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019.
|
| 234 |
+
[66] Wortsman, M., Ilharco, G., Li, M., Kim, J. W., Hajishirzi, H., Farhadi, A., Namkoong, H., and Schmidt, L. Robust fine-tuning of zero-shot models. arXiv preprint arXiv:2109.01903, 2021.
|
| 235 |
+
[67] Wortsman, M., Ilharco, G., Gadre, S. Y., Roelofs, R., Gontijo-Lopes, R., Morcos, A. S., Namkoong, H., Farhadi, A., Carmon, Y., Kornblith, S., and Schmidt, L. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. arXiv preprint arXiv:2203.05482, 2022.
|
| 236 |
+
[68] Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. How transferable are features in deep neural networks? In Advances in neural information processing systems, 2014.
|
| 237 |
+
|
| 238 |
+
# Checklist
|
| 239 |
+
|
| 240 |
+
1. For all authors...
|
| 241 |
+
|
| 242 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 243 |
+
(b) Did you describe the limitations of your work? [Yes] Settings where merging underperforms traditional gradient-based training are covered in section 3 and also discuss caveats of our method in section 2.4.
|
| 244 |
+
(c) Did you discuss any potential negative societal impacts of your work? [No]
|
| 245 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 246 |
+
|
| 247 |
+
2. If you are including theoretical results...
|
| 248 |
+
|
| 249 |
+
(a) Did you state the full set of assumptions of all theoretical results? [Yes] We state that merging assumes a shared architecture and initialization at various points in the paper and explain why in section 2.4.
|
| 250 |
+
(b) Did you include complete proofs of all theoretical results? [N/A] No theoretical advances were sufficiently complex to warrant proof.
|
| 251 |
+
|
| 252 |
+
3. If you ran experiments...
|
| 253 |
+
|
| 254 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] Code is provided in the supplementary. All datasets are public and are downloadable with our code.
|
| 255 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See appendix C and appendix D
|
| 256 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See tables A1 to A3.
|
| 257 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]
|
| 258 |
+
|
| 259 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 260 |
+
|
| 261 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 262 |
+
(b) Did you mention the license of the assets? [No] All datasets are widely-used (hundreds or thousands of citations) public datasets.
|
| 263 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We include code in the supplemental.
|
| 264 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
|
| 265 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
|
| 266 |
+
|
| 267 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 268 |
+
|
| 269 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
|
| 270 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 271 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
|
parse/dev/LSKlp_aceOC/LSKlp_aceOC_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Loek7hfb46P/Loek7hfb46P.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Loek7hfb46P/Loek7hfb46P_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Loek7hfb46P/Loek7hfb46P_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Loek7hfb46P/Loek7hfb46P_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/MbCAOMGsZXC/MbCAOMGsZXC.md
ADDED
|
@@ -0,0 +1,239 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training
|
| 2 |
+
|
| 3 |
+
Renrui Zhang1,2, Ziyu Guo2, Rongyao Fang1, Bin Zhao2, Dong Wang2, Yu Qiao2, Hongsheng $\mathbf { L i ^ { 1 , 3 } }$ , Peng GaoB2
|
| 4 |
+
|
| 5 |
+
1 CUHK-SenseTime Joint Laboratory, The Chinese University of Hong Kong, 2 Shanghai AI Laboratory, 3 Centre for Perceptual and Interactive Intelligence Limited
|
| 6 |
+
|
| 7 |
+
{zhangrenrui, gaopeng}@pjlab.org.cnhsli@ee.cuhk.edu.hk
|
| 8 |
+
|
| 9 |
+
# Abstract
|
| 10 |
+
|
| 11 |
+
Masked Autoencoders (MAE) have shown great potentials in self-supervised pretraining for language and 2D image transformers. However, it still remains an open question on how to exploit masked autoencoding for learning 3D representations of irregular point clouds. In this paper, we propose Point-M2AE, a strong Multi-scale MAE pre-training framework for hierarchical self-supervised learning of 3D point clouds. Unlike the standard transformer in MAE, we modify the encoder and decoder into pyramid architectures to progressively model spatial geometries and capture both fine-grained and high-level semantics of 3D shapes. For the encoder that downsamples point tokens by stages, we design a multi-scale masking strategy to generate consistent visible regions across scales, and adopt a local spatial self-attention mechanism during fine-tuning to focus on neighboring patterns. By multi-scale token propagation, the lightweight decoder gradually upsamples point tokens with complementary skip connections from the encoder, which further promotes the reconstruction from a global-to-local perspective. Extensive experiments demonstrate the state-of-the-art performance of Point-M2AE for 3D representation learning. With a frozen encoder after pretraining, Point-M2AE achieves $9 2 . 9 \%$ accuracy for linear SVM on ModelNet40, even surpassing some fully trained methods. By fine-tuning on downstream tasks, Point-M2AE achieves $8 6 . 4 3 \%$ accuracy on ScanObjectNN, $+ 3 . 3 6 \%$ to the secondbest, and largely benefits the few-shot classification, part segmentation and 3D object detection with the hierarchical pre-training scheme. Code is available at https://github.com/ZrrSkywalker/Point-M2AE.
|
| 12 |
+
|
| 13 |
+
# 1 Introduction
|
| 14 |
+
|
| 15 |
+
Learning to represent from unlabeled data without annotations, known as self-supervised learning, has attained great success in natural language processing [10, 32, 33, 5], computer vision [19, 7, 8, 18] and multi-modality learning [31, 50, 21]. By pre-training on the large-scale raw data, the networks are endowed with robust representation abilities and can significantly benefit downstream tasks with fine-tuning. Motivated by masked language modeling [32, 10], MAE [18] and some other methods [46, 53, 3] adopt asymmetric encoder-decoder transformers [13] to apply masked autoencoding for self-supervised learning on 2D images. They represent the input image as multiple local patches, and randomly mask them with a high ratio to build the pretext task for reconstruction. Specifically, the encoder aims at capturing high-level latent representations from limited visible patches, and the lightweight decoder is forced to reconstruct the RGB values of masked patches on top. Despite its superiority on grid-based 2D images, we ask the question: can MAE-style masked autoencoding be adapted to irregular point clouds as a powerful 3D representation learner?
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
Figure 1: Comparison of MAE (Top) and our Point-M2AE (Bottom). MAE for 2D image pretraining adopts standard transformer of the plain encoder and decoder, while Point-M2AE introduces a hierarchical transformer with skip connections for multi-scale point cloud pre-training.
|
| 19 |
+
|
| 20 |
+
To tackle this challenge, we propose Multi-scale Masked autoencoders for learning the hierarchical representations of point clouds via self-supervised pre-training, termed as Point-M2AE. We represent a point cloud as a set of point tokens depicting different spatial local regions, and inherit MAE’s pipeline to first encode visible point tokens and then reconstruct the masked 3D coordinates. Different from 2D images, masked autoencoding for 3D point clouds has three characteristics to be specially considered. Firstly, it is critical to understand the relations between local parts and the overall 3D shapes, which have strong geometric and semantic dependence. As examples, the network can recognize an airplane starting from its wing, or segment the wing’s part from the airplane’s global feature. Therefore, we regard the standard transformer with the plain encoder and decoder is sub-optimal for capturing such local-global spatial relations in 3D, which directly downsamples the input into a low-resolution representation as shown in Figure 1 (Top). We modify both the encoder and decoder into multi-stage hierarchies for progressively encoding multi-scale features of point clouds, constructing an asymmetric U-Net [35] like architecture in Figure 1 (Bottom). Secondly, as our Point-M2AE encodes multi-scale point clouds unlike the single-scale 2D images, the unmasked visible regions are required to be both block-wise within one scale and consistent across scales, which are respectively for reserving complete local geometries and ensuring coherent feature learning for the network. For this, we introduce a multi-scale masking strategy, which generates random masks at the final scale with a high ratio (e.g., $80 \%$ ), and back-projects the unmasked positions to all preceding scales. Thirdly, to better reconstruct 3D geometries from a local-to-global perspective, we utilize skip connections to complement the decoder with fine-grained information from the corresponding stages of the encoder. During fine-tuning on downstream tasks, we also adopt a local spatial self-attention mechanism with increasing attention scopes for point tokens at different stages of the encoder, which refocus each token within neighboring detailed structures.
|
| 21 |
+
|
| 22 |
+
By the multi-scale pre-training, Point-M2AE can encode point clouds from local-to-global hierarchies and then reconstructs the masked coordinates from global-to-local perspectives, which learns powerful 3D representations and performs superior transfer ability. After self-supervised pre-training on ShapeNet [6], Point-M2AE achieves $9 2 . 9 \%$ classification accuracy for linear SVM on ModelNet40 [44] with the frozen encoder, which surpasses the runner-up CrossPoint [2] by $+ 1 . 2 \%$ and even outperforms some fully supervised methods. By fine-tuning on various downstream tasks, Point-M2AE achieves $8 6 . 4 3 \%$ $( + 3 . 3 6 \% )$ accuracy on ScanObjectNN [38] and $9 4 . 0 \%$ $( + 0 . 8 \% )$ accuracy on ModelNet40 [44] for shape classification, $8 6 . 5 1 \%$ $( + 0 . 9 1 \% )$ instance mIoU on ShapeNetPart [48] for part segmentation, and $9 5 . 0 \%$ $( + 2 . 7 \% )$ accuracy on 10-way 20-shot ModelNet40 for few-shot classification. Our multi-scale masked autoencoding also benefits the 3D object detection on ScanNetV2 [9] by $+ 1 . 3 \%$ $\mathsf { A P } _ { 2 5 }$ and $+ 1 . 3 \%$ $\mathrm { { A P } _ { 5 0 } }$ , which provides the detection backbone with a hierarchical understanding of the point clouds.
|
| 23 |
+
|
| 24 |
+
We summarize the contributions of our paper as follows:
|
| 25 |
+
|
| 26 |
+
1. We propose Point-M2AE, a strong masked autoencoding framework, which conducts hierarchical point cloud encoding and reconstruction for better learning multi-scale spatial geometries of 3D shapes.
|
| 27 |
+
2. We introduce a U-Net like transformer architecture for MAE-style pre-training on point clouds, and adopt a multi-scale masking strategy to generate consistent visible regions across scales.
|
| 28 |
+
3. Point-M2AE achieves state-of-the-art performance for transfer learning on various downstream tasks, which indicates our approach to be a powerful representation learner for 3D point clouds.
|
| 29 |
+
|
| 30 |
+
# 2 Related Work
|
| 31 |
+
|
| 32 |
+
Pre-training by Masked Modeling. Compared to contrastive learning methods [19, 7, 8] that learn from inter-sample relations, self-supervised pre-training by masked autoencoding builds the pretext tasks to predict the masked parts of the input signals. The series of GPT [32, 33, 5] and BERT [11] apply masked modeling to natural language processing and achieve extraordinary performance boost on downstream tasks with fine-tuning. Inspired by this, BEiT [4] proposes to match image patches with discrete tokens via dVAE [34] and pre-train a standard vision transformer [13, 49] by masked image modeling. On top of that, MAE [18] directly reconstructs the raw pixel values of masked tokens and performs great efficiency with a high mask ratio. The follow-up works further improve the performance of MAE by momentum encoder [53], contrastive learning [3], and modified reconstruction targets [42]. For self-supervised pre-training on 3D point clouds, the masked autoencoding has not been widely adopted. Similar to BEiT, Point-BERT [49] utilizes dVAE to map 3D patches to tokens for masked point modeling, but heavily relies on constrastive learning [19], complicated data augmentation, and the costly two-stage pre-training. In contrast, our Point-M2AE is a pure masked autoencoding method of one-stage pre-training, and follows MAE to reconstruct the input signals without dVAE mapping. Different from previous MAE methods adopting standard plain transformer, we propose a hierarchical transformer architecture along with the multi-scale masking strategy to better learn a strong and generic representation for 3D point clouds.
|
| 33 |
+
|
| 34 |
+
Self-supervised Learning for Point Clouds. 3D representation learning without annotations has been widely studied in recent years. Mainstream methods mainly build the pretext tasks to reconstruct the transformed input point cloud based on the encoded latent vectors, such as rotation [28], deformation [1], rearranged parts [36] and occlusion [40]. From another perspective, PointContrast [45] utilizes contrastive learning between features of the same points from different views to learn discriminative 3D representations. DepthContrast [51] further extends the contrast for depth maps of different augmentations. CrossPoint [2] conducts cross-modality contrastive learning between point clouds and their corresponding rendering images to acquire rich self-supervised signals. Point-BERT [49] and Point-MAE [27] respectively introduce BERT-style [10] and MAE-style [18] pre-training schemes for 3D point clouds with standard transformer networks and performs competitively on various downstream tasks, but both of them can only encode point clouds with a single resolution and ignores the local-global relations between 3D shapes. In this paper, we propose Point-M2AE, an MAE-style framework with a hierarchical transformer for multi-scale point cloud pre-training. We achieve state-of-the-art downstream performance by learning the multi-scale representation of point clouds.
|
| 35 |
+
|
| 36 |
+
# 3 Method
|
| 37 |
+
|
| 38 |
+
The overall pipeline of Point-M2AE is shown in Figure 2, where we encode and reconstruct the point cloud by a hierarchical network architecture. In Section 3.1, We first introduce the masking strategy of Point-M2AE with multi-scale representations of point clouds. Then in Section 3.2 and Section 3.3, we present the details of our encoder and decoder with multi-stage hierarchies.
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
Figure 2: Overall pipeline of Point-M2AE. After the multi-scale masking, we embed point tokens at the 1-st scale and feed the visible ones into a hierarchical encoder-decoder transformer, which captures both high-level semantics and fine-grained patterns of the point cloud during pre-training.
|
| 42 |
+
|
| 43 |
+
# 3.1 Multi-scale Masking
|
| 44 |
+
|
| 45 |
+
To build a U-Net [35] like masked autoencoder for hierarchical learning, we encode the point cloud by $S$ scales with different number of points at each scale, and correspondingly modify the standard plain encoder into the $S$ -stage architecture. Following MAE, we embed the point cloud into discrete point tokens and randomly mask them for reconstruction. Importantly, for irregular-distributed points in the multi-scale architecture, the unmasked visible spatial regions are required to be consistent not only within one scale, but also across different scales. This is because the block-wise parts of 3D shapes tend to preserve more complete fine-grained geometries, and the unmasked positions are better to be shared across all scales for coherent feature learning of the encoder. Therefore, as shown in Figure 3, we first construct the $S$ -scale coordinate representations of the input point cloud and back-project the random masks from the final $S$ -th scale to the earlier scales to avoid fragmented visible parts.
|
| 46 |
+
|
| 47 |
+
$S$ -scale Representations. We denote the input point cloud as $P \in \mathbb { R } ^ { N \times 3 }$ and regard it as the 0-th scale. For the $i$ -th scale, $1 \leq i \leq S$ , we utilize Furthest Point Sampling (FPS) to downsample the points from the $( i - 1 )$ -th scale, which produces seed points $P _ { i } \in \mathbb { R } ^ { N _ { i } \times 3 }$ for scale $i$ of $N _ { i }$ points. Then, we adopt $k$ Nearest-Neighbour ( $k$ -NN) to aggregate the neighboring $k$ points for each seed point and obtain the neighbor indices $I _ { i } \in \mathbb { R } ^ { N _ { i } \times k }$ . By successively downsampling and grouping, we acquire the $S$ -scale representations $\{ P _ { i } , I _ { i } \} _ { i = 1 } ^ { S }$ of the input point cloud, where the number of points $N _ { i }$ gradually decreases and the inclusion relations between scales are recorded in $I _ { i }$ .
|
| 48 |
+
|
| 49 |
+
Back-projecting Visible Positions. For seed points $P _ { S }$ at the final $S$ -th scale, we randomly mask them with a large proportion (e.g., $80 \%$ ) and denote the remaining visible points as $P _ { S } ^ { v } \in \dot { \mathbb { R } } ^ { N _ { S } ^ { v } \times 3 }$ of $N _ { S }$ points. We then back-project the unmasked positions $P _ { S } ^ { v }$ to ensure the consistent visible regions across scales. For the $i$ -th scale, $1 \leq i < S$ , we retrieve all the $k$ nearest neighbors of $P _ { i + 1 } ^ { v }$ from the indices $I _ { i + 1 }$ to serve as the visible positions $P _ { i } ^ { v }$ , and mask the others. By recursively back-projecting, we obtain the visible and masked positions of all v m $S$ scales, denoted as $\{ P _ { i } ^ { v } , P _ { i } ^ { m } \} _ { i = 1 } ^ { S }$ , where $P _ { i } ^ { v } \in \mathbb { R } ^ { N _ { i } ^ { v } \times 3 }$ , $P _ { i } ^ { m } \in \mathbb { R } ^ { N _ { i } ^ { m } \times 3 }$ and $N _ { i } = N _ { i } ^ { v } + N _ { i } ^ { m }$ .
|
| 50 |
+
|
| 51 |
+
# 3.2 Hierarchical Encoder
|
| 52 |
+
|
| 53 |
+
Based on the multi-scale masking, we embed the initial tokens of visible points $P _ { 1 } ^ { v }$ for the 1-st scale and them into the hierarchical encoder with $S$ stages. Every stage is equipped with $K$ stacked encoder blocks, and each block contains a self-attention layer and a Feed Forward Network (FFN) of MLP layers. Between every two consecutive stages, we introduce spatial token merging modules to aggregate adjacent visible tokens and enlarge receptive fields for downsampling the point clouds.
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
Figure 3: Multi-scale masking strategy. To obtain a consistent visible regions across scales, we first represent the input point cloud by multi-scale coordinates and generate the random mask at the highest one. Then, we back-project the unmasked visible positions to all earlier scales.
|
| 57 |
+
|
| 58 |
+
Token Embedding and Merging. Indexed by $I _ { 1 }$ , we utilize a mini-PointNet [29] to extract and fuse the features of every seed point from $P _ { 1 } ^ { v } \in \mathbb { R } ^ { N _ { 1 } ^ { v } \times 3 }$ with its $k$ nearest neighbors. After that, we obtain the initial point tokens $T _ { 1 } ^ { v } \in \mathbb { R } ^ { N _ { 1 } ^ { v } \times C _ { 1 } }$ for the 1-st stage of the encoder, which embeds $N _ { 1 } ^ { e }$ local patterns of the 3D shape. Between the $( i - 1 )$ -th and $i$ -th stages, $1 < i \leq S$ , we merge $T _ { i - 1 } ^ { \dot { v } } \in \mathbb { R } ^ { \mathsf { \tilde { N } } _ { i - 1 } \times C _ { i - 1 } }$ to acquire the downsampled point tokens for the $i$ -th stage. We utilize MLP layers and a max pooling to integrate every $k$ tokens nearest to $P _ { i } ^ { v }$ indexed by $I _ { i }$ , which outputs $T _ { i } ^ { v } \in \mathbb { R } ^ { N _ { i } \times C _ { i } }$ . Due to our multi-scale masking, the merged $T _ { i } ^ { v }$ corresponds to the same visible parts of $T _ { i - 1 } ^ { v }$ , which enables the consistent feature encoding across different scales. For larger $i$ of deeper stages, we set higher feature dimension $C _ { i }$ to encode spatial geometries with richer semantics.
|
| 59 |
+
|
| 60 |
+
Local Spatial Self-Attention. During pre-training, we expect point tokens in the multi-stage encoder to capture global cues for 3D shapes, which benefits the reconstruction of masked parts. However, when fine-tuning on downstream tasks without masked autoencoding, point tokens in the shallower stages are better to mainly focus on local information and not to be disturbed by long-range signals, referring to the inductive bias of 3D locality [30]. Thus, during fine-tuning, we modify the original self-attention layer in the encoder with a local spatial constraint that only neighboring tokens within a ball query would be available for attention calculation. As the point tokens are downsampled by stages, we set increasing radii $\{ r _ { i } \} _ { i = 1 } ^ { S }$ of multi-scale ball queries for gradually expanding the attention scopes, which fulfills the local-to-global feature aggregation scheme.
|
| 61 |
+
|
| 62 |
+
# 3.3 Hierarchical Decoder
|
| 63 |
+
|
| 64 |
+
Via the hierarchical encoder, we obtain the encoded visible tokens $\{ T _ { i } ^ { v } \} _ { i = 1 } ^ { S }$ of all scales. Starting from the highest and concatenate $S$ -th scale, we assign a sharedem with the visible tokens arnable mask token to. We denote them as ked positions with coordin $P _ { S } ^ { m }$ $T _ { S } ^ { v }$ $\{ H _ { 1 } ^ { v } , H _ { 1 } ^ { m } \}$ $\{ P _ { S } ^ { v } , P _ { S } ^ { m } \}$ S , which serve as the input of the hierarchical decoder. We design the decoder to be lightweight with $S - 1$ stages and only one decoder block for each stage, which enforces the encoder to embed more semantics of the point clouds. Each decoder block consists of a vanilla self-attention layer and an FFN. We do not apply the local constraint to the attention in the decoder, since a global understanding between visible and mask tokens is crucial to the reconstruction.
|
| 65 |
+
|
| 66 |
+
Point Token Upsampling. We upsample the point tokens between stages to progressively recover the fine-grained geometries of 3D shapes before reconstruction. We regulate that the $j$ -th stage of the decoder corresponds to the $( S + 1 - j )$ -th stage of the encoder, both of which contain point tokens of the same $( S + 1 - j )$ -th scale with the feature dimension $C _ { S + 1 - j }$ . Between the $( j - 1 )$ - th and $j$ -th stage, $1 < j \le S - 1$ , we upsample the tokens $\{ H _ { j - 1 } ^ { v } , H _ { j - 1 } ^ { m } \}$ from the coordinates $\{ P _ { S + 2 - j } ^ { v } , P _ { S + 2 - j } ^ { m } \}$ into $\{ P _ { S + 1 - j } ^ { v } , P _ { S + 1 - j } ^ { m } \}$ via the token propagation module. Specifically, we obtain the $k$ nearest neighbors of each point token in $\{ H _ { j - 1 } ^ { v } , H _ { { \underline { { j } } } - 1 } ^ { m } \}$ indexed by $I _ { S + 2 - j }$ , and recover their neighbors’ features by weighted interpolation referring to PointNet+ $^ { \cdot + }$ [30], which generates the tokens $\{ \bar { H } _ { j } ^ { v } , H _ { j } ^ { m } \}$ of the $j$ -th stage.
|
| 67 |
+
|
| 68 |
+
Table 1: Linear evaluation on ModelNet40 [44] by SVM. We report different self-supervised learning methods and underline the second-best one.
|
| 69 |
+
|
| 70 |
+
<table><tr><td rowspan=1 colspan=2>Method Acc. (%)</td></tr><tr><td rowspan=1 colspan=1></td><td rowspan=3 colspan=1>3D-GAN [43] 83.3Latent-GAN [39] 85.7</td></tr><tr><td rowspan=5 colspan=1></td><td rowspan=5 colspan=1>Latent-GAN [39] 85.7SO-Net [22] 87.3FoldingNet [47] 88.4MAP-VAE[17] 88.4VIP-GAN[16] 90.2</td></tr><tr><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>87.3</td></tr><tr><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>88.4</td></tr><tr><td rowspan=4 colspan=1></td><td rowspan=2 colspan=1>DGCNN + Jiasaw [37] 90.6DGCNN + OcCo [40] 90.7DGCNN + CrossPoint [2] 91.2</td></tr><tr><td rowspan=1 colspan=1>90.7</td></tr><tr><td rowspan=1 colspan=1>Transformer + OcCo [49] 89.6</td></tr><tr><td rowspan=1 colspan=1>Point-BERT[49] 87.4</td></tr><tr><td rowspan=1 colspan=2>Point-M2AE 92.9</td></tr><tr><td rowspan=1 colspan=2>Improvement +1.7</td></tr></table>
|
| 71 |
+
|
| 72 |
+
Table 2: Shape classification on ModelNet40 [44]. ‘#points’ and ‘Acc.’ denote the number of points for training and the overall accuracy. [S] represents finetuning after self-supervised pre-training.
|
| 73 |
+
|
| 74 |
+
<table><tr><td>Method</td><td>#points</td><td>Acc. (%)</td></tr><tr><td>PointNet [29] PointNet++ [30]</td><td>1k</td><td>89.2</td></tr><tr><td></td><td>1k</td><td>90.5</td></tr><tr><td>PointCNN [23]</td><td>1k</td><td>92.2</td></tr><tr><td>[S] SO-Net [22] DGCNN [41]</td><td>5k</td><td>92.5</td></tr><tr><td>PCT[15]</td><td>1k</td><td>92.9</td></tr><tr><td>Point Transformer [52]</td><td>1k</td><td>93.2</td></tr><tr><td></td><td>-</td><td>93.7</td></tr><tr><td>Transformer [49]</td><td>1k</td><td>91.4</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>1k</td><td>92.1</td></tr><tr><td>[S] Point-BERT[49]</td><td>1k</td><td>93.2</td></tr><tr><td>[S] Point-BERT</td><td>4k</td><td>93.4</td></tr><tr><td>[S] Point-BERT</td><td>8k</td><td>93.8</td></tr><tr><td>[S] Point-M2AE</td><td>1k</td><td>94.0</td></tr></table>
|
| 75 |
+
|
| 76 |
+
Skip Connections. To further complement the fine-grained geometries, we channel-wisely concatenate the visible tokens $H _ { j } ^ { v } \in \mathbb { R } ^ { N _ { S + 1 - j } \times C _ { S + 1 - j } }$ of the decoder with $T _ { S + 1 - j } ^ { v } \in \mathbb { R } ^ { N _ { S + 1 - j } \times \check { C } _ { S + 1 - j } }$ from the corresponding $( S + 1 - j )$ -th stage of the encoder via skip connections, and adopt a linear projection layer to fuse their features. For the mask tokens $H _ { j } ^ { m }$ , we keep them unchanged, since the encoder only contains visible tokens without the masked ones.
|
| 77 |
+
|
| 78 |
+
Point Reconstruction. After $S - 1$ stages of the decoder, we acquire $\{ H _ { S - 1 } ^ { v } , H _ { S - 1 } ^ { m } \}$ with coordinates $\{ P _ { 2 } ^ { v } , P _ { 2 } ^ { m } \}$ and reconstruct the masked values from the mask tokens $H _ { S - 1 } ^ { m }$ . Other than predicting values at the 0-th scale of the input point cloud $P$ , we reconstruct the coordinates of $P _ { 1 } ^ { m }$ , namely, recovering the masked positions of the 1-st scale $P _ { 1 } ^ { m } \in \mathbb { R } ^ { N _ { 1 } ^ { m } \times 3 }$ from the 2-nd scale $\bar { P _ { 2 } ^ { m } } \in \mathbb { R } ^ { N _ { 2 } ^ { m } \times 3 }$ . This is because $\{ P _ { 1 } ^ { v } , P _ { 1 } ^ { m } \}$ of the 1-st scale could well represent the overall 3D shape and simultaneously preserve enough local patterns, which already constructs a comparatively challenging pretext task for pre-training. If we further upsample $\{ H _ { S - 1 } ^ { \bar { v } } , H _ { S - 1 } ^ { m } \}$ into $\{ \bar { H } _ { S } ^ { v } , H _ { S } ^ { m } \}$ and reconstruct the masked raw points from $P _ { 1 } ^ { m }$ , the extra spatial noises and computational overhead would adversely influence our performance and efficiency. Therefore, for every token in $H _ { S - 1 } ^ { m } \in \mathbb { R } ^ { N _ { 2 } ^ { m } \times C _ { 2 } }$ , we reconstruct its $k$ nearest neighbors recorded in $I _ { 2 }$ by a reconstruction head of one linear projection layer and compute the loss by $l _ { 2 }$ Chamfer Distance [14], formulated as,
|
| 79 |
+
|
| 80 |
+
$$
|
| 81 |
+
\begin{array} { r l } & { \widehat { P } _ { 2 \to 1 } ^ { m } = \mathrm { L i n e a r } ( H _ { S - 1 } ^ { m } ) , \mathrm { ~ w h e r e ~ } \widehat { P } _ { 2 \to 1 } ^ { m } \in \mathbb { R } ^ { N _ { 2 } ^ { m } \times k \times 3 } , } \\ & { \mathcal { L } _ { C D } = \mathrm { C h a m f e r D i s t a n c e } ( P _ { 2 \to 1 } ^ { m } , \widehat { P } _ { 2 \to 1 } ^ { m } ) , } \end{array}
|
| 82 |
+
$$
|
| 83 |
+
|
| 84 |
+
where $\widehat { P } _ { 2 1 } ^ { m }$ and $P _ { 2 1 } ^ { m }$ denote the predicted and ground-truth reconstruction coordinates from the 2-nd scale to the 1-st scale. We only utilize $\mathcal { L } _ { C D }$ for supervision without contrastive loss to conduct a pure masked autoencoding for self-supervised pre-training.
|
| 85 |
+
|
| 86 |
+
# 4 Experiments
|
| 87 |
+
|
| 88 |
+
In Section 4.1 and Section 4.2, we introduce the pre-training experiments of Point-M2AE and report the fine-tuning performance on various downstream tasks. We also conduct ablation studies in Section 4.3 to validate the effectiveness of our approach.
|
| 89 |
+
|
| 90 |
+
# 4.1 Self-supervised Pre-training
|
| 91 |
+
|
| 92 |
+
Settings. We pre-train our Point-M2AE on ShapeNet [6] dataset, which contains 57,448 synthetic 3D shapes of 55 categories. We set the stage number $S$ as 3, and construct a 3-stage encoder and a
|
| 93 |
+
|
| 94 |
+
Table 3: Shape classification on ScanObjectNN [38]. We report the accuracy $( \% )$ on the three splits of ScanObjectNN. [S] represents fine-tuning after self-supervised pre-training.
|
| 95 |
+
|
| 96 |
+
<table><tr><td>Method</td><td>OBJ-BG</td><td>OBJ-ONLY</td><td>PB-T50-RS</td></tr><tr><td>PointNet [29]</td><td>73.3</td><td>79.2</td><td>68.0</td></tr><tr><td>PointNet++ [30]</td><td>82.3</td><td>84.3</td><td>77.9</td></tr><tr><td>DGCNN [41]</td><td>82.8</td><td>86.2</td><td>78.1</td></tr><tr><td>PointCNN [23]</td><td>86.1</td><td>85.5</td><td>78.5</td></tr><tr><td>Transformer [49]</td><td>79.86</td><td>80.55</td><td>77.24</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>84.85</td><td>85.54</td><td>78.79</td></tr><tr><td>[S] Point-BERT[49]</td><td>87.43</td><td>88.12</td><td>83.07</td></tr><tr><td>[S] Point-M2AE</td><td>91.22</td><td>88.81</td><td>86.43</td></tr><tr><td>Improvement</td><td>+3.79</td><td>+0.69</td><td>+3.36</td></tr></table>
|
| 97 |
+
|
| 98 |
+
2-stage decoder for hierarchical learning. We adopt 5 blocks in each encoder stage, but only 1 block per stage for the lightweight decoder. For the 3-scale point clouds, we set the point numbers and token dimensions respectively as {512, 256, 64} and {96, 192, 384}. We also set different $k$ for the $k$ -NN at different scales, which are {16, 8, 8}. We mask the highest scale of point clouds with a high ratio of $80 \%$ and set 6 heads for all the attention modules. The detailed training settings are in Appendix.
|
| 99 |
+
|
| 100 |
+
Linear SVM. After pre-training on ShapeNet, we test the 3D representation capability of PointM2AE via linear evaluation on ModelNet40 [44]. We sample 1,024 points from each 3D shape of ModelNet40 and utilize our frozen encoder to extract their features. On top of that, we train a linear SVM and report the classification accuracy in Table 1. As shown, Point-M2AE achieves the best performance among all existing self-supervised methods for point clouds, and surpasses the second-best CrossPoint [2] by $+ 1 . 7 \%$ . Point-M2AE also exceeds Point-BERT [49] by $+ 5 . 5 \%$ , which is a masked point modeling method with a MoCo loss [19] but adopts a standard transformer and conducts single-scale learning. It is worth noting that even if we freeze all our parameters, Point-M2AE with $9 2 . 9 \%$ accuracy still outperforms many fully trained methods on ModelNet40, e.g., $9 0 . 5 \%$ by PointNet $^ { + + }$ [30], $9 2 . 8 \%$ by DensePoint [24], etc. The experiments fully demonstrate the superior 3D representation capacity of our Point-M2AE.
|
| 101 |
+
|
| 102 |
+
# 4.2 Downstream Tasks
|
| 103 |
+
|
| 104 |
+
For fine-tuning on downstream tasks, we discard the hierarchical decoder in pre-training and append different heads onto the hierarchical encoder for different tasks.
|
| 105 |
+
|
| 106 |
+
Shape Classification. We fine-tune Point-M2AE on two shape classification datasets: the widely adopted ModelNet40 [44] and the challenging ScanObjectNN [38]. For local spatial attention layers, we set the ball queries’ radii of 3-scale point clouds as {0.32, 0.64, 1.28}. We follow Point-BERT to use the voting strategy [25] for fair comparison on ModelNet40. To handle the noisy spatial structures, we increase $k$ of $k$ -NN into {32, 16, 16} for ScanObjectNN to encode local patterns with larger receptive fields. As reported in Table 2, Point-M2AE achieves $9 4 . 0 \%$ accuracy on ModelNet40 with 1024 points per sample, which surpasses Point-BERT fine-tuned with 1024 points by $+ 0 . 8 \%$ and 8192 points by $+ 0 . 2 \%$ . For ScanObjectNN in Table 3, our Point-M2AE outperforms the secondbest Point-BERT by a significant margin, $+ 3 . 7 9 \%$ , $+ 0 . 6 9 \%$ and $+ 3 . 3 6 \%$ , respectively for the three splits, indicating our great advantages under complex circumstances by multi-scale encoding. As ScanObjectNN of real-world scenes has a large semantic gap with the pre-trained synthetic ShapeNet, Point-M2AE also exerts strong transfer ability to understand point clouds of another domain.
|
| 107 |
+
|
| 108 |
+
Part Segmentation. We evaluate Point-M2AE for part segmentation on ShapeNetPart [48], which predicts per-point part labels and requires detailed understanding for local patterns. We adopt an extremely simple segmentation head to validate the effectiveness of our pre-training for well capturing both high-level semantics and fine-grained details. By the hierarchical encoder, we obtain
|
| 109 |
+
|
| 110 |
+
Table 4: Few-shot classification on ModelNet40 [44]. We report the average accuracy $( \% )$ and standard deviation $( \% )$ of 10 independent experiments.
|
| 111 |
+
|
| 112 |
+
<table><tr><td rowspan="2">Method</td><td colspan="2">5-way</td><td colspan="2">10-way</td></tr><tr><td>10-shot</td><td>20-shot</td><td>10-shot</td><td>20-shot</td></tr><tr><td>DGCNN [41]</td><td>91.8 ± 3.7</td><td>93.4 ± 3.2</td><td>86.3 ± 6.2</td><td>90.9 ± 5.1</td></tr><tr><td>[S] DGCNN + OcCo [40]</td><td>91.9 ± 3.3</td><td>93.9 ± 3.1</td><td>86.4 ± 5.4</td><td>91.3 ± 4.6</td></tr><tr><td>Transformer [49]</td><td>87.8 ±5.2</td><td>93.3 ± 4.3</td><td>84.6 ± 5.5</td><td>89.4 ± 6.3</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>94.0±3.6</td><td>95.9 ± 2.3</td><td>89.4 ± 5.1</td><td>92.4 ± 4.6</td></tr><tr><td>[S] Point-BERT[49]</td><td>94.6 ± 3.1</td><td>96.3 ± 2.7</td><td>91.0 ± 5.4</td><td>92.7 ± 5.1</td></tr><tr><td>[S] Point-M2AE</td><td>96.8 ± 1.8</td><td>98.3 ± 1.4</td><td>92.3 ± 4.5</td><td>95.0 ± 3.0</td></tr><tr><td>Improvement</td><td>+2.2</td><td>+2.0</td><td>+1.3</td><td>+2.3</td></tr></table>
|
| 113 |
+
|
| 114 |
+
Table 5: Part segmentation on ShapeNetPart [48]. $\mathbf { \hat { m } } \mathbf { I o U } _ { C }$ ’ $( \% )$ and $\mathbf { \dot { m l o U } } _ { I } ,$ $( \% )$ denote the mean IoU across all part categories and all instances in the dataset, respectively.
|
| 115 |
+
|
| 116 |
+
<table><tr><td>Method</td><td>mIoUc</td><td>mIoU1</td></tr><tr><td>PointNet [29] PointNet++ [30]</td><td>80.39 81.85</td><td>83.70 85.10</td></tr><tr><td>DGCNN [41]</td><td>82.33</td><td>85.20</td></tr><tr><td>Transformer [49]</td><td>83.42</td><td>85.10</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>83.42</td><td>85.10</td></tr><tr><td>[S] Point-BERT[49]</td><td>84.11</td><td>85.60</td></tr><tr><td>[S] Point-M2AE</td><td>84.86</td><td>86.51</td></tr><tr><td>Improvement</td><td>+0.75</td><td>+0.91</td></tr></table>
|
| 117 |
+
|
| 118 |
+
Table 6: 3D object detection on ScanNetV2 [9]. We report the performance $( \% )$ of self-supervised learning methods based on VoteNet [12] and 3DETR-m [26].
|
| 119 |
+
|
| 120 |
+
<table><tr><td>Method</td><td>AP25</td><td>AP50</td></tr><tr><td>VoteNet [12] [S] STRL [20] [S] PointContrast [45]</td><td>58.6 59.5 59.2</td><td>33.5 38.4 38.0</td></tr><tr><td>[S] DepthContrast [51] 3DETR[26]</td><td>61.3 62.1</td><td>1 37.9</td></tr><tr><td>3DETR-m [26]</td><td>65.0</td><td>47.0</td></tr><tr><td>[S] Point-M2AE</td><td>66.3</td><td>48.3</td></tr><tr><td>Improvement</td><td>+1.3</td><td>+1.3</td></tr></table>
|
| 121 |
+
|
| 122 |
+
3-scale point tokens of {512, 256, 64} points, and perform feature propagation in PointNet $^ { - + }$ [30] to independently upsample the tokens into 2048 points of the input point cloud. Then, we concatenate the upsampled 3-scale features for each point and predict the part label by stacked linear projection layers. As reported in Table 4.2, Point-M2AE achieves the best $8 6 . 5 1 \%$ instance mIoU with the simple segmentation head, surpassing the second-best Point-BERT by $+ 0 . 9 1 \%$ . Note that Point-BERT [49] and other methods [29, 30, 41] adopt hierarchical segmentation heads to progressively upsample the point features from intermediate layers, while our head contains no hierarchical structure and only relies on the pre-trained encoder to capture the multi-scale information of point clouds. The results fully demonstrate the significance of Point-M2AE’s multi-scale pre-training to segmentation tasks.
|
| 123 |
+
|
| 124 |
+
Few-shot Classification. We conduct experiments for few-shot classification on ModelNet40 [44] to evaluate the performance of Point-M2AE with limited fine-tuning data. As reported in Table 4.2, Point-M2AE achieves the best performance for all four settings, and surpasses Point-BERT by $+ 2 . 2 \%$ , $+ 2 . 0 \%$ , $+ 1 . 3 \%$ , and $+ 2 . 7 \%$ , respectively. Our approach also shows smaller deviations than other transformer-based methods, which indicates Point-M2AE has learned to produce more universal 3D representations for well adapting to downstream tasks under low-data regimes.
|
| 125 |
+
|
| 126 |
+
3D Object Detection To further evaluate our hierarchical pre-training on 3D object detection, we apply Point-M2AE to serving as the feature backbone on the indoor ScanNetV2 [9] dataset. We select 3DETR-m [26] as our baseline, which consists of a 3-block encoder and a transformer decoder. Considering the quite different dataset statistics, e.g., 2k input points for ShapeNet [6] and $5 0 \mathrm { k }$ input points for ScanNetV2, we adopt the same encoder architecture with that of 3DETR-m, and keep our hierarchical decoder with skip connections unchanged for self-supervised pre-training on ScanNetV2. More details of models and training are in Appendix. As reported in Table 4.2, compared to training from scratch, our hierarchical pre-training boosts the performance of 3DETR-m by $+ 1 . 3 4 \%$ $\mathrm { A P _ { 2 5 } }$ and $+ 1 . 2 9 \%$ $\mathsf { A P } _ { 5 0 }$ . The experiments demonstrate the effectiveness of Point-M2AE to learn multi-scale point cloud encoding for object detection and its potential to benefit a wider range of 3D applications.
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
Figure 4: Visualization of fine-grained information. We denote the outputs from hierarchical and non-hierarchical architectures as [NH] and [H], respectively. For an input point cloud (Middle), we visualize its extracted features (Left) and reconstruction results (Right).
|
| 130 |
+
|
| 131 |
+
Table 7: Hierarchical Modules. ‘H’ represents the encoder and decoder with multi-stage hierarchies. ‘Skip C.’ denotes the skip connections.
|
| 132 |
+
|
| 133 |
+
<table><tr><td>Encoder</td><td>Decoder</td><td>Skip C.</td><td>Acc. (%)</td></tr><tr><td>H</td><td>H</td><td>√</td><td>92.9</td></tr><tr><td></td><td>1</td><td>√</td><td>90.7</td></tr><tr><td>1</td><td>H</td><td>√</td><td>91.5</td></tr><tr><td>H</td><td>1</td><td>√</td><td>92.2</td></tr><tr><td>H</td><td>H</td><td></td><td>92.1</td></tr></table>
|
| 134 |
+
|
| 135 |
+
Table 8: Different Masking Strategy. ‘MS Mask’ and ‘Ratio’ denote the multi-scale masking and the mask ratio.
|
| 136 |
+
|
| 137 |
+
<table><tr><td>MS Mask</td><td>Ratio</td><td>Acc. (%)</td></tr><tr><td>了</td><td>0.8</td><td>92.9</td></tr><tr><td>1</td><td>0.8</td><td>88.4</td></tr><tr><td>√</td><td>0.6</td><td>92.3</td></tr><tr><td><</td><td>0.7</td><td>92.7</td></tr><tr><td>√</td><td>0.9</td><td>92.5</td></tr></table>
|
| 138 |
+
|
| 139 |
+
# 4.3 Ablation Study
|
| 140 |
+
|
| 141 |
+
We conduct ablation study by modifying one of the components at a time during pre-training and explore the best masking strategy. We report the classification accuracy on ModelNet40 [44] by linear SVM to evaluate the pre-trained representations. For downstream tasks, we train the network from scratch to validate the significance of our hierarchical pre-training.
|
| 142 |
+
|
| 143 |
+
Hierarchical Modules. As reported in Table 7, on top of our final solution, Point-M2AE, in the first row, we respectively experiment with removing the hierarchical encoder, hierarchical decoder, and skip connections from our framework. Specifically, we replace our encoder and decoder with 1-stage plain architectures similar to MAE, which contains 15 and 2 vanilla transformer blocks, respectively. We observe the absence of multi-stage structures either in encoder or decoder hurts the performance, and the hierarchical encoder plays a better role than the decoder. Also, the skip connections well benefits the accuracy by providing complementary information for the decoder.
|
| 144 |
+
|
| 145 |
+
Masking Strategy. In Table 8, we report Point-M2AE with different mask settings. Without the multi-scale masking, we randomly generate masks at each scale, which leads to fragmented visible regions for all scales. With this strategy, the network would ‘peek’ different parts of the point cloud at different stages, which disturbs the representation learning and harms the performance by $- 4 . 5 \%$ accuracy. For different mask ratios, we find the $8 0 \%$ ratio performs the best to build a properly challenging pretext task for self-supervised pre-training.
|
| 146 |
+
|
| 147 |
+
With and Without Pre-training. We report the performance of Point-M2AE on downstream tasks with and without the pre-training in Table 9. For ‘w/o’, we randomly initialize the parameters and train the network from scratch. As shown, the pre-training can largely boost the performance on four datasets respectively by $+ 1 . 5 \%$ , $+ 2 . 5 \%$ , $+ 3 . 8 \%$ , and $+ 1 . 1 \bar { \% }$ , which indicates the superiority and significance of our hierarchical pre-training.
|
| 148 |
+
|
| 149 |
+
Table 9: With and without the pre-training. ‘ModelNet40-FS’ denotes the few-shot classification on 10-way 20-shot ModelNet40 [44].
|
| 150 |
+
|
| 151 |
+
<table><tr><td>Dataset</td><td>w/o (%)</td><td>w (%)</td></tr><tr><td>ModelNet40 [44]</td><td>92.5</td><td>94.0</td></tr><tr><td>ScanObjectNN[38]</td><td>83.9</td><td>86.4</td></tr><tr><td>ModelNet40-FS [44]</td><td>91.2</td><td>95.0</td></tr><tr><td>ShapeNetPart [48]</td><td>85.4</td><td>86.5</td></tr></table>
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
Figure 5: Visualization of multi-scale point clouds. In each row, we visualize the input point clouds, their multi-scale representations, the reconstructed coordinates, and multi-scale masked point clouds.
|
| 155 |
+
|
| 156 |
+
# 5 Visualization
|
| 157 |
+
|
| 158 |
+
Multi-scale Masking. To ease the understanding of our multi-scale masking strategy, we visualize the input point cloud, the 3-scale representations, the reconstructed point cloud, and 3-scale masked point clouds, respectively in each row of Figure 5. As shown, different scales can represent different levels of geometric details and semantics for point clouds. By the multi-scale masking strategy, we observe the visible positions of masked point clouds are block-wise within one scale and consistent across scales, which is significant for our hierarchical pre-training.
|
| 159 |
+
|
| 160 |
+
Fine-grained Information. The fine-grained 3D structures, e.g., thin branches of a plant, fingers of a human, engines of a plane, are significant to distinguish similar shapes and can be well encoded by our hierarchical representations. In Figure 4, we compare our Point-M2AE with multi-stage, [H], and single-scale, [NH], architectures by visualizing their extracted point features and reconstructed point clouds during pre-training. In contarst to the single-scale network, the multi-scale one indicates higher feature responses in the fine-grained structures and reconstructs more accurate spatial details.
|
| 161 |
+
|
| 162 |
+
# 6 Conclusion
|
| 163 |
+
|
| 164 |
+
We propose Point-M2AE, a multi-scale masked autoencoder for self-supervised pre-training on 3D point clouds. With a hierarchical architecture, Point-M2AE learns to produce powerful 3D representations by encoding multi-scale point clouds and reconstructing the masked coordinates from a global-to-local upsampling scheme. Extensive experiments have demonstrated the superiority of Point-M2AE to be a strong 3D representation learner. For limitations and future work, we will focus on applying Point-M2AE for wider 3D applications, e.g., outdoor and open-world scene understanding. We do not foresee negative social impact from the proposed work.
|
| 165 |
+
|
| 166 |
+
Acknowledgement. This work is supported by the National Natural Science Foundation of China (Grant No. 62206272), Shanghai Committee of Science and Technology (Grant No. 21DZ1100100), Centre for Perceptual and Interactive Intelligence Limited, and the General Research Fund through the Research Grants Council of Hong Kong (Grant No. 14204021, 14207319).
|
| 167 |
+
|
| 168 |
+
# References
|
| 169 |
+
|
| 170 |
+
[1] Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 123–133, 2021. 3
|
| 171 |
+
[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. arXiv preprint arXiv:2203.00680, 2022. 2, 3, 6, 7
|
| 172 |
+
[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. 1, 3
|
| 173 |
+
[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3
|
| 174 |
+
[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020. 1, 3
|
| 175 |
+
[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6, 8
|
| 176 |
+
[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020. 1, 3
|
| 177 |
+
[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750–15758, 2021. 1, 3
|
| 178 |
+
[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 3, 8
|
| 179 |
+
[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 3
|
| 180 |
+
[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3
|
| 181 |
+
[12] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 202–210. Springer, 2019. 8
|
| 182 |
+
[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3
|
| 183 |
+
[14] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017. 6
|
| 184 |
+
[15] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187–199, 2021. 6
|
| 185 |
+
[16] Zhizhong Han, Mingyang Shang, Yu-Shen Liu, and Matthias Zwicker. View inter-prediction gan: Unsupervised representation learning for 3d shapes by learning global shape memories to support local view predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8376–8384, 2019. 6
|
| 186 |
+
[17] Zhizhong Han, Xiyang Wang, Yu-Shen Liu, and Matthias Zwicker. Multi-angle point cloudvae: Unsupervised feature learning for 3d point clouds from multiple angles by joint selfreconstruction and half-to-half prediction. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10441–10450. IEEE, 2019. 6
|
| 187 |
+
[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 3
|
| 188 |
+
[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020. 1, 3, 7 [20] Siyuan Huang, Song ervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6535–6545, 2021. 8 [21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, YunHsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages
|
| 189 |
+
4904–4916. PMLR, 2021. 1 [22] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 6 [23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on $\mathbf { X }$ -transformed points. Advances in neural information processing systems,
|
| 190 |
+
31:820–830, 2018. 6, 7 [24] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5239–5248,
|
| 191 |
+
2019. 7 [25] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8895–8904, 2019. 7 [26] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2906–2917, October 2021. 8 [27] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. arXiv preprint arXiv:2203.06604, 2022.
|
| 192 |
+
3 [28] Omid Poursaeed, Tianxing Jiang, Han Qiao, Nayun Xu, and Vladimir G Kim. Self-supervised learning of point clouds via orientation estimation. In 2020 International Conference on 3D Vision (3DV), pages 1018–1028. IEEE, 2020. 3 [29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017. 5, 6, 7, 8 [30] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 5, 6, 7, 8 [31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021. 1 [32] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 1, 3 [33] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1, 3 [34] Jason Tyler Rolfe. Discrete variational autoencoders. arXiv preprint arXiv:1609.02200, 2016. 3 [35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. 2, 4 [36] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 3 [37] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 6 [38] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1588–1597, 2019. 2, 7, 9 [39] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized representations of point clouds with graph-convolutional generative adversarial networks. IEEE Transactions on Multimedia, 23:402–414, 2020. 6 [40] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9782–9792, 2021. 3, 6, 8
|
| 193 |
+
[41] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1–12, 2019. 6, 7, 8
|
| 194 |
+
[42] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 3
|
| 195 |
+
[43] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 6
|
| 196 |
+
[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912–1920, 2015. 2, 6, 7, 8, 9
|
| 197 |
+
[45] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In European conference on computer vision, pages 574–591. Springer, 2020. 3, 8
|
| 198 |
+
[46] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 1
|
| 199 |
+
[47] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 6
|
| 200 |
+
[48] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1–12, 2016. 2, 7, 8, 9
|
| 201 |
+
[49] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Pointbert: Pre-training 3d point cloud transformers with masked point modeling. arXiv preprint arXiv:2111.14819, 2021. 3, 6, 7, 8
|
| 202 |
+
[50] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413, 2021. 1
|
| 203 |
+
[51] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252–10263, 2021. 3, 8
|
| 204 |
+
[52] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259– 16268, 2021. 6
|
| 205 |
+
[53] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 3
|
| 206 |
+
|
| 207 |
+
# Checklist
|
| 208 |
+
|
| 209 |
+
1. For all authors...
|
| 210 |
+
|
| 211 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 212 |
+
(b) Did you describe the limitations of your work? [Yes]
|
| 213 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes]
|
| 214 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 215 |
+
|
| 216 |
+
2. If you are including theoretical results...
|
| 217 |
+
|
| 218 |
+
(a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes]
|
| 219 |
+
|
| 220 |
+
3. If you ran experiments...
|
| 221 |
+
|
| 222 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes]
|
| 223 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes]
|
| 224 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes]
|
| 225 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes]
|
| 226 |
+
|
| 227 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 228 |
+
|
| 229 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 230 |
+
(b) Did you mention the license of the assets? [N/A]
|
| 231 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]
|
| 232 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
|
| 233 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
|
| 234 |
+
|
| 235 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 236 |
+
|
| 237 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
|
| 238 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 239 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
|
parse/dev/MbCAOMGsZXC/MbCAOMGsZXC_content_list.json
ADDED
|
@@ -0,0 +1,1053 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
210,
|
| 8 |
+
122,
|
| 9 |
+
789,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Renrui Zhang1,2, Ziyu Guo2, Rongyao Fang1, Bin Zhao2, Dong Wang2, Yu Qiao2, Hongsheng $\\mathbf { L i ^ { 1 , 3 } }$ , Peng GaoB2 ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
267,
|
| 19 |
+
224,
|
| 20 |
+
732,
|
| 21 |
+
260
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "1 CUHK-SenseTime Joint Laboratory, The Chinese University of Hong Kong, 2 Shanghai AI Laboratory, 3 Centre for Perceptual and Interactive Intelligence Limited ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
214,
|
| 30 |
+
268,
|
| 31 |
+
782,
|
| 32 |
+
297
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "{zhangrenrui, gaopeng}@pjlab.org.cnhsli@ee.cuhk.edu.hk",
|
| 39 |
+
"bbox": [
|
| 40 |
+
328,
|
| 41 |
+
301,
|
| 42 |
+
669,
|
| 43 |
+
328
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Abstract ",
|
| 50 |
+
"text_level": 1,
|
| 51 |
+
"bbox": [
|
| 52 |
+
462,
|
| 53 |
+
364,
|
| 54 |
+
535,
|
| 55 |
+
381
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "Masked Autoencoders (MAE) have shown great potentials in self-supervised pretraining for language and 2D image transformers. However, it still remains an open question on how to exploit masked autoencoding for learning 3D representations of irregular point clouds. In this paper, we propose Point-M2AE, a strong Multi-scale MAE pre-training framework for hierarchical self-supervised learning of 3D point clouds. Unlike the standard transformer in MAE, we modify the encoder and decoder into pyramid architectures to progressively model spatial geometries and capture both fine-grained and high-level semantics of 3D shapes. For the encoder that downsamples point tokens by stages, we design a multi-scale masking strategy to generate consistent visible regions across scales, and adopt a local spatial self-attention mechanism during fine-tuning to focus on neighboring patterns. By multi-scale token propagation, the lightweight decoder gradually upsamples point tokens with complementary skip connections from the encoder, which further promotes the reconstruction from a global-to-local perspective. Extensive experiments demonstrate the state-of-the-art performance of Point-M2AE for 3D representation learning. With a frozen encoder after pretraining, Point-M2AE achieves $9 2 . 9 \\%$ accuracy for linear SVM on ModelNet40, even surpassing some fully trained methods. By fine-tuning on downstream tasks, Point-M2AE achieves $8 6 . 4 3 \\%$ accuracy on ScanObjectNN, $+ 3 . 3 6 \\%$ to the secondbest, and largely benefits the few-shot classification, part segmentation and 3D object detection with the hierarchical pre-training scheme. Code is available at https://github.com/ZrrSkywalker/Point-M2AE. ",
|
| 62 |
+
"bbox": [
|
| 63 |
+
233,
|
| 64 |
+
397,
|
| 65 |
+
766,
|
| 66 |
+
700
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "1 Introduction ",
|
| 73 |
+
"text_level": 1,
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
728,
|
| 77 |
+
310,
|
| 78 |
+
746
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Learning to represent from unlabeled data without annotations, known as self-supervised learning, has attained great success in natural language processing [10, 32, 33, 5], computer vision [19, 7, 8, 18] and multi-modality learning [31, 50, 21]. By pre-training on the large-scale raw data, the networks are endowed with robust representation abilities and can significantly benefit downstream tasks with fine-tuning. Motivated by masked language modeling [32, 10], MAE [18] and some other methods [46, 53, 3] adopt asymmetric encoder-decoder transformers [13] to apply masked autoencoding for self-supervised learning on 2D images. They represent the input image as multiple local patches, and randomly mask them with a high ratio to build the pretext task for reconstruction. Specifically, the encoder aims at capturing high-level latent representations from limited visible patches, and the lightweight decoder is forced to reconstruct the RGB values of masked patches on top. Despite its superiority on grid-based 2D images, we ask the question: can MAE-style masked autoencoding be adapted to irregular point clouds as a powerful 3D representation learner? ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
761,
|
| 88 |
+
825,
|
| 89 |
+
898
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/3e2685f04f793b25c13779172a5bb0a5596cb22169a2fbd5a4fb91ad34eb5be0.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: Comparison of MAE (Top) and our Point-M2AE (Bottom). MAE for 2D image pretraining adopts standard transformer of the plain encoder and decoder, while Point-M2AE introduces a hierarchical transformer with skip connections for multi-scale point cloud pre-training. "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
176,
|
| 102 |
+
112,
|
| 103 |
+
800,
|
| 104 |
+
333
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "",
|
| 111 |
+
"bbox": [
|
| 112 |
+
173,
|
| 113 |
+
428,
|
| 114 |
+
821,
|
| 115 |
+
457
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "To tackle this challenge, we propose Multi-scale Masked autoencoders for learning the hierarchical representations of point clouds via self-supervised pre-training, termed as Point-M2AE. We represent a point cloud as a set of point tokens depicting different spatial local regions, and inherit MAE’s pipeline to first encode visible point tokens and then reconstruct the masked 3D coordinates. Different from 2D images, masked autoencoding for 3D point clouds has three characteristics to be specially considered. Firstly, it is critical to understand the relations between local parts and the overall 3D shapes, which have strong geometric and semantic dependence. As examples, the network can recognize an airplane starting from its wing, or segment the wing’s part from the airplane’s global feature. Therefore, we regard the standard transformer with the plain encoder and decoder is sub-optimal for capturing such local-global spatial relations in 3D, which directly downsamples the input into a low-resolution representation as shown in Figure 1 (Top). We modify both the encoder and decoder into multi-stage hierarchies for progressively encoding multi-scale features of point clouds, constructing an asymmetric U-Net [35] like architecture in Figure 1 (Bottom). Secondly, as our Point-M2AE encodes multi-scale point clouds unlike the single-scale 2D images, the unmasked visible regions are required to be both block-wise within one scale and consistent across scales, which are respectively for reserving complete local geometries and ensuring coherent feature learning for the network. For this, we introduce a multi-scale masking strategy, which generates random masks at the final scale with a high ratio (e.g., $80 \\%$ ), and back-projects the unmasked positions to all preceding scales. Thirdly, to better reconstruct 3D geometries from a local-to-global perspective, we utilize skip connections to complement the decoder with fine-grained information from the corresponding stages of the encoder. During fine-tuning on downstream tasks, we also adopt a local spatial self-attention mechanism with increasing attention scopes for point tokens at different stages of the encoder, which refocus each token within neighboring detailed structures. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
173,
|
| 124 |
+
463,
|
| 125 |
+
825,
|
| 126 |
+
780
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "By the multi-scale pre-training, Point-M2AE can encode point clouds from local-to-global hierarchies and then reconstructs the masked coordinates from global-to-local perspectives, which learns powerful 3D representations and performs superior transfer ability. After self-supervised pre-training on ShapeNet [6], Point-M2AE achieves $9 2 . 9 \\%$ classification accuracy for linear SVM on ModelNet40 [44] with the frozen encoder, which surpasses the runner-up CrossPoint [2] by $+ 1 . 2 \\%$ and even outperforms some fully supervised methods. By fine-tuning on various downstream tasks, Point-M2AE achieves $8 6 . 4 3 \\%$ $( + 3 . 3 6 \\% )$ accuracy on ScanObjectNN [38] and $9 4 . 0 \\%$ $( + 0 . 8 \\% )$ accuracy on ModelNet40 [44] for shape classification, $8 6 . 5 1 \\%$ $( + 0 . 9 1 \\% )$ instance mIoU on ShapeNetPart [48] for part segmentation, and $9 5 . 0 \\%$ $( + 2 . 7 \\% )$ accuracy on 10-way 20-shot ModelNet40 for few-shot classification. Our multi-scale masked autoencoding also benefits the 3D object detection on ScanNetV2 [9] by $+ 1 . 3 \\%$ $\\mathsf { A P } _ { 2 5 }$ and $+ 1 . 3 \\%$ $\\mathrm { { A P } _ { 5 0 } }$ , which provides the detection backbone with a hierarchical understanding of the point clouds. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
786,
|
| 136 |
+
825,
|
| 137 |
+
911
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "",
|
| 144 |
+
"bbox": [
|
| 145 |
+
176,
|
| 146 |
+
90,
|
| 147 |
+
825,
|
| 148 |
+
133
|
| 149 |
+
],
|
| 150 |
+
"page_idx": 2
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"type": "text",
|
| 154 |
+
"text": "We summarize the contributions of our paper as follows: ",
|
| 155 |
+
"bbox": [
|
| 156 |
+
174,
|
| 157 |
+
140,
|
| 158 |
+
544,
|
| 159 |
+
154
|
| 160 |
+
],
|
| 161 |
+
"page_idx": 2
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"type": "text",
|
| 165 |
+
"text": "1. We propose Point-M2AE, a strong masked autoencoding framework, which conducts hierarchical point cloud encoding and reconstruction for better learning multi-scale spatial geometries of 3D shapes. \n2. We introduce a U-Net like transformer architecture for MAE-style pre-training on point clouds, and adopt a multi-scale masking strategy to generate consistent visible regions across scales. \n3. Point-M2AE achieves state-of-the-art performance for transfer learning on various downstream tasks, which indicates our approach to be a powerful representation learner for 3D point clouds. ",
|
| 166 |
+
"bbox": [
|
| 167 |
+
210,
|
| 168 |
+
165,
|
| 169 |
+
825,
|
| 170 |
+
304
|
| 171 |
+
],
|
| 172 |
+
"page_idx": 2
|
| 173 |
+
},
|
| 174 |
+
{
|
| 175 |
+
"type": "text",
|
| 176 |
+
"text": "2 Related Work ",
|
| 177 |
+
"text_level": 1,
|
| 178 |
+
"bbox": [
|
| 179 |
+
174,
|
| 180 |
+
324,
|
| 181 |
+
321,
|
| 182 |
+
342
|
| 183 |
+
],
|
| 184 |
+
"page_idx": 2
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"type": "text",
|
| 188 |
+
"text": "Pre-training by Masked Modeling. Compared to contrastive learning methods [19, 7, 8] that learn from inter-sample relations, self-supervised pre-training by masked autoencoding builds the pretext tasks to predict the masked parts of the input signals. The series of GPT [32, 33, 5] and BERT [11] apply masked modeling to natural language processing and achieve extraordinary performance boost on downstream tasks with fine-tuning. Inspired by this, BEiT [4] proposes to match image patches with discrete tokens via dVAE [34] and pre-train a standard vision transformer [13, 49] by masked image modeling. On top of that, MAE [18] directly reconstructs the raw pixel values of masked tokens and performs great efficiency with a high mask ratio. The follow-up works further improve the performance of MAE by momentum encoder [53], contrastive learning [3], and modified reconstruction targets [42]. For self-supervised pre-training on 3D point clouds, the masked autoencoding has not been widely adopted. Similar to BEiT, Point-BERT [49] utilizes dVAE to map 3D patches to tokens for masked point modeling, but heavily relies on constrastive learning [19], complicated data augmentation, and the costly two-stage pre-training. In contrast, our Point-M2AE is a pure masked autoencoding method of one-stage pre-training, and follows MAE to reconstruct the input signals without dVAE mapping. Different from previous MAE methods adopting standard plain transformer, we propose a hierarchical transformer architecture along with the multi-scale masking strategy to better learn a strong and generic representation for 3D point clouds. ",
|
| 189 |
+
"bbox": [
|
| 190 |
+
174,
|
| 191 |
+
358,
|
| 192 |
+
825,
|
| 193 |
+
592
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "Self-supervised Learning for Point Clouds. 3D representation learning without annotations has been widely studied in recent years. Mainstream methods mainly build the pretext tasks to reconstruct the transformed input point cloud based on the encoded latent vectors, such as rotation [28], deformation [1], rearranged parts [36] and occlusion [40]. From another perspective, PointContrast [45] utilizes contrastive learning between features of the same points from different views to learn discriminative 3D representations. DepthContrast [51] further extends the contrast for depth maps of different augmentations. CrossPoint [2] conducts cross-modality contrastive learning between point clouds and their corresponding rendering images to acquire rich self-supervised signals. Point-BERT [49] and Point-MAE [27] respectively introduce BERT-style [10] and MAE-style [18] pre-training schemes for 3D point clouds with standard transformer networks and performs competitively on various downstream tasks, but both of them can only encode point clouds with a single resolution and ignores the local-global relations between 3D shapes. In this paper, we propose Point-M2AE, an MAE-style framework with a hierarchical transformer for multi-scale point cloud pre-training. We achieve state-of-the-art downstream performance by learning the multi-scale representation of point clouds. ",
|
| 200 |
+
"bbox": [
|
| 201 |
+
174,
|
| 202 |
+
609,
|
| 203 |
+
825,
|
| 204 |
+
803
|
| 205 |
+
],
|
| 206 |
+
"page_idx": 2
|
| 207 |
+
},
|
| 208 |
+
{
|
| 209 |
+
"type": "text",
|
| 210 |
+
"text": "3 Method ",
|
| 211 |
+
"text_level": 1,
|
| 212 |
+
"bbox": [
|
| 213 |
+
174,
|
| 214 |
+
823,
|
| 215 |
+
269,
|
| 216 |
+
840
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "The overall pipeline of Point-M2AE is shown in Figure 2, where we encode and reconstruct the point cloud by a hierarchical network architecture. In Section 3.1, We first introduce the masking strategy of Point-M2AE with multi-scale representations of point clouds. Then in Section 3.2 and Section 3.3, we present the details of our encoder and decoder with multi-stage hierarchies. ",
|
| 223 |
+
"bbox": [
|
| 224 |
+
176,
|
| 225 |
+
856,
|
| 226 |
+
825,
|
| 227 |
+
911
|
| 228 |
+
],
|
| 229 |
+
"page_idx": 2
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"type": "image",
|
| 233 |
+
"img_path": "images/edb748d5a03aeab9c2928941bc8ab907c1bfafa8355882c3a4e1bd26f171de50.jpg",
|
| 234 |
+
"image_caption": [
|
| 235 |
+
"Figure 2: Overall pipeline of Point-M2AE. After the multi-scale masking, we embed point tokens at the 1-st scale and feed the visible ones into a hierarchical encoder-decoder transformer, which captures both high-level semantics and fine-grained patterns of the point cloud during pre-training. "
|
| 236 |
+
],
|
| 237 |
+
"image_footnote": [],
|
| 238 |
+
"bbox": [
|
| 239 |
+
174,
|
| 240 |
+
94,
|
| 241 |
+
821,
|
| 242 |
+
314
|
| 243 |
+
],
|
| 244 |
+
"page_idx": 3
|
| 245 |
+
},
|
| 246 |
+
{
|
| 247 |
+
"type": "text",
|
| 248 |
+
"text": "3.1 Multi-scale Masking ",
|
| 249 |
+
"text_level": 1,
|
| 250 |
+
"bbox": [
|
| 251 |
+
174,
|
| 252 |
+
393,
|
| 253 |
+
356,
|
| 254 |
+
409
|
| 255 |
+
],
|
| 256 |
+
"page_idx": 3
|
| 257 |
+
},
|
| 258 |
+
{
|
| 259 |
+
"type": "text",
|
| 260 |
+
"text": "To build a U-Net [35] like masked autoencoder for hierarchical learning, we encode the point cloud by $S$ scales with different number of points at each scale, and correspondingly modify the standard plain encoder into the $S$ -stage architecture. Following MAE, we embed the point cloud into discrete point tokens and randomly mask them for reconstruction. Importantly, for irregular-distributed points in the multi-scale architecture, the unmasked visible spatial regions are required to be consistent not only within one scale, but also across different scales. This is because the block-wise parts of 3D shapes tend to preserve more complete fine-grained geometries, and the unmasked positions are better to be shared across all scales for coherent feature learning of the encoder. Therefore, as shown in Figure 3, we first construct the $S$ -scale coordinate representations of the input point cloud and back-project the random masks from the final $S$ -th scale to the earlier scales to avoid fragmented visible parts. ",
|
| 261 |
+
"bbox": [
|
| 262 |
+
173,
|
| 263 |
+
419,
|
| 264 |
+
825,
|
| 265 |
+
571
|
| 266 |
+
],
|
| 267 |
+
"page_idx": 3
|
| 268 |
+
},
|
| 269 |
+
{
|
| 270 |
+
"type": "text",
|
| 271 |
+
"text": "$S$ -scale Representations. We denote the input point cloud as $P \\in \\mathbb { R } ^ { N \\times 3 }$ and regard it as the 0-th scale. For the $i$ -th scale, $1 \\leq i \\leq S$ , we utilize Furthest Point Sampling (FPS) to downsample the points from the $( i - 1 )$ -th scale, which produces seed points $P _ { i } \\in \\mathbb { R } ^ { N _ { i } \\times 3 }$ for scale $i$ of $N _ { i }$ points. Then, we adopt $k$ Nearest-Neighbour ( $k$ -NN) to aggregate the neighboring $k$ points for each seed point and obtain the neighbor indices $I _ { i } \\in \\mathbb { R } ^ { N _ { i } \\times k }$ . By successively downsampling and grouping, we acquire the $S$ -scale representations $\\{ P _ { i } , I _ { i } \\} _ { i = 1 } ^ { S }$ of the input point cloud, where the number of points $N _ { i }$ gradually decreases and the inclusion relations between scales are recorded in $I _ { i }$ . ",
|
| 272 |
+
"bbox": [
|
| 273 |
+
174,
|
| 274 |
+
585,
|
| 275 |
+
825,
|
| 276 |
+
683
|
| 277 |
+
],
|
| 278 |
+
"page_idx": 3
|
| 279 |
+
},
|
| 280 |
+
{
|
| 281 |
+
"type": "text",
|
| 282 |
+
"text": "Back-projecting Visible Positions. For seed points $P _ { S }$ at the final $S$ -th scale, we randomly mask them with a large proportion (e.g., $80 \\%$ ) and denote the remaining visible points as $P _ { S } ^ { v } \\in \\dot { \\mathbb { R } } ^ { N _ { S } ^ { v } \\times 3 }$ of $N _ { S }$ points. We then back-project the unmasked positions $P _ { S } ^ { v }$ to ensure the consistent visible regions across scales. For the $i$ -th scale, $1 \\leq i < S$ , we retrieve all the $k$ nearest neighbors of $P _ { i + 1 } ^ { v }$ from the indices $I _ { i + 1 }$ to serve as the visible positions $P _ { i } ^ { v }$ , and mask the others. By recursively back-projecting, we obtain the visible and masked positions of all v m $S$ scales, denoted as $\\{ P _ { i } ^ { v } , P _ { i } ^ { m } \\} _ { i = 1 } ^ { S }$ , where $P _ { i } ^ { v } \\in \\mathbb { R } ^ { N _ { i } ^ { v } \\times 3 }$ , $P _ { i } ^ { m } \\in \\mathbb { R } ^ { N _ { i } ^ { m } \\times 3 }$ and $N _ { i } = N _ { i } ^ { v } + N _ { i } ^ { m }$ . ",
|
| 283 |
+
"bbox": [
|
| 284 |
+
173,
|
| 285 |
+
696,
|
| 286 |
+
825,
|
| 287 |
+
801
|
| 288 |
+
],
|
| 289 |
+
"page_idx": 3
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"type": "text",
|
| 293 |
+
"text": "3.2 Hierarchical Encoder ",
|
| 294 |
+
"text_level": 1,
|
| 295 |
+
"bbox": [
|
| 296 |
+
176,
|
| 297 |
+
815,
|
| 298 |
+
362,
|
| 299 |
+
830
|
| 300 |
+
],
|
| 301 |
+
"page_idx": 3
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"type": "text",
|
| 305 |
+
"text": "Based on the multi-scale masking, we embed the initial tokens of visible points $P _ { 1 } ^ { v }$ for the 1-st scale and them into the hierarchical encoder with $S$ stages. Every stage is equipped with $K$ stacked encoder blocks, and each block contains a self-attention layer and a Feed Forward Network (FFN) of MLP layers. Between every two consecutive stages, we introduce spatial token merging modules to aggregate adjacent visible tokens and enlarge receptive fields for downsampling the point clouds. ",
|
| 306 |
+
"bbox": [
|
| 307 |
+
174,
|
| 308 |
+
842,
|
| 309 |
+
825,
|
| 310 |
+
911
|
| 311 |
+
],
|
| 312 |
+
"page_idx": 3
|
| 313 |
+
},
|
| 314 |
+
{
|
| 315 |
+
"type": "image",
|
| 316 |
+
"img_path": "images/ebfbe50f633e3247458dfe15d4dd00e5e63faa29175b365476efad0435636ab8.jpg",
|
| 317 |
+
"image_caption": [
|
| 318 |
+
"Figure 3: Multi-scale masking strategy. To obtain a consistent visible regions across scales, we first represent the input point cloud by multi-scale coordinates and generate the random mask at the highest one. Then, we back-project the unmasked visible positions to all earlier scales. "
|
| 319 |
+
],
|
| 320 |
+
"image_footnote": [],
|
| 321 |
+
"bbox": [
|
| 322 |
+
187,
|
| 323 |
+
87,
|
| 324 |
+
803,
|
| 325 |
+
233
|
| 326 |
+
],
|
| 327 |
+
"page_idx": 4
|
| 328 |
+
},
|
| 329 |
+
{
|
| 330 |
+
"type": "text",
|
| 331 |
+
"text": "Token Embedding and Merging. Indexed by $I _ { 1 }$ , we utilize a mini-PointNet [29] to extract and fuse the features of every seed point from $P _ { 1 } ^ { v } \\in \\mathbb { R } ^ { N _ { 1 } ^ { v } \\times 3 }$ with its $k$ nearest neighbors. After that, we obtain the initial point tokens $T _ { 1 } ^ { v } \\in \\mathbb { R } ^ { N _ { 1 } ^ { v } \\times C _ { 1 } }$ for the 1-st stage of the encoder, which embeds $N _ { 1 } ^ { e }$ local patterns of the 3D shape. Between the $( i - 1 )$ -th and $i$ -th stages, $1 < i \\leq S$ , we merge $T _ { i - 1 } ^ { \\dot { v } } \\in \\mathbb { R } ^ { \\mathsf { \\tilde { N } } _ { i - 1 } \\times C _ { i - 1 } }$ to acquire the downsampled point tokens for the $i$ -th stage. We utilize MLP layers and a max pooling to integrate every $k$ tokens nearest to $P _ { i } ^ { v }$ indexed by $I _ { i }$ , which outputs $T _ { i } ^ { v } \\in \\mathbb { R } ^ { N _ { i } \\times C _ { i } }$ . Due to our multi-scale masking, the merged $T _ { i } ^ { v }$ corresponds to the same visible parts of $T _ { i - 1 } ^ { v }$ , which enables the consistent feature encoding across different scales. For larger $i$ of deeper stages, we set higher feature dimension $C _ { i }$ to encode spatial geometries with richer semantics. ",
|
| 332 |
+
"bbox": [
|
| 333 |
+
173,
|
| 334 |
+
320,
|
| 335 |
+
825,
|
| 336 |
+
450
|
| 337 |
+
],
|
| 338 |
+
"page_idx": 4
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"type": "text",
|
| 342 |
+
"text": "Local Spatial Self-Attention. During pre-training, we expect point tokens in the multi-stage encoder to capture global cues for 3D shapes, which benefits the reconstruction of masked parts. However, when fine-tuning on downstream tasks without masked autoencoding, point tokens in the shallower stages are better to mainly focus on local information and not to be disturbed by long-range signals, referring to the inductive bias of 3D locality [30]. Thus, during fine-tuning, we modify the original self-attention layer in the encoder with a local spatial constraint that only neighboring tokens within a ball query would be available for attention calculation. As the point tokens are downsampled by stages, we set increasing radii $\\{ r _ { i } \\} _ { i = 1 } ^ { S }$ of multi-scale ball queries for gradually expanding the attention scopes, which fulfills the local-to-global feature aggregation scheme. ",
|
| 343 |
+
"bbox": [
|
| 344 |
+
173,
|
| 345 |
+
472,
|
| 346 |
+
825,
|
| 347 |
+
598
|
| 348 |
+
],
|
| 349 |
+
"page_idx": 4
|
| 350 |
+
},
|
| 351 |
+
{
|
| 352 |
+
"type": "text",
|
| 353 |
+
"text": "3.3 Hierarchical Decoder ",
|
| 354 |
+
"text_level": 1,
|
| 355 |
+
"bbox": [
|
| 356 |
+
176,
|
| 357 |
+
621,
|
| 358 |
+
361,
|
| 359 |
+
636
|
| 360 |
+
],
|
| 361 |
+
"page_idx": 4
|
| 362 |
+
},
|
| 363 |
+
{
|
| 364 |
+
"type": "text",
|
| 365 |
+
"text": "Via the hierarchical encoder, we obtain the encoded visible tokens $\\{ T _ { i } ^ { v } \\} _ { i = 1 } ^ { S }$ of all scales. Starting from the highest and concatenate $S$ -th scale, we assign a sharedem with the visible tokens arnable mask token to. We denote them as ked positions with coordin $P _ { S } ^ { m }$ $T _ { S } ^ { v }$ $\\{ H _ { 1 } ^ { v } , H _ { 1 } ^ { m } \\}$ $\\{ P _ { S } ^ { v } , P _ { S } ^ { m } \\}$ S , which serve as the input of the hierarchical decoder. We design the decoder to be lightweight with $S - 1$ stages and only one decoder block for each stage, which enforces the encoder to embed more semantics of the point clouds. Each decoder block consists of a vanilla self-attention layer and an FFN. We do not apply the local constraint to the attention in the decoder, since a global understanding between visible and mask tokens is crucial to the reconstruction. ",
|
| 366 |
+
"bbox": [
|
| 367 |
+
173,
|
| 368 |
+
648,
|
| 369 |
+
825,
|
| 370 |
+
761
|
| 371 |
+
],
|
| 372 |
+
"page_idx": 4
|
| 373 |
+
},
|
| 374 |
+
{
|
| 375 |
+
"type": "text",
|
| 376 |
+
"text": "Point Token Upsampling. We upsample the point tokens between stages to progressively recover the fine-grained geometries of 3D shapes before reconstruction. We regulate that the $j$ -th stage of the decoder corresponds to the $( S + 1 - j )$ -th stage of the encoder, both of which contain point tokens of the same $( S + 1 - j )$ -th scale with the feature dimension $C _ { S + 1 - j }$ . Between the $( j - 1 )$ - th and $j$ -th stage, $1 < j \\le S - 1$ , we upsample the tokens $\\{ H _ { j - 1 } ^ { v } , H _ { j - 1 } ^ { m } \\}$ from the coordinates $\\{ P _ { S + 2 - j } ^ { v } , P _ { S + 2 - j } ^ { m } \\}$ into $\\{ P _ { S + 1 - j } ^ { v } , P _ { S + 1 - j } ^ { m } \\}$ via the token propagation module. Specifically, we obtain the $k$ nearest neighbors of each point token in $\\{ H _ { j - 1 } ^ { v } , H _ { { \\underline { { j } } } - 1 } ^ { m } \\}$ indexed by $I _ { S + 2 - j }$ , and recover their neighbors’ features by weighted interpolation referring to PointNet+ $^ { \\cdot + }$ [30], which generates the tokens $\\{ \\bar { H } _ { j } ^ { v } , H _ { j } ^ { m } \\}$ of the $j$ -th stage. ",
|
| 377 |
+
"bbox": [
|
| 378 |
+
174,
|
| 379 |
+
782,
|
| 380 |
+
825,
|
| 381 |
+
912
|
| 382 |
+
],
|
| 383 |
+
"page_idx": 4
|
| 384 |
+
},
|
| 385 |
+
{
|
| 386 |
+
"type": "table",
|
| 387 |
+
"img_path": "images/586e91fd650989908d3aa070a1629b0d02487d92914c7d527402a7298246efcf.jpg",
|
| 388 |
+
"table_caption": [
|
| 389 |
+
"Table 1: Linear evaluation on ModelNet40 [44] by SVM. We report different self-supervised learning methods and underline the second-best one. "
|
| 390 |
+
],
|
| 391 |
+
"table_footnote": [],
|
| 392 |
+
"table_body": "<table><tr><td rowspan=1 colspan=2>Method Acc. (%)</td></tr><tr><td rowspan=1 colspan=1></td><td rowspan=3 colspan=1>3D-GAN [43] 83.3Latent-GAN [39] 85.7</td></tr><tr><td rowspan=5 colspan=1></td><td rowspan=5 colspan=1>Latent-GAN [39] 85.7SO-Net [22] 87.3FoldingNet [47] 88.4MAP-VAE[17] 88.4VIP-GAN[16] 90.2</td></tr><tr><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>87.3</td></tr><tr><td rowspan=1 colspan=1></td></tr><tr><td rowspan=1 colspan=1>88.4</td></tr><tr><td rowspan=4 colspan=1></td><td rowspan=2 colspan=1>DGCNN + Jiasaw [37] 90.6DGCNN + OcCo [40] 90.7DGCNN + CrossPoint [2] 91.2</td></tr><tr><td rowspan=1 colspan=1>90.7</td></tr><tr><td rowspan=1 colspan=1>Transformer + OcCo [49] 89.6</td></tr><tr><td rowspan=1 colspan=1>Point-BERT[49] 87.4</td></tr><tr><td rowspan=1 colspan=2>Point-M2AE 92.9</td></tr><tr><td rowspan=1 colspan=2>Improvement +1.7</td></tr></table>",
|
| 393 |
+
"bbox": [
|
| 394 |
+
174,
|
| 395 |
+
150,
|
| 396 |
+
446,
|
| 397 |
+
378
|
| 398 |
+
],
|
| 399 |
+
"page_idx": 5
|
| 400 |
+
},
|
| 401 |
+
{
|
| 402 |
+
"type": "table",
|
| 403 |
+
"img_path": "images/1e1c289b35c86a2e855b835d498838e93654f362fad95b4c902fa89a8f2289f9.jpg",
|
| 404 |
+
"table_caption": [
|
| 405 |
+
"Table 2: Shape classification on ModelNet40 [44]. ‘#points’ and ‘Acc.’ denote the number of points for training and the overall accuracy. [S] represents finetuning after self-supervised pre-training. "
|
| 406 |
+
],
|
| 407 |
+
"table_footnote": [],
|
| 408 |
+
"table_body": "<table><tr><td>Method</td><td>#points</td><td>Acc. (%)</td></tr><tr><td>PointNet [29] PointNet++ [30]</td><td>1k</td><td>89.2</td></tr><tr><td></td><td>1k</td><td>90.5</td></tr><tr><td>PointCNN [23]</td><td>1k</td><td>92.2</td></tr><tr><td>[S] SO-Net [22] DGCNN [41]</td><td>5k</td><td>92.5</td></tr><tr><td>PCT[15]</td><td>1k</td><td>92.9</td></tr><tr><td>Point Transformer [52]</td><td>1k</td><td>93.2</td></tr><tr><td></td><td>-</td><td>93.7</td></tr><tr><td>Transformer [49]</td><td>1k</td><td>91.4</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>1k</td><td>92.1</td></tr><tr><td>[S] Point-BERT[49]</td><td>1k</td><td>93.2</td></tr><tr><td>[S] Point-BERT</td><td>4k</td><td>93.4</td></tr><tr><td>[S] Point-BERT</td><td>8k</td><td>93.8</td></tr><tr><td>[S] Point-M2AE</td><td>1k</td><td>94.0</td></tr></table>",
|
| 409 |
+
"bbox": [
|
| 410 |
+
462,
|
| 411 |
+
150,
|
| 412 |
+
825,
|
| 413 |
+
378
|
| 414 |
+
],
|
| 415 |
+
"page_idx": 5
|
| 416 |
+
},
|
| 417 |
+
{
|
| 418 |
+
"type": "text",
|
| 419 |
+
"text": "Skip Connections. To further complement the fine-grained geometries, we channel-wisely concatenate the visible tokens $H _ { j } ^ { v } \\in \\mathbb { R } ^ { N _ { S + 1 - j } \\times C _ { S + 1 - j } }$ of the decoder with $T _ { S + 1 - j } ^ { v } \\in \\mathbb { R } ^ { N _ { S + 1 - j } \\times \\check { C } _ { S + 1 - j } }$ from the corresponding $( S + 1 - j )$ -th stage of the encoder via skip connections, and adopt a linear projection layer to fuse their features. For the mask tokens $H _ { j } ^ { m }$ , we keep them unchanged, since the encoder only contains visible tokens without the masked ones. ",
|
| 420 |
+
"bbox": [
|
| 421 |
+
173,
|
| 422 |
+
405,
|
| 423 |
+
825,
|
| 424 |
+
477
|
| 425 |
+
],
|
| 426 |
+
"page_idx": 5
|
| 427 |
+
},
|
| 428 |
+
{
|
| 429 |
+
"type": "text",
|
| 430 |
+
"text": "Point Reconstruction. After $S - 1$ stages of the decoder, we acquire $\\{ H _ { S - 1 } ^ { v } , H _ { S - 1 } ^ { m } \\}$ with coordinates $\\{ P _ { 2 } ^ { v } , P _ { 2 } ^ { m } \\}$ and reconstruct the masked values from the mask tokens $H _ { S - 1 } ^ { m }$ . Other than predicting values at the 0-th scale of the input point cloud $P$ , we reconstruct the coordinates of $P _ { 1 } ^ { m }$ , namely, recovering the masked positions of the 1-st scale $P _ { 1 } ^ { m } \\in \\mathbb { R } ^ { N _ { 1 } ^ { m } \\times 3 }$ from the 2-nd scale $\\bar { P _ { 2 } ^ { m } } \\in \\mathbb { R } ^ { N _ { 2 } ^ { m } \\times 3 }$ . This is because $\\{ P _ { 1 } ^ { v } , P _ { 1 } ^ { m } \\}$ of the 1-st scale could well represent the overall 3D shape and simultaneously preserve enough local patterns, which already constructs a comparatively challenging pretext task for pre-training. If we further upsample $\\{ H _ { S - 1 } ^ { \\bar { v } } , H _ { S - 1 } ^ { m } \\}$ into $\\{ \\bar { H } _ { S } ^ { v } , H _ { S } ^ { m } \\}$ and reconstruct the masked raw points from $P _ { 1 } ^ { m }$ , the extra spatial noises and computational overhead would adversely influence our performance and efficiency. Therefore, for every token in $H _ { S - 1 } ^ { m } \\in \\mathbb { R } ^ { N _ { 2 } ^ { m } \\times C _ { 2 } }$ , we reconstruct its $k$ nearest neighbors recorded in $I _ { 2 }$ by a reconstruction head of one linear projection layer and compute the loss by $l _ { 2 }$ Chamfer Distance [14], formulated as, ",
|
| 431 |
+
"bbox": [
|
| 432 |
+
173,
|
| 433 |
+
491,
|
| 434 |
+
825,
|
| 435 |
+
650
|
| 436 |
+
],
|
| 437 |
+
"page_idx": 5
|
| 438 |
+
},
|
| 439 |
+
{
|
| 440 |
+
"type": "equation",
|
| 441 |
+
"img_path": "images/b94072f30ee88e378652230950f3aaaea01e064eeb6ccdc36b935809dff609b9.jpg",
|
| 442 |
+
"text": "$$\n\\begin{array} { r l } & { \\widehat { P } _ { 2 \\to 1 } ^ { m } = \\mathrm { L i n e a r } ( H _ { S - 1 } ^ { m } ) , \\mathrm { ~ w h e r e ~ } \\widehat { P } _ { 2 \\to 1 } ^ { m } \\in \\mathbb { R } ^ { N _ { 2 } ^ { m } \\times k \\times 3 } , } \\\\ & { \\mathcal { L } _ { C D } = \\mathrm { C h a m f e r D i s t a n c e } ( P _ { 2 \\to 1 } ^ { m } , \\widehat { P } _ { 2 \\to 1 } ^ { m } ) , } \\end{array}\n$$",
|
| 443 |
+
"text_format": "latex",
|
| 444 |
+
"bbox": [
|
| 445 |
+
320,
|
| 446 |
+
655,
|
| 447 |
+
676,
|
| 448 |
+
700
|
| 449 |
+
],
|
| 450 |
+
"page_idx": 5
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"type": "text",
|
| 454 |
+
"text": "where $\\widehat { P } _ { 2 1 } ^ { m }$ and $P _ { 2 1 } ^ { m }$ denote the predicted and ground-truth reconstruction coordinates from the 2-nd scale to the 1-st scale. We only utilize $\\mathcal { L } _ { C D }$ for supervision without contrastive loss to conduct a pure masked autoencoding for self-supervised pre-training. ",
|
| 455 |
+
"bbox": [
|
| 456 |
+
174,
|
| 457 |
+
707,
|
| 458 |
+
823,
|
| 459 |
+
750
|
| 460 |
+
],
|
| 461 |
+
"page_idx": 5
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"type": "text",
|
| 465 |
+
"text": "4 Experiments ",
|
| 466 |
+
"text_level": 1,
|
| 467 |
+
"bbox": [
|
| 468 |
+
174,
|
| 469 |
+
768,
|
| 470 |
+
312,
|
| 471 |
+
786
|
| 472 |
+
],
|
| 473 |
+
"page_idx": 5
|
| 474 |
+
},
|
| 475 |
+
{
|
| 476 |
+
"type": "text",
|
| 477 |
+
"text": "In Section 4.1 and Section 4.2, we introduce the pre-training experiments of Point-M2AE and report the fine-tuning performance on various downstream tasks. We also conduct ablation studies in Section 4.3 to validate the effectiveness of our approach. ",
|
| 478 |
+
"bbox": [
|
| 479 |
+
174,
|
| 480 |
+
799,
|
| 481 |
+
825,
|
| 482 |
+
842
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 5
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "text",
|
| 488 |
+
"text": "4.1 Self-supervised Pre-training ",
|
| 489 |
+
"text_level": 1,
|
| 490 |
+
"bbox": [
|
| 491 |
+
176,
|
| 492 |
+
857,
|
| 493 |
+
408,
|
| 494 |
+
872
|
| 495 |
+
],
|
| 496 |
+
"page_idx": 5
|
| 497 |
+
},
|
| 498 |
+
{
|
| 499 |
+
"type": "text",
|
| 500 |
+
"text": "Settings. We pre-train our Point-M2AE on ShapeNet [6] dataset, which contains 57,448 synthetic 3D shapes of 55 categories. We set the stage number $S$ as 3, and construct a 3-stage encoder and a ",
|
| 501 |
+
"bbox": [
|
| 502 |
+
174,
|
| 503 |
+
882,
|
| 504 |
+
821,
|
| 505 |
+
911
|
| 506 |
+
],
|
| 507 |
+
"page_idx": 5
|
| 508 |
+
},
|
| 509 |
+
{
|
| 510 |
+
"type": "table",
|
| 511 |
+
"img_path": "images/10d053b9556291555d3bf2f681806cefc0bf8d2d1961351079013f015533f296.jpg",
|
| 512 |
+
"table_caption": [
|
| 513 |
+
"Table 3: Shape classification on ScanObjectNN [38]. We report the accuracy $( \\% )$ on the three splits of ScanObjectNN. [S] represents fine-tuning after self-supervised pre-training. "
|
| 514 |
+
],
|
| 515 |
+
"table_footnote": [],
|
| 516 |
+
"table_body": "<table><tr><td>Method</td><td>OBJ-BG</td><td>OBJ-ONLY</td><td>PB-T50-RS</td></tr><tr><td>PointNet [29]</td><td>73.3</td><td>79.2</td><td>68.0</td></tr><tr><td>PointNet++ [30]</td><td>82.3</td><td>84.3</td><td>77.9</td></tr><tr><td>DGCNN [41]</td><td>82.8</td><td>86.2</td><td>78.1</td></tr><tr><td>PointCNN [23]</td><td>86.1</td><td>85.5</td><td>78.5</td></tr><tr><td>Transformer [49]</td><td>79.86</td><td>80.55</td><td>77.24</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>84.85</td><td>85.54</td><td>78.79</td></tr><tr><td>[S] Point-BERT[49]</td><td>87.43</td><td>88.12</td><td>83.07</td></tr><tr><td>[S] Point-M2AE</td><td>91.22</td><td>88.81</td><td>86.43</td></tr><tr><td>Improvement</td><td>+3.79</td><td>+0.69</td><td>+3.36</td></tr></table>",
|
| 517 |
+
"bbox": [
|
| 518 |
+
258,
|
| 519 |
+
130,
|
| 520 |
+
736,
|
| 521 |
+
310
|
| 522 |
+
],
|
| 523 |
+
"page_idx": 6
|
| 524 |
+
},
|
| 525 |
+
{
|
| 526 |
+
"type": "text",
|
| 527 |
+
"text": "2-stage decoder for hierarchical learning. We adopt 5 blocks in each encoder stage, but only 1 block per stage for the lightweight decoder. For the 3-scale point clouds, we set the point numbers and token dimensions respectively as {512, 256, 64} and {96, 192, 384}. We also set different $k$ for the $k$ -NN at different scales, which are {16, 8, 8}. We mask the highest scale of point clouds with a high ratio of $80 \\%$ and set 6 heads for all the attention modules. The detailed training settings are in Appendix. ",
|
| 528 |
+
"bbox": [
|
| 529 |
+
173,
|
| 530 |
+
337,
|
| 531 |
+
825,
|
| 532 |
+
420
|
| 533 |
+
],
|
| 534 |
+
"page_idx": 6
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"type": "text",
|
| 538 |
+
"text": "Linear SVM. After pre-training on ShapeNet, we test the 3D representation capability of PointM2AE via linear evaluation on ModelNet40 [44]. We sample 1,024 points from each 3D shape of ModelNet40 and utilize our frozen encoder to extract their features. On top of that, we train a linear SVM and report the classification accuracy in Table 1. As shown, Point-M2AE achieves the best performance among all existing self-supervised methods for point clouds, and surpasses the second-best CrossPoint [2] by $+ 1 . 7 \\%$ . Point-M2AE also exceeds Point-BERT [49] by $+ 5 . 5 \\%$ , which is a masked point modeling method with a MoCo loss [19] but adopts a standard transformer and conducts single-scale learning. It is worth noting that even if we freeze all our parameters, Point-M2AE with $9 2 . 9 \\%$ accuracy still outperforms many fully trained methods on ModelNet40, e.g., $9 0 . 5 \\%$ by PointNet $^ { + + }$ [30], $9 2 . 8 \\%$ by DensePoint [24], etc. The experiments fully demonstrate the superior 3D representation capacity of our Point-M2AE. ",
|
| 539 |
+
"bbox": [
|
| 540 |
+
173,
|
| 541 |
+
433,
|
| 542 |
+
825,
|
| 543 |
+
585
|
| 544 |
+
],
|
| 545 |
+
"page_idx": 6
|
| 546 |
+
},
|
| 547 |
+
{
|
| 548 |
+
"type": "text",
|
| 549 |
+
"text": "4.2 Downstream Tasks ",
|
| 550 |
+
"text_level": 1,
|
| 551 |
+
"bbox": [
|
| 552 |
+
176,
|
| 553 |
+
603,
|
| 554 |
+
344,
|
| 555 |
+
617
|
| 556 |
+
],
|
| 557 |
+
"page_idx": 6
|
| 558 |
+
},
|
| 559 |
+
{
|
| 560 |
+
"type": "text",
|
| 561 |
+
"text": "For fine-tuning on downstream tasks, we discard the hierarchical decoder in pre-training and append different heads onto the hierarchical encoder for different tasks. ",
|
| 562 |
+
"bbox": [
|
| 563 |
+
174,
|
| 564 |
+
628,
|
| 565 |
+
823,
|
| 566 |
+
656
|
| 567 |
+
],
|
| 568 |
+
"page_idx": 6
|
| 569 |
+
},
|
| 570 |
+
{
|
| 571 |
+
"type": "text",
|
| 572 |
+
"text": "Shape Classification. We fine-tune Point-M2AE on two shape classification datasets: the widely adopted ModelNet40 [44] and the challenging ScanObjectNN [38]. For local spatial attention layers, we set the ball queries’ radii of 3-scale point clouds as {0.32, 0.64, 1.28}. We follow Point-BERT to use the voting strategy [25] for fair comparison on ModelNet40. To handle the noisy spatial structures, we increase $k$ of $k$ -NN into {32, 16, 16} for ScanObjectNN to encode local patterns with larger receptive fields. As reported in Table 2, Point-M2AE achieves $9 4 . 0 \\%$ accuracy on ModelNet40 with 1024 points per sample, which surpasses Point-BERT fine-tuned with 1024 points by $+ 0 . 8 \\%$ and 8192 points by $+ 0 . 2 \\%$ . For ScanObjectNN in Table 3, our Point-M2AE outperforms the secondbest Point-BERT by a significant margin, $+ 3 . 7 9 \\%$ , $+ 0 . 6 9 \\%$ and $+ 3 . 3 6 \\%$ , respectively for the three splits, indicating our great advantages under complex circumstances by multi-scale encoding. As ScanObjectNN of real-world scenes has a large semantic gap with the pre-trained synthetic ShapeNet, Point-M2AE also exerts strong transfer ability to understand point clouds of another domain. ",
|
| 573 |
+
"bbox": [
|
| 574 |
+
174,
|
| 575 |
+
672,
|
| 576 |
+
825,
|
| 577 |
+
839
|
| 578 |
+
],
|
| 579 |
+
"page_idx": 6
|
| 580 |
+
},
|
| 581 |
+
{
|
| 582 |
+
"type": "text",
|
| 583 |
+
"text": "Part Segmentation. We evaluate Point-M2AE for part segmentation on ShapeNetPart [48], which predicts per-point part labels and requires detailed understanding for local patterns. We adopt an extremely simple segmentation head to validate the effectiveness of our pre-training for well capturing both high-level semantics and fine-grained details. By the hierarchical encoder, we obtain ",
|
| 584 |
+
"bbox": [
|
| 585 |
+
176,
|
| 586 |
+
856,
|
| 587 |
+
823,
|
| 588 |
+
911
|
| 589 |
+
],
|
| 590 |
+
"page_idx": 6
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"type": "table",
|
| 594 |
+
"img_path": "images/e1d81688b9b97f951e5c18508db1c70bdbc15de5219e21e7a42ec3d1eabecf92.jpg",
|
| 595 |
+
"table_caption": [
|
| 596 |
+
"Table 4: Few-shot classification on ModelNet40 [44]. We report the average accuracy $( \\% )$ and standard deviation $( \\% )$ of 10 independent experiments. "
|
| 597 |
+
],
|
| 598 |
+
"table_footnote": [],
|
| 599 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">5-way</td><td colspan=\"2\">10-way</td></tr><tr><td>10-shot</td><td>20-shot</td><td>10-shot</td><td>20-shot</td></tr><tr><td>DGCNN [41]</td><td>91.8 ± 3.7</td><td>93.4 ± 3.2</td><td>86.3 ± 6.2</td><td>90.9 ± 5.1</td></tr><tr><td>[S] DGCNN + OcCo [40]</td><td>91.9 ± 3.3</td><td>93.9 ± 3.1</td><td>86.4 ± 5.4</td><td>91.3 ± 4.6</td></tr><tr><td>Transformer [49]</td><td>87.8 ±5.2</td><td>93.3 ± 4.3</td><td>84.6 ± 5.5</td><td>89.4 ± 6.3</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>94.0±3.6</td><td>95.9 ± 2.3</td><td>89.4 ± 5.1</td><td>92.4 ± 4.6</td></tr><tr><td>[S] Point-BERT[49]</td><td>94.6 ± 3.1</td><td>96.3 ± 2.7</td><td>91.0 ± 5.4</td><td>92.7 ± 5.1</td></tr><tr><td>[S] Point-M2AE</td><td>96.8 ± 1.8</td><td>98.3 ± 1.4</td><td>92.3 ± 4.5</td><td>95.0 ± 3.0</td></tr><tr><td>Improvement</td><td>+2.2</td><td>+2.0</td><td>+1.3</td><td>+2.3</td></tr></table>",
|
| 600 |
+
"bbox": [
|
| 601 |
+
236,
|
| 602 |
+
123,
|
| 603 |
+
756,
|
| 604 |
+
279
|
| 605 |
+
],
|
| 606 |
+
"page_idx": 7
|
| 607 |
+
},
|
| 608 |
+
{
|
| 609 |
+
"type": "table",
|
| 610 |
+
"img_path": "images/8d24e5bccef0424131fa70385ec0e1059951a03cad7bfb9a3ba5475560f578ca.jpg",
|
| 611 |
+
"table_caption": [
|
| 612 |
+
"Table 5: Part segmentation on ShapeNetPart [48]. $\\mathbf { \\hat { m } } \\mathbf { I o U } _ { C }$ ’ $( \\% )$ and $\\mathbf { \\dot { m l o U } } _ { I } ,$ $( \\% )$ denote the mean IoU across all part categories and all instances in the dataset, respectively. "
|
| 613 |
+
],
|
| 614 |
+
"table_footnote": [],
|
| 615 |
+
"table_body": "<table><tr><td>Method</td><td>mIoUc</td><td>mIoU1</td></tr><tr><td>PointNet [29] PointNet++ [30]</td><td>80.39 81.85</td><td>83.70 85.10</td></tr><tr><td>DGCNN [41]</td><td>82.33</td><td>85.20</td></tr><tr><td>Transformer [49]</td><td>83.42</td><td>85.10</td></tr><tr><td>[S] Transformer + OcCo [49]</td><td>83.42</td><td>85.10</td></tr><tr><td>[S] Point-BERT[49]</td><td>84.11</td><td>85.60</td></tr><tr><td>[S] Point-M2AE</td><td>84.86</td><td>86.51</td></tr><tr><td>Improvement</td><td>+0.75</td><td>+0.91</td></tr></table>",
|
| 616 |
+
"bbox": [
|
| 617 |
+
183,
|
| 618 |
+
363,
|
| 619 |
+
513,
|
| 620 |
+
507
|
| 621 |
+
],
|
| 622 |
+
"page_idx": 7
|
| 623 |
+
},
|
| 624 |
+
{
|
| 625 |
+
"type": "table",
|
| 626 |
+
"img_path": "images/4ea2bde4566908731cf820b6c7ac92360f464aae4b1ee8d1e2f7e7e4b5ba2888.jpg",
|
| 627 |
+
"table_caption": [
|
| 628 |
+
"Table 6: 3D object detection on ScanNetV2 [9]. We report the performance $( \\% )$ of self-supervised learning methods based on VoteNet [12] and 3DETR-m [26]. "
|
| 629 |
+
],
|
| 630 |
+
"table_footnote": [],
|
| 631 |
+
"table_body": "<table><tr><td>Method</td><td>AP25</td><td>AP50</td></tr><tr><td>VoteNet [12] [S] STRL [20] [S] PointContrast [45]</td><td>58.6 59.5 59.2</td><td>33.5 38.4 38.0</td></tr><tr><td>[S] DepthContrast [51] 3DETR[26]</td><td>61.3 62.1</td><td>1 37.9</td></tr><tr><td>3DETR-m [26]</td><td>65.0</td><td>47.0</td></tr><tr><td>[S] Point-M2AE</td><td>66.3</td><td>48.3</td></tr><tr><td>Improvement</td><td>+1.3</td><td>+1.3</td></tr></table>",
|
| 632 |
+
"bbox": [
|
| 633 |
+
539,
|
| 634 |
+
363,
|
| 635 |
+
808,
|
| 636 |
+
507
|
| 637 |
+
],
|
| 638 |
+
"page_idx": 7
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"type": "text",
|
| 642 |
+
"text": "3-scale point tokens of {512, 256, 64} points, and perform feature propagation in PointNet $^ { - + }$ [30] to independently upsample the tokens into 2048 points of the input point cloud. Then, we concatenate the upsampled 3-scale features for each point and predict the part label by stacked linear projection layers. As reported in Table 4.2, Point-M2AE achieves the best $8 6 . 5 1 \\%$ instance mIoU with the simple segmentation head, surpassing the second-best Point-BERT by $+ 0 . 9 1 \\%$ . Note that Point-BERT [49] and other methods [29, 30, 41] adopt hierarchical segmentation heads to progressively upsample the point features from intermediate layers, while our head contains no hierarchical structure and only relies on the pre-trained encoder to capture the multi-scale information of point clouds. The results fully demonstrate the significance of Point-M2AE’s multi-scale pre-training to segmentation tasks. ",
|
| 643 |
+
"bbox": [
|
| 644 |
+
173,
|
| 645 |
+
530,
|
| 646 |
+
825,
|
| 647 |
+
656
|
| 648 |
+
],
|
| 649 |
+
"page_idx": 7
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"text": "Few-shot Classification. We conduct experiments for few-shot classification on ModelNet40 [44] to evaluate the performance of Point-M2AE with limited fine-tuning data. As reported in Table 4.2, Point-M2AE achieves the best performance for all four settings, and surpasses Point-BERT by $+ 2 . 2 \\%$ , $+ 2 . 0 \\%$ , $+ 1 . 3 \\%$ , and $+ 2 . 7 \\%$ , respectively. Our approach also shows smaller deviations than other transformer-based methods, which indicates Point-M2AE has learned to produce more universal 3D representations for well adapting to downstream tasks under low-data regimes. ",
|
| 654 |
+
"bbox": [
|
| 655 |
+
173,
|
| 656 |
+
672,
|
| 657 |
+
825,
|
| 658 |
+
756
|
| 659 |
+
],
|
| 660 |
+
"page_idx": 7
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "text",
|
| 664 |
+
"text": "3D Object Detection To further evaluate our hierarchical pre-training on 3D object detection, we apply Point-M2AE to serving as the feature backbone on the indoor ScanNetV2 [9] dataset. We select 3DETR-m [26] as our baseline, which consists of a 3-block encoder and a transformer decoder. Considering the quite different dataset statistics, e.g., 2k input points for ShapeNet [6] and $5 0 \\mathrm { k }$ input points for ScanNetV2, we adopt the same encoder architecture with that of 3DETR-m, and keep our hierarchical decoder with skip connections unchanged for self-supervised pre-training on ScanNetV2. More details of models and training are in Appendix. As reported in Table 4.2, compared to training from scratch, our hierarchical pre-training boosts the performance of 3DETR-m by $+ 1 . 3 4 \\%$ $\\mathrm { A P _ { 2 5 } }$ and $+ 1 . 2 9 \\%$ $\\mathsf { A P } _ { 5 0 }$ . The experiments demonstrate the effectiveness of Point-M2AE to learn multi-scale point cloud encoding for object detection and its potential to benefit a wider range of 3D applications. ",
|
| 665 |
+
"bbox": [
|
| 666 |
+
174,
|
| 667 |
+
772,
|
| 668 |
+
826,
|
| 669 |
+
911
|
| 670 |
+
],
|
| 671 |
+
"page_idx": 7
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"type": "image",
|
| 675 |
+
"img_path": "images/d83a0bd222b589cf439c0f2002621535a464f8104af642b612bc9cad6f3d1937.jpg",
|
| 676 |
+
"image_caption": [
|
| 677 |
+
"Figure 4: Visualization of fine-grained information. We denote the outputs from hierarchical and non-hierarchical architectures as [NH] and [H], respectively. For an input point cloud (Middle), we visualize its extracted features (Left) and reconstruction results (Right). "
|
| 678 |
+
],
|
| 679 |
+
"image_footnote": [],
|
| 680 |
+
"bbox": [
|
| 681 |
+
173,
|
| 682 |
+
1,
|
| 683 |
+
823,
|
| 684 |
+
185
|
| 685 |
+
],
|
| 686 |
+
"page_idx": 8
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"type": "table",
|
| 690 |
+
"img_path": "images/6f2d8c6fb4a096090e1e7b52de9a2eb088ba0c5abd86c101cb8f1a3514828a88.jpg",
|
| 691 |
+
"table_caption": [
|
| 692 |
+
"Table 7: Hierarchical Modules. ‘H’ represents the encoder and decoder with multi-stage hierarchies. ‘Skip C.’ denotes the skip connections. "
|
| 693 |
+
],
|
| 694 |
+
"table_footnote": [],
|
| 695 |
+
"table_body": "<table><tr><td>Encoder</td><td>Decoder</td><td>Skip C.</td><td>Acc. (%)</td></tr><tr><td>H</td><td>H</td><td>√</td><td>92.9</td></tr><tr><td></td><td>1</td><td>√</td><td>90.7</td></tr><tr><td>1</td><td>H</td><td>√</td><td>91.5</td></tr><tr><td>H</td><td>1</td><td>√</td><td>92.2</td></tr><tr><td>H</td><td>H</td><td></td><td>92.1</td></tr></table>",
|
| 696 |
+
"bbox": [
|
| 697 |
+
181,
|
| 698 |
+
313,
|
| 699 |
+
493,
|
| 700 |
+
426
|
| 701 |
+
],
|
| 702 |
+
"page_idx": 8
|
| 703 |
+
},
|
| 704 |
+
{
|
| 705 |
+
"type": "table",
|
| 706 |
+
"img_path": "images/54bc8ec63e6b61f2a7c02f8d6b3bb30aec771338444edf34826fa3933ce10c5e.jpg",
|
| 707 |
+
"table_caption": [
|
| 708 |
+
"Table 8: Different Masking Strategy. ‘MS Mask’ and ‘Ratio’ denote the multi-scale masking and the mask ratio. "
|
| 709 |
+
],
|
| 710 |
+
"table_footnote": [],
|
| 711 |
+
"table_body": "<table><tr><td>MS Mask</td><td>Ratio</td><td>Acc. (%)</td></tr><tr><td>了</td><td>0.8</td><td>92.9</td></tr><tr><td>1</td><td>0.8</td><td>88.4</td></tr><tr><td>√</td><td>0.6</td><td>92.3</td></tr><tr><td><</td><td>0.7</td><td>92.7</td></tr><tr><td>√</td><td>0.9</td><td>92.5</td></tr></table>",
|
| 712 |
+
"bbox": [
|
| 713 |
+
552,
|
| 714 |
+
313,
|
| 715 |
+
781,
|
| 716 |
+
426
|
| 717 |
+
],
|
| 718 |
+
"page_idx": 8
|
| 719 |
+
},
|
| 720 |
+
{
|
| 721 |
+
"type": "text",
|
| 722 |
+
"text": "4.3 Ablation Study ",
|
| 723 |
+
"text_level": 1,
|
| 724 |
+
"bbox": [
|
| 725 |
+
174,
|
| 726 |
+
449,
|
| 727 |
+
318,
|
| 728 |
+
464
|
| 729 |
+
],
|
| 730 |
+
"page_idx": 8
|
| 731 |
+
},
|
| 732 |
+
{
|
| 733 |
+
"type": "text",
|
| 734 |
+
"text": "We conduct ablation study by modifying one of the components at a time during pre-training and explore the best masking strategy. We report the classification accuracy on ModelNet40 [44] by linear SVM to evaluate the pre-trained representations. For downstream tasks, we train the network from scratch to validate the significance of our hierarchical pre-training. ",
|
| 735 |
+
"bbox": [
|
| 736 |
+
174,
|
| 737 |
+
477,
|
| 738 |
+
825,
|
| 739 |
+
532
|
| 740 |
+
],
|
| 741 |
+
"page_idx": 8
|
| 742 |
+
},
|
| 743 |
+
{
|
| 744 |
+
"type": "text",
|
| 745 |
+
"text": "Hierarchical Modules. As reported in Table 7, on top of our final solution, Point-M2AE, in the first row, we respectively experiment with removing the hierarchical encoder, hierarchical decoder, and skip connections from our framework. Specifically, we replace our encoder and decoder with 1-stage plain architectures similar to MAE, which contains 15 and 2 vanilla transformer blocks, respectively. We observe the absence of multi-stage structures either in encoder or decoder hurts the performance, and the hierarchical encoder plays a better role than the decoder. Also, the skip connections well benefits the accuracy by providing complementary information for the decoder. ",
|
| 746 |
+
"bbox": [
|
| 747 |
+
173,
|
| 748 |
+
553,
|
| 749 |
+
825,
|
| 750 |
+
650
|
| 751 |
+
],
|
| 752 |
+
"page_idx": 8
|
| 753 |
+
},
|
| 754 |
+
{
|
| 755 |
+
"type": "text",
|
| 756 |
+
"text": "Masking Strategy. In Table 8, we report Point-M2AE with different mask settings. Without the multi-scale masking, we randomly generate masks at each scale, which leads to fragmented visible regions for all scales. With this strategy, the network would ‘peek’ different parts of the point cloud at different stages, which disturbs the representation learning and harms the performance by $- 4 . 5 \\%$ accuracy. For different mask ratios, we find the $8 0 \\%$ ratio performs the best to build a properly challenging pretext task for self-supervised pre-training. ",
|
| 757 |
+
"bbox": [
|
| 758 |
+
173,
|
| 759 |
+
669,
|
| 760 |
+
825,
|
| 761 |
+
753
|
| 762 |
+
],
|
| 763 |
+
"page_idx": 8
|
| 764 |
+
},
|
| 765 |
+
{
|
| 766 |
+
"type": "text",
|
| 767 |
+
"text": "With and Without Pre-training. We report the performance of Point-M2AE on downstream tasks with and without the pre-training in Table 9. For ‘w/o’, we randomly initialize the parameters and train the network from scratch. As shown, the pre-training can largely boost the performance on four datasets respectively by $+ 1 . 5 \\%$ , $+ 2 . 5 \\%$ , $+ 3 . 8 \\%$ , and $+ 1 . 1 \\bar { \\% }$ , which indicates the superiority and significance of our hierarchical pre-training. ",
|
| 768 |
+
"bbox": [
|
| 769 |
+
174,
|
| 770 |
+
772,
|
| 771 |
+
485,
|
| 772 |
+
911
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 8
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "table",
|
| 778 |
+
"img_path": "images/86571a688ddc98648184788e97e40b241d191c8f832c4c878d9dd6b6eccc667d.jpg",
|
| 779 |
+
"table_caption": [
|
| 780 |
+
"Table 9: With and without the pre-training. ‘ModelNet40-FS’ denotes the few-shot classification on 10-way 20-shot ModelNet40 [44]. "
|
| 781 |
+
],
|
| 782 |
+
"table_footnote": [],
|
| 783 |
+
"table_body": "<table><tr><td>Dataset</td><td>w/o (%)</td><td>w (%)</td></tr><tr><td>ModelNet40 [44]</td><td>92.5</td><td>94.0</td></tr><tr><td>ScanObjectNN[38]</td><td>83.9</td><td>86.4</td></tr><tr><td>ModelNet40-FS [44]</td><td>91.2</td><td>95.0</td></tr><tr><td>ShapeNetPart [48]</td><td>85.4</td><td>86.5</td></tr></table>",
|
| 784 |
+
"bbox": [
|
| 785 |
+
508,
|
| 786 |
+
816,
|
| 787 |
+
807,
|
| 788 |
+
911
|
| 789 |
+
],
|
| 790 |
+
"page_idx": 8
|
| 791 |
+
},
|
| 792 |
+
{
|
| 793 |
+
"type": "image",
|
| 794 |
+
"img_path": "images/4b7ee7b02566fc0abb0850c788e6552a429abeea9d9b8a45a9cb93270f865658.jpg",
|
| 795 |
+
"image_caption": [
|
| 796 |
+
"Figure 5: Visualization of multi-scale point clouds. In each row, we visualize the input point clouds, their multi-scale representations, the reconstructed coordinates, and multi-scale masked point clouds. "
|
| 797 |
+
],
|
| 798 |
+
"image_footnote": [],
|
| 799 |
+
"bbox": [
|
| 800 |
+
169,
|
| 801 |
+
88,
|
| 802 |
+
818,
|
| 803 |
+
436
|
| 804 |
+
],
|
| 805 |
+
"page_idx": 9
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "text",
|
| 809 |
+
"text": "5 Visualization ",
|
| 810 |
+
"text_level": 1,
|
| 811 |
+
"bbox": [
|
| 812 |
+
174,
|
| 813 |
+
497,
|
| 814 |
+
313,
|
| 815 |
+
513
|
| 816 |
+
],
|
| 817 |
+
"page_idx": 9
|
| 818 |
+
},
|
| 819 |
+
{
|
| 820 |
+
"type": "text",
|
| 821 |
+
"text": "Multi-scale Masking. To ease the understanding of our multi-scale masking strategy, we visualize the input point cloud, the 3-scale representations, the reconstructed point cloud, and 3-scale masked point clouds, respectively in each row of Figure 5. As shown, different scales can represent different levels of geometric details and semantics for point clouds. By the multi-scale masking strategy, we observe the visible positions of masked point clouds are block-wise within one scale and consistent across scales, which is significant for our hierarchical pre-training. ",
|
| 822 |
+
"bbox": [
|
| 823 |
+
174,
|
| 824 |
+
527,
|
| 825 |
+
825,
|
| 826 |
+
612
|
| 827 |
+
],
|
| 828 |
+
"page_idx": 9
|
| 829 |
+
},
|
| 830 |
+
{
|
| 831 |
+
"type": "text",
|
| 832 |
+
"text": "Fine-grained Information. The fine-grained 3D structures, e.g., thin branches of a plant, fingers of a human, engines of a plane, are significant to distinguish similar shapes and can be well encoded by our hierarchical representations. In Figure 4, we compare our Point-M2AE with multi-stage, [H], and single-scale, [NH], architectures by visualizing their extracted point features and reconstructed point clouds during pre-training. In contarst to the single-scale network, the multi-scale one indicates higher feature responses in the fine-grained structures and reconstructs more accurate spatial details. ",
|
| 833 |
+
"bbox": [
|
| 834 |
+
174,
|
| 835 |
+
622,
|
| 836 |
+
825,
|
| 837 |
+
705
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 9
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "text",
|
| 843 |
+
"text": "6 Conclusion ",
|
| 844 |
+
"text_level": 1,
|
| 845 |
+
"bbox": [
|
| 846 |
+
174,
|
| 847 |
+
720,
|
| 848 |
+
299,
|
| 849 |
+
738
|
| 850 |
+
],
|
| 851 |
+
"page_idx": 9
|
| 852 |
+
},
|
| 853 |
+
{
|
| 854 |
+
"type": "text",
|
| 855 |
+
"text": "We propose Point-M2AE, a multi-scale masked autoencoder for self-supervised pre-training on 3D point clouds. With a hierarchical architecture, Point-M2AE learns to produce powerful 3D representations by encoding multi-scale point clouds and reconstructing the masked coordinates from a global-to-local upsampling scheme. Extensive experiments have demonstrated the superiority of Point-M2AE to be a strong 3D representation learner. For limitations and future work, we will focus on applying Point-M2AE for wider 3D applications, e.g., outdoor and open-world scene understanding. We do not foresee negative social impact from the proposed work. ",
|
| 856 |
+
"bbox": [
|
| 857 |
+
174,
|
| 858 |
+
752,
|
| 859 |
+
825,
|
| 860 |
+
849
|
| 861 |
+
],
|
| 862 |
+
"page_idx": 9
|
| 863 |
+
},
|
| 864 |
+
{
|
| 865 |
+
"type": "text",
|
| 866 |
+
"text": "Acknowledgement. This work is supported by the National Natural Science Foundation of China (Grant No. 62206272), Shanghai Committee of Science and Technology (Grant No. 21DZ1100100), Centre for Perceptual and Interactive Intelligence Limited, and the General Research Fund through the Research Grants Council of Hong Kong (Grant No. 14204021, 14207319). ",
|
| 867 |
+
"bbox": [
|
| 868 |
+
176,
|
| 869 |
+
856,
|
| 870 |
+
825,
|
| 871 |
+
911
|
| 872 |
+
],
|
| 873 |
+
"page_idx": 9
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "text",
|
| 877 |
+
"text": "References ",
|
| 878 |
+
"text_level": 1,
|
| 879 |
+
"bbox": [
|
| 880 |
+
174,
|
| 881 |
+
90,
|
| 882 |
+
266,
|
| 883 |
+
106
|
| 884 |
+
],
|
| 885 |
+
"page_idx": 10
|
| 886 |
+
},
|
| 887 |
+
{
|
| 888 |
+
"type": "text",
|
| 889 |
+
"text": "[1] Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 123–133, 2021. 3 \n[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. arXiv preprint arXiv:2203.00680, 2022. 2, 3, 6, 7 \n[3] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. arXiv preprint arXiv:2202.03555, 2022. 1, 3 \n[4] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3 \n[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020. 1, 3 \n[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 6, 8 \n[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020. 1, 3 \n[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750–15758, 2021. 1, 3 \n[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 3, 8 \n[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1, 3 \n[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 \n[12] Zhipeng Ding, Xu Han, and Marc Niethammer. Votenet: A deep learning label fusion method for multi-atlas segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 202–210. Springer, 2019. 8 \n[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3 \n[14] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017. 6 \n[15] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187–199, 2021. 6 \n[16] Zhizhong Han, Mingyang Shang, Yu-Shen Liu, and Matthias Zwicker. View inter-prediction gan: Unsupervised representation learning for 3d shapes by learning global shape memories to support local view predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8376–8384, 2019. 6 \n[17] Zhizhong Han, Xiyang Wang, Yu-Shen Liu, and Matthias Zwicker. Multi-angle point cloudvae: Unsupervised feature learning for 3d point clouds from multiple angles by joint selfreconstruction and half-to-half prediction. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10441–10450. IEEE, 2019. 6 \n[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. 1, 3 \n[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020. 1, 3, 7 [20] Siyuan Huang, Song ervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6535–6545, 2021. 8 [21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, YunHsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages \n4904–4916. PMLR, 2021. 1 [22] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 6 [23] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on $\\mathbf { X }$ -transformed points. Advances in neural information processing systems, \n31:820–830, 2018. 6, 7 [24] Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud processing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5239–5248, \n2019. 7 [25] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8895–8904, 2019. 7 [26] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2906–2917, October 2021. 8 [27] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. arXiv preprint arXiv:2203.06604, 2022. \n3 [28] Omid Poursaeed, Tianxing Jiang, Han Qiao, Nayun Xu, and Vladimir G Kim. Self-supervised learning of point clouds via orientation estimation. In 2020 International Conference on 3D Vision (3DV), pages 1018–1028. IEEE, 2020. 3 [29] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660, 2017. 5, 6, 7, 8 [30] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 5, 6, 7, 8 [31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021. 1 [32] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 1, 3 [33] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1, 3 [34] Jason Tyler Rolfe. Discrete variational autoencoders. arXiv preprint arXiv:1609.02200, 2016. 3 [35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. 2, 4 [36] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 3 [37] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. Advances in Neural Information Processing Systems, 32, 2019. 6 [38] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1588–1597, 2019. 2, 7, 9 [39] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized representations of point clouds with graph-convolutional generative adversarial networks. IEEE Transactions on Multimedia, 23:402–414, 2020. 6 [40] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9782–9792, 2021. 3, 6, 8 \n[41] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1–12, 2019. 6, 7, 8 \n[42] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. arXiv preprint arXiv:2112.09133, 2021. 3 \n[43] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 6 \n[44] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912–1920, 2015. 2, 6, 7, 8, 9 \n[45] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In European conference on computer vision, pages 574–591. Springer, 2020. 3, 8 \n[46] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. 1 \n[47] Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Foldingnet: Point cloud auto-encoder via deep grid deformation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 206–215, 2018. 6 \n[48] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1–12, 2016. 2, 7, 8, 9 \n[49] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Pointbert: Pre-training 3d point cloud transformers with masked point modeling. arXiv preprint arXiv:2111.14819, 2021. 3, 6, 7, 8 \n[50] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413, 2021. 1 \n[51] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10252–10263, 2021. 3, 8 \n[52] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259– 16268, 2021. 6 \n[53] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 3 ",
|
| 890 |
+
"bbox": [
|
| 891 |
+
173,
|
| 892 |
+
136,
|
| 893 |
+
828,
|
| 894 |
+
912
|
| 895 |
+
],
|
| 896 |
+
"page_idx": 10
|
| 897 |
+
},
|
| 898 |
+
{
|
| 899 |
+
"type": "text",
|
| 900 |
+
"text": "",
|
| 901 |
+
"bbox": [
|
| 902 |
+
171,
|
| 903 |
+
101,
|
| 904 |
+
828,
|
| 905 |
+
893
|
| 906 |
+
],
|
| 907 |
+
"page_idx": 11
|
| 908 |
+
},
|
| 909 |
+
{
|
| 910 |
+
"type": "text",
|
| 911 |
+
"text": "",
|
| 912 |
+
"bbox": [
|
| 913 |
+
171,
|
| 914 |
+
92,
|
| 915 |
+
826,
|
| 916 |
+
617
|
| 917 |
+
],
|
| 918 |
+
"page_idx": 12
|
| 919 |
+
},
|
| 920 |
+
{
|
| 921 |
+
"type": "text",
|
| 922 |
+
"text": "Checklist ",
|
| 923 |
+
"text_level": 1,
|
| 924 |
+
"bbox": [
|
| 925 |
+
174,
|
| 926 |
+
640,
|
| 927 |
+
254,
|
| 928 |
+
656
|
| 929 |
+
],
|
| 930 |
+
"page_idx": 12
|
| 931 |
+
},
|
| 932 |
+
{
|
| 933 |
+
"type": "text",
|
| 934 |
+
"text": "1. For all authors... ",
|
| 935 |
+
"bbox": [
|
| 936 |
+
214,
|
| 937 |
+
666,
|
| 938 |
+
339,
|
| 939 |
+
681
|
| 940 |
+
],
|
| 941 |
+
"page_idx": 12
|
| 942 |
+
},
|
| 943 |
+
{
|
| 944 |
+
"type": "text",
|
| 945 |
+
"text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
|
| 946 |
+
"bbox": [
|
| 947 |
+
238,
|
| 948 |
+
685,
|
| 949 |
+
825,
|
| 950 |
+
776
|
| 951 |
+
],
|
| 952 |
+
"page_idx": 12
|
| 953 |
+
},
|
| 954 |
+
{
|
| 955 |
+
"type": "text",
|
| 956 |
+
"text": "2. If you are including theoretical results... ",
|
| 957 |
+
"bbox": [
|
| 958 |
+
214,
|
| 959 |
+
780,
|
| 960 |
+
493,
|
| 961 |
+
795
|
| 962 |
+
],
|
| 963 |
+
"page_idx": 12
|
| 964 |
+
},
|
| 965 |
+
{
|
| 966 |
+
"type": "text",
|
| 967 |
+
"text": "(a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes] ",
|
| 968 |
+
"bbox": [
|
| 969 |
+
236,
|
| 970 |
+
799,
|
| 971 |
+
735,
|
| 972 |
+
830
|
| 973 |
+
],
|
| 974 |
+
"page_idx": 12
|
| 975 |
+
},
|
| 976 |
+
{
|
| 977 |
+
"type": "text",
|
| 978 |
+
"text": "3. If you ran experiments... ",
|
| 979 |
+
"bbox": [
|
| 980 |
+
212,
|
| 981 |
+
835,
|
| 982 |
+
393,
|
| 983 |
+
849
|
| 984 |
+
],
|
| 985 |
+
"page_idx": 12
|
| 986 |
+
},
|
| 987 |
+
{
|
| 988 |
+
"type": "text",
|
| 989 |
+
"text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] ",
|
| 990 |
+
"bbox": [
|
| 991 |
+
238,
|
| 992 |
+
853,
|
| 993 |
+
825,
|
| 994 |
+
911
|
| 995 |
+
],
|
| 996 |
+
"page_idx": 12
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "text",
|
| 1000 |
+
"text": "",
|
| 1001 |
+
"bbox": [
|
| 1002 |
+
238,
|
| 1003 |
+
92,
|
| 1004 |
+
825,
|
| 1005 |
+
150
|
| 1006 |
+
],
|
| 1007 |
+
"page_idx": 13
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "text",
|
| 1011 |
+
"text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
|
| 1012 |
+
"bbox": [
|
| 1013 |
+
214,
|
| 1014 |
+
154,
|
| 1015 |
+
825,
|
| 1016 |
+
169
|
| 1017 |
+
],
|
| 1018 |
+
"page_idx": 13
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"type": "text",
|
| 1022 |
+
"text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [N/A] \n(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ",
|
| 1023 |
+
"bbox": [
|
| 1024 |
+
238,
|
| 1025 |
+
172,
|
| 1026 |
+
823,
|
| 1027 |
+
281
|
| 1028 |
+
],
|
| 1029 |
+
"page_idx": 13
|
| 1030 |
+
},
|
| 1031 |
+
{
|
| 1032 |
+
"type": "text",
|
| 1033 |
+
"text": "5. If you used crowdsourcing or conducted research with human subjects... ",
|
| 1034 |
+
"bbox": [
|
| 1035 |
+
214,
|
| 1036 |
+
284,
|
| 1037 |
+
705,
|
| 1038 |
+
299
|
| 1039 |
+
],
|
| 1040 |
+
"page_idx": 13
|
| 1041 |
+
},
|
| 1042 |
+
{
|
| 1043 |
+
"type": "text",
|
| 1044 |
+
"text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
|
| 1045 |
+
"bbox": [
|
| 1046 |
+
238,
|
| 1047 |
+
303,
|
| 1048 |
+
825,
|
| 1049 |
+
392
|
| 1050 |
+
],
|
| 1051 |
+
"page_idx": 13
|
| 1052 |
+
}
|
| 1053 |
+
]
|
parse/dev/MbCAOMGsZXC/MbCAOMGsZXC_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/MbCAOMGsZXC/MbCAOMGsZXC_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/O476oWmiNNp/O476oWmiNNp.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/O476oWmiNNp/O476oWmiNNp_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|