Add files using upload-large-folder tool

ground_truth/ui_vision/analysis/prediction_results_gt_uivsion_analysis.json

@@ -0,0 +1,60 @@
+{
+  "total_samples": 30192,
+  "buckets": {
+    "[0.00MP, 2.00MP)": {
+      "count": 6019,
+      "accuracy": 0.17245389599601263,
+      "avg_latency": 0.7228303633944069,
+      "avg_tokens_per_s": 35.64420101106884,
+      "avg_gpu_mem": 7451.553949521178,
+      "avg_cpu_mem": 0.0,
+      "avg_dist_error": 325.9399871849913,
+      "avg_text_tokens": 1490.8657584316331
+    },
+    "[2.00MP, 4.00MP)": {
+      "count": 6580,
+      "accuracy": 0.19027355623100303,
+      "avg_latency": 0.7962758200172346,
+      "avg_tokens_per_s": 32.82464988539824,
+      "avg_gpu_mem": 7704.550098324112,
+      "avg_cpu_mem": -0.010238768047112462,
+      "avg_dist_error": 439.99651716602176,
+      "avg_text_tokens": 2749.0281155015196
+    },
+    "[4.00MP, 6.00MP)": {
+      "count": 5905,
+      "accuracy": 0.19949195596951735,
+      "avg_latency": 0.9608905909212336,
+      "avg_tokens_per_s": 27.606283359329687,
+      "avg_gpu_mem": 8224.805557145031,
+      "avg_cpu_mem": 0.0,
+      "avg_dist_error": 584.1049039158897,
+      "avg_text_tokens": 5332.177138018628
+    },
+    "[6.00MP, 8.00MP)": {
+      "count": 5899,
+      "accuracy": 0.1971520596711307,
+      "avg_latency": 1.133657247166254,
+      "avg_tokens_per_s": 23.547970196679127,
+      "avg_gpu_mem": 8737.519954471229,
+      "avg_cpu_mem": 0.1417904517714867,
+      "avg_dist_error": 719.2336327187414,
+      "avg_text_tokens": 7885.908628581115
+    },
+    "[8.00MP, 10.00MP)": {
+      "count": 5789,
+      "accuracy": 0.18690620141647954,
+      "avg_latency": 1.3251224298502797,
+      "avg_tokens_per_s": 20.181068190818337,
+      "avg_gpu_mem": 9253.31703657057,
+      "avg_cpu_mem": 0.0,
+      "avg_dist_error": 837.830567449717,
+      "avg_text_tokens": 10446.569873898774
+    }
+  },
+  "resize_impact": {},
+  "summary": {
+    "overall_accuracy": 0.18922231054583996,
+    "total_resized": 24403
+  }
+}

ground_truth/ui_vision/analysis/prediction_results_gt_uivsion_analysis.txt

@@ -0,0 +1,234 @@
+
+==========================================================================================
+EXECUTIVE SUMMARY
+==========================================================================================
+Total samples analyzed: 30192
+Resized samples: 24403 (80.8%)
+Overall accuracy: 0.189
+Unresized accuracy: 0.192
+
+Best accuracy: 0.199 at 4.0MP
+Most efficient: 1.0MP (best accuracy per compute)
+
+PERFORMANCE BY PIXEL BUCKET
+==========================================================================================
+Pixel Bucket            Count   Accuracy   Latency(s)   Tokens/s    GPU(MB)
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         6019      0.172        0.723       35.6     7451.6
+[2.00MP, 4.00MP)         6580      0.190        0.796       32.8     7704.6
+[4.00MP, 6.00MP)         5905      0.199        0.961       27.6     8224.8
+[6.00MP, 8.00MP)         5899      0.197        1.134       23.5     8737.5
+[8.00MP, 10.00MP)        5789      0.187        1.325       20.2     9253.3
+Unknown                     0      0.000        0.000        0.0        0.0
+
+RESIZE-ONLY PERFORMANCE BY PIXEL BUCKET
+==========================================================================================
+Pixel Bucket            Count   Accuracy   Latency(s)   Tokens/s    GPU(MB)
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         4738      0.174        0.723       35.7     7453.1
+[2.00MP, 4.00MP)         2626      0.197        0.799       32.7     7708.8
+[4.00MP, 6.00MP)         5593      0.195        0.961       27.6     8223.8
+[6.00MP, 8.00MP)         5657      0.192        1.134       23.5     8740.0
+[8.00MP, 10.00MP)        5789      0.187        1.325       20.2     9253.3
+Unknown                     0      0.000        0.000        0.0        0.0
+
+NO-RESIZE PERFORMANCE BY PIXEL BUCKET
+==========================================================================================
+Pixel Bucket            Count   Accuracy   Latency(s)   Tokens/s    GPU(MB)
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         1281      0.165        0.724       35.6     7445.9
+[2.00MP, 4.00MP)         3954      0.186        0.795       32.9     7701.7
+[4.00MP, 6.00MP)          312      0.288        0.967       27.4     8241.9
+[6.00MP, 8.00MP)          242      0.310        1.119       23.8     8680.4
+[8.00MP, 10.00MP)           0      0.000        0.000        0.0        0.0
+Unknown                     0      0.000        0.000        0.0        0.0
+
+RESIZE IMPACT ANALYSIS
+==========================================================================================
+
+UPSCALING (Lower → Higher Resolution)
+------------------------------------------------------------------------------------------
+Resize Direction        Count   Δ Accuracy   Δ Latency(s)   Δ Tokens/s
+------------------------------------------------------------------------------------------
+1MP→2MP                  1281       +0.000         -0.181         +4.5
+1MP→4MP                  1281       +0.006         -0.016         -0.6
+1MP→6MP                  1281       +0.004         +0.155         -4.7
+1MP→8MP                  1281       -0.005         +0.345         -8.0
+2MP→4MP                  3954       +0.008         -0.021         -0.5
+2MP→6MP                  3954       +0.006         +0.154         -4.6
+2MP→8MP                  3954       +0.002         +0.345         -8.0
+4MP→6MP                   312       +0.001         +0.144         -4.3
+4MP→8MP                   312       -0.035         +0.341         -7.8
+6MP→8MP                   242       -0.006         +0.329         -7.5
+
+DOWNSCALING (Higher → Lower Resolution)
+------------------------------------------------------------------------------------------
+Resize Direction        Count   Δ Accuracy   Δ Latency(s)   Δ Tokens/s
+------------------------------------------------------------------------------------------
+2MP→1MP                  3954       -0.019         -0.257         +7.5
+4MP→1MP                   312       -0.025         -0.267         +7.7
+4MP→2MP                   312       +0.013         -0.192         +4.9
+6MP→1MP                   242       -0.039         -0.279         +8.1
+6MP→2MP                   242       -0.002         -0.202         +5.2
+6MP→4MP                   242       +0.011         -0.038         -0.0
+
+SUMMARY:
+Upscaling operations: 10 resize patterns
+Downscaling operations: 6 resize patterns
+
+LOCALIZATION ERROR ANALYSIS
+==========================================================================================
+
+ABSOLUTE PIXEL ERROR - ALL PREDICTIONS
+Pixel Bucket            Count   Avg Error(px)    Median Error      Std Dev
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         6019           325.9           165.1        353.8
+[2.00MP, 4.00MP)         6580           440.0           214.9        487.7
+[4.00MP, 6.00MP)         5905           584.1           272.7        660.6
+[6.00MP, 8.00MP)         5899           719.2           316.5        836.5
+[8.00MP, 10.00MP)        5789           837.8           374.4        973.1
+
+ABSOLUTE PIXEL ERROR - INCORRECT PREDICTIONS ONLY
+Pixel Bucket            Count   Avg Error(px)    Median Error      Std Dev
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         4981           383.1           248.5        360.7
+[2.00MP, 4.00MP)         5328           528.8           352.7        497.9
+[4.00MP, 6.00MP)         4727           709.4           456.5        677.2
+[6.00MP, 8.00MP)         4736           871.1           515.3        861.3
+[8.00MP, 10.00MP)        4707          1002.9           594.6       1001.4
+
+NORMALIZED ERROR RATIO (ERROR/IMAGE_DIMENSION) - ALL PREDICTIONS
+Pixel Bucket            Count     Avg Ratio %  Median Ratio %    Std Dev %
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         6019           21.33           10.90        23.14
+[2.00MP, 4.00MP)         6580           20.45            9.96        22.62
+[4.00MP, 6.00MP)         5905           19.02            8.90        21.50
+[6.00MP, 8.00MP)         5899           19.16            8.38        22.27
+[8.00MP, 10.00MP)        5789           19.30            8.63        22.39
+
+NORMALIZED ERROR RATIO (%) - INCORRECT PREDICTIONS ONLY
+Pixel Bucket            Count     Avg Ratio %  Median Ratio %    Std Dev %
+------------------------------------------------------------------------------------------
+[0.00MP, 2.00MP)         4981           25.08           16.16        23.59
+[2.00MP, 4.00MP)         5328           24.58           16.35        23.09
+[4.00MP, 6.00MP)         4727           23.10           14.75        22.04
+[6.00MP, 8.00MP)         4736           23.22           13.73        22.93
+[8.00MP, 10.00MP)        4707           23.10           13.57        23.04
+
+SCALING LAWS ANALYSIS
+==========================================================================================
+Latency scaling: 0.685 × (megapixels)^0.29
+Memory scaling:  7305.5 × (megapixels)^0.10
+
+Interpretation:
+- Latency shows sublinear scaling (exponent=0.29)
+
+ILLUSTRATIVE EXAMPLES FOR MANUAL VERIFICATION
+==========================================================================================
+
+Bucket: [0.00MP, 2.00MP)
+------------------------------------------------------------
+  Example 1: element_grounding/22b8b830bc15b848d1ce4482d7e4403d.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 1.0MP (Resized: YES)
+    Error: 1504.4 pixels
+    Latency: 0.753s
+
+  Example 2: element_grounding/41ce086a03b9c1a9c79b0e2b950cd952.png
+    Result: ✗ INCORRECT
+    Original: 6.0MP → Resized: 1.0MP (Resized: YES)
+    Error: 154.5 pixels
+    Latency: 0.734s
+
+  Example 3: element_grounding/5c64f9e79d68374320b3dbccc525d32b.png
+    Result: ✓ CORRECT
+    Original: 2.0MP → Resized: 1.0MP (Resized: YES)
+    Error: 0.0 pixels
+    Latency: 0.754s
+
+
+Bucket: [2.00MP, 4.00MP)
+------------------------------------------------------------
+  Example 1: element_grounding/5be9de76d720d3291af343ebf7c6d1d3.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 2.0MP (Resized: NO)
+    Error: 2037.1 pixels
+    Latency: 0.773s
+
+  Example 2: element_grounding/954abfb07cfa543d81238a94641e49bf.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 2.0MP (Resized: NO)
+    Error: 195.5 pixels
+    Latency: 0.801s
+
+  Example 3: element_grounding/f51c978400b8a800b79a27424b5d73c4.png
+    Result: ✓ CORRECT
+    Original: 2.0MP → Resized: 2.0MP (Resized: NO)
+    Error: 0.0 pixels
+    Latency: 0.737s
+
+
+Bucket: [4.00MP, 6.00MP)
+------------------------------------------------------------
+  Example 1: element_grounding/9ac7d92f48439825c27b3eb62375489b.png
+    Result: ✗ INCORRECT
+    Original: 4.0MP → Resized: 4.0MP (Resized: NO)
+    Error: 2836.4 pixels
+    Latency: 1.032s
+
+  Example 2: element_grounding/9e676172ad02175c6cee8dce1ed702c9.png
+    Result: ✗ INCORRECT
+    Original: 1.0MP → Resized: 4.0MP (Resized: YES)
+    Error: 243.0 pixels
+    Latency: 0.946s
+
+  Example 3: element_grounding/d7ca612e6465f73feac5d38b91967f66.png
+    Result: ✓ CORRECT
+    Original: 4.0MP → Resized: 4.0MP (Resized: NO)
+    Error: 0.5 pixels
+    Latency: 0.927s
+
+
+Bucket: [6.00MP, 8.00MP)
+------------------------------------------------------------
+  Example 1: element_grounding/3985eda5ab6844f31de2dd353ff0ffe0.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 6.0MP (Resized: YES)
+    Error: 3617.8 pixels
+    Latency: 1.128s
+
+  Example 2: element_grounding/93865eda9a9bc0da084161b3baeec4be.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 6.0MP (Resized: YES)
+    Error: 280.7 pixels
+    Latency: 1.118s
+
+  Example 3: element_grounding/1072c1ef503a0d3220f791068db3d637.png
+    Result: ✓ CORRECT
+    Original: 2.0MP → Resized: 6.0MP (Resized: YES)
+    Error: 0.5 pixels
+    Latency: 1.082s
+
+
+Bucket: [8.00MP, 10.00MP)
+------------------------------------------------------------
+  Example 1: element_grounding/3985eda5ab6844f31de2dd353ff0ffe0.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 8.0MP (Resized: YES)
+    Error: 4183.6 pixels
+    Latency: 1.323s
+
+  Example 2: element_grounding/b41e14773e346550953a2676ce0c5be8.png
+    Result: ✗ INCORRECT
+    Original: 2.0MP → Resized: 8.0MP (Resized: YES)
+    Error: 339.2 pixels
+    Latency: 1.349s
+
+  Example 3: element_grounding/1c6e9319b349abb98c2a881e98bc4b0a.png
+    Result: ✓ CORRECT
+    Original: 2.0MP → Resized: 8.0MP (Resized: YES)
+    Error: 0.0 pixels
+    Latency: 1.378s
+
+Total examples selected: 15
+==========================================================================================

ground_truth/ui_vision/benchmark_summary_gt_uivsion.json

@@ -0,0 +1,87 @@
+{
+  "total_samples": 30192,
+  "correct": 5713,
+  "errors": 24484,
+  "accuracy": 0.18922231054583996,
+  "latency_avg_s": 0.981148901689561,
+  "latency_p50_s": 0.9517914992757142,
+  "latency_p95_s": 1.3430759178940206,
+  "avg_tokens_per_s": 28.129352984739107,
+  "avg_text_input_tokens": 5483.002119766826,
+  "avg_generated_tokens": 26.32203895071542,
+  "avg_gpu_mem_MB": 8254.653689244016,
+  "avg_cpu_mem_MB": 0.02546778756995728,
+  "oom_samples": 0,
+  "resolution_stats": {
+    "1.0mpx": {
+      "samples": 6019,
+      "correct": 1038,
+      "accuracy": 0.17245389599601263,
+      "avg_latency": 0.7228303633944069
+    },
+    "2.0mpx": {
+      "samples": 6580,
+      "correct": 1252,
+      "accuracy": 0.19027355623100303,
+      "avg_latency": 0.7962758200172346
+    },
+    "4.0mpx": {
+      "samples": 5905,
+      "correct": 1178,
+      "accuracy": 0.19949195596951735,
+      "avg_latency": 0.9608905909212336
+    },
+    "6.0mpx": {
+      "samples": 5899,
+      "correct": 1163,
+      "accuracy": 0.1971520596711307,
+      "avg_latency": 1.133657247166254
+    },
+    "8.0mpx": {
+      "samples": 5789,
+      "correct": 1082,
+      "accuracy": 0.18690620141647954,
+      "avg_latency": 1.3251224298502797
+    }
+  },
+  "variant_stats": {
+    "element_basic": {
+      "samples": 9260,
+      "correct": 2298,
+      "accuracy": 0.24816414686825053,
+      "avg_latency": 0.9715121370901154
+    },
+    "element_functional": {
+      "samples": 9260,
+      "correct": 1765,
+      "accuracy": 0.19060475161987042,
+      "avg_latency": 0.9791719729506227
+    },
+    "element_spatial": {
+      "samples": 10031,
+      "correct": 512,
+      "accuracy": 0.05104177051141461,
+      "avg_latency": 0.9984707653792855
+    },
+    "layout": {
+      "samples": 1641,
+      "correct": 1138,
+      "accuracy": 0.6934795856185253,
+      "avg_latency": 0.9407998374859092
+    }
+  },
+  "notes": {
+    "mpx_buckets": [
+      1.0,
+      2.0,
+      4.0,
+      6.0,
+      8.0
+    ],
+    "model_path": "/share_1/users/eddie/models/Qwen2.5-VL-3B-UI-R1-E",
+    "test_name": "gt_uivsion",
+    "dataset_name": "ui_vision",
+    "ui_vision_root": "/share_1/users/eddie/data/ui-vision",
+    "max_batch_size": 1
+  }
+}