metadata
license: apache-2.0
3D-IDE Training Data
Complete, verified training data for 3D-IDE (CVPR 2026). This is the exact data used in our experiments — byte-identical to our training server. Use md5_manifest.txt / md5_manifest_scannet.txt to verify your local copy without re-downloading.
ScanNet notice:
scannet/archives and parts of the metadata are derived from ScanNet. By downloading them you agree to the ScanNet Terms of Use.
Contents
| File | Size | Contents (archive-internal paths) | Extract into |
|---|---|---|---|
3d_ide_data.zip |
4.3 GB | processed/, metadata/, benchmark/, balanced/, embodiedscan/, multi.yaml |
data/ |
scannet/scannet_posed_images_part1.zip |
~25 GB | posed_images/scene0000_00 … (first half) |
data/scannet/ |
scannet/scannet_posed_images_part2.zip |
~25 GB | posed_images/… (second half) |
data/scannet/ |
scannet/scannet_mask_pcd.zip |
~10 GB | mask/, pcd_with_object_aabbs/ |
data/scannet/ |
vggt_features/vggt_features_part{1..9}.zip |
9 × ~43 GB | scene…/vggt_sliced.npy (no wrapper dir!) |
data/scannet/posed_images_3d_feature_vggt/ |
md5_manifest.txt |
— | MD5 of every file in 3d_ide_data.zip |
— |
md5_manifest_scannet.txt |
— | MD5 of every raw file under scannet/ |
— |
The loose processed/, metadata/, benchmark/, balanced/ folders in this repo mirror the contents of 3d_ide_data.zip for convenient single-file access.
Extraction
Run from the repo root of 3D-IDE:
cd 3D-IDE
# 1. main data
unzip 3d_ide_data.zip -d data/
# 2. scannet raw data
unzip scannet_posed_images_part1.zip -d data/scannet/
unzip scannet_posed_images_part2.zip -d data/scannet/
unzip scannet_mask_pcd.zip -d data/scannet/
# 3. VGGT features (note: archives contain scene dirs directly, no wrapper)
mkdir -p data/scannet/posed_images_3d_feature_vggt
for i in 1 2 3 4 5 6 7 8 9; do
unzip vggt_features_part$i.zip -d data/scannet/posed_images_3d_feature_vggt/
done
Verify (instead of re-downloading)
If you already prepared data and want to check it matches ours exactly:
cd 3D-IDE/data
md5sum -c md5_manifest.txt # json / metadata / embodiedscan
md5sum -c md5_manifest_scannet.txt # posed_images / mask / pcd
Any FAILED line points at a file that differs from our training copy.
Not included (download separately)
| Item | Where |
|---|---|
data/models/LLaVA-Video-7B-Qwen2/ |
lmms-lab/LLaVA-Video-7B-Qwen2 |
VGGT_checkpoints/model.pt |
facebook/VGGT-1B |
Final layout
3D-IDE/
├── data/
│ ├── balanced/
│ ├── benchmark/
│ ├── embodiedscan/
│ ├── metadata/
│ ├── models/LLaVA-Video-7B-Qwen2/ # separate download
│ ├── processed/
│ ├── multi.yaml
│ └── scannet/
│ ├── mask/
│ ├── pcd_with_object_aabbs/
│ ├── posed_images/
│ └── posed_images_3d_feature_vggt/
└── VGGT_checkpoints/model.pt # separate download