spatial-training/objaverse_vida

Objaverse VIDA Dataset

Processed 3D assets from Objaverse for use with AI2-THOR and ProcTHOR.

Contents

Component	Description	Size
processed_2023_07_28/	~40K processed 3D objects with textures	27 GB
houses_2023_07_28/	160K+ procedurally generated house layouts	3.7 GB
procthor_databases_2023_07_28/	Asset databases, materials, placement rules	70 MB

Quick Start

# Install dependencies
pip install huggingface_hub[hf_transfer]

# Download the unpack script
wget https://huggingface.co/datasets/spatial-training/objaverse_vida/raw/main/unpack.py

# Run it (downloads ~30GB and extracts to ./objaverse_vida)
python unpack.py ./objaverse_vida

Or specify a custom path:

python unpack.py /path/to/destination

Data Format

Processed Objects

Each object directory contains:

{object_id}/
├── {object_id}.pkl.gz    # 3D mesh data (gzip pickle)
├── albedo.jpg            # Diffuse texture
├── normal.jpg            # Normal map
├── emission.jpg          # Emission map
└── thor_metadata.json    # AI2-THOR metadata

House Layouts

• train.jsonl.gz / test.jsonl.gz / val.jsonl.gz - Full house definitions
• train/ / test/ / val/ - Individual house JSON files

Asset Databases

• asset-database.json - Asset catalog
• material-database.json - Materials
• placement-annotations.json - Placement rules
• receptacles.json - Receptacle definitions

Manual Download

If you prefer not to use the script:

# Clone the dataset
huggingface-cli download spatial-training/objaverse_vida --repo-type dataset --local-dir ./objaverse_vida

# Extract processed shards
cd objaverse_vida/processed_2023_07_28
for f in shard_*.tar; do tar -xf "$f" && rm "$f"; done
rm manifest.json

# Extract house files
cd ../houses_2023_07_28
for split in train test val; do
  tar -xf "${split}_individual.tar" -C "$split/" && rm "${split}_individual.tar"
done

Citation

@inproceedings{deitke2023objaverse,
  title={Objaverse: A Universe of Annotated 3D Objects},
  author={Deitke, Matt and others},
  booktitle={CVPR},
  year={2023}
}

@inproceedings{deitke2022procthor,
  title={ProcTHOR: Large-Scale Embodied AI Using Procedural Generation},
  author={Deitke, Matt and others},
  booktitle={NeurIPS},
  year={2022}
}