tracywei/scannet_3drae

ScanNet Preprocessed for 3DRAE

This repository contains the processed ScanNet data used by the released 3DRAE training pipeline. It stores RGB frames together with calibrated camera intrinsics and extrinsics in the directory format expected by the official code.

Links

• Paper
• Project page
• Code
• Model weights

Download and Extraction

The dataset is stored as split TAR parts. Download all parts with the Hugging Face CLI:

hf download tracywei/scannet_3drae \
  --repo-type dataset \
  --include "scannet.tar.part-*" \
  --local-dir data/scannet_3drae_download

Concatenate the parts in filename order and stream the archive directly to tar:

mkdir -p data
cat data/scannet_3drae_download/scannet.tar.part-* | \
  tar --numeric-owner -xpf - -C data/

The archive extracts to data/scannet, which is the default path used by the released training configurations.

Data Structure

data/scannet/
├── train/
│   └── <scene_name>/
│       ├── images/
│       │   ├── 000000.png
│       │   ├── 000001.png
│       │   └── ...
│       ├── intrinsics.npy
│       ├── extrinsics.npy
│       ├── frame_ids.npy
│       └── metadata.json
└── test/
    └── <scene_name>/...

The test directory is constructed from the official ScanNet v2 validation scene list.

Each scene contains:

• images/*.png: RGB frames, indexed consecutively after preprocessing.
• intrinsics.npy: per-frame camera intrinsic matrices with shape [N, 3, 3].
• extrinsics.npy: per-frame OpenCV world-to-camera matrices with shape [N, 4, 4].
• frame_ids.npy: original ScanNet frame IDs corresponding to the processed frames.
• metadata.json: scene name, split, sequence length, image resolution, filename format, and other loading metadata.

This release contains RGB images and camera parameters. It does not contain depth maps or semantic annotations.

Preprocessing

The dataset was produced from the official ScanNet v2 export using tools/prepare_scannet.py. The preprocessing procedure:

1. Matches RGB frames with valid camera poses.
2. Converts the original camera-to-world poses to OpenCV world-to-camera matrices.
3. Resizes RGB images so that the longest side is at most 560 pixels while preserving aspect ratio.
4. Updates camera intrinsics according to the resized image resolution.
5. Saves RGB frames as PNG and stores camera parameters as NumPy arrays.

To reproduce the released format from an authorized ScanNet download:

python tools/prepare_scannet.py \
  --scans-root <scannet_root>/scans \
  --train-list <scannet_root>/Tasks/Benchmark/scannetv2_train.txt \
  --test-list <scannet_root>/Tasks/Benchmark/scannetv2_val.txt \
  --output data/scannet \
  --workers 8

The source scenes must first be exported with color/, pose/, and intrinsic/intrinsic_color.txt.

Usage

Clone the official implementation and place the extracted dataset at data/scannet:

git clone https://github.com/WU-CVGL/3DRAE.git
cd 3DRAE

The released Stage 1 and Stage 2 training configurations already use this path. See the 3DRAE repository for training commands and configuration details.

License and Terms of Use

This release is derived from ScanNet and remains subject to the ScanNet Terms of Use. Users are responsible for obtaining ScanNet access and complying with its terms, including its restrictions on usage and data sharing.

Citation

If you use this processed dataset, please cite both 3DRAE and ScanNet:

@article{wei20263drae,
  author  = {Wei, Dongxu and Xu, Qi and Li, Zhiqi and Zhou, Hangning and Qiu, Cong and Qin, Hailong and Yang, Mu and Cui, Zhaopeng and Liu, Peidong},
  title   = {Any 3D Scene is Worth 1K Tokens: 3D-Grounded Representation for Scene Generation at Scale},
  journal = {arXiv},
  year    = {2026}
}

@inproceedings{dai2017scannet,
  title     = {ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes},
  author    = {Dai, Angela and Chang, Angel X. and Savva, Manolis and Halber, Maciej and Funkhouser, Thomas and Nie{\ss}ner, Matthias},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year      = {2017}
}