What Makes Good Synthetic Training Data for Zero-Shot Stereo Matching? (WMGStereo)

Paper | GitHub

WMGStereo is a procedural dataset generator specifically optimized for zero-shot stereo matching performance. This repository contains the WMGStereo-150k dataset, a large-scale synthetic training dataset featuring indoor, nature, and dense "flying" scenes.

Dataset Download

You can download the dataset using the huggingface-cli:

pip install huggingface-cli
huggingface-cli download pvl-lab/WMGStereo --repo-type dataset

Dataset Structure

The dataset file structure is as follows:

.
└── WMGStereo/
    ├── indoor/
    │   └── seed_num/
    │       └── frames/
    │           ├── Image/
    │           │   ├── camera_0
    │           │   └── camera_1
    │           ├── camview/
    │           │   ├── camera_0
    │           │   └── camera_1
    │           ├── disparity/
    │           │   └── camera_0
    │           ├── occ_mask/
    │           │   └── camera_0
    │           └── sky_mask/
    │               └── camera_0
    ├── flying/
    │   └── ...
    └── nature/
        └── ...

• Camera 0 and 1: correspond to left and right camera frames, respectively.
• Ground Truth: We provide disparity, occlusion, and sky-region masks for the left camera.
• camview: contains .npz files that contain a dictionary with indices K, T, HW, corresponding to calibration, translation, and resolution matrices.

Citation

If you find WMGStereo useful for your work, please consider citing the academic paper:

@misc{yan2025proceduraldatasetgenerationzeroshot,
      title={What Makes Good Synthetic Training Data for Zero-Shot Stereo Matching?}, 
      author={David Yan and Alexander Raistrick and Jia Deng},
      year={2025},
      eprint={2504.16930},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2504.16930}, 
}