FFS Stereo4D

[Project Page] [Paper] [Code]

Disparity maps for stereo matching, generated from the Stereo4D dataset using FoundationStereo.

Dataset Structure

data/train/
  metadata.csv
  0000000.zip   (first 50,000 images)
  0000001.zip   (next 50,000 images)
  ...
  0000025.zip

Each zip contains disparity PNG files named {vid_id}_frame_{frame_idx:06d}.png.

• Disparity images: 3-channel uint8 784×784 PNG files encoding per-pixel disparity. Decode with: disp = (R * 255*255 + G * 255 + B) / 1000.0. See also: https://github.com/NVlabs/FoundationStereo/blob/master/scripts/vis_dataset.py
• metadata.csv: Links each disparity image back to its source YouTube video, with a zip_file column indicating which zip contains the image.

Metadata Columns

Retrieving Source RGB Frames

This dataset contains disparity maps only. Due to the copyrights of these videos, users need to download on your own behalf. The corresponding left/right RGB stereo pairs can be recovered by:

1. Following stereo4d toolkit to download the YouTube video using youtube_video_id.
2. Seek to timestamp_sec (or video_frame_index) to locate the source frame.
3. Apply equirectangular rectification using the Stereo4D calibration .npz files to obtain the left and right perspective images.

Generation Pipeline

1. Source: YouTube 360 videos from the Stereo4D dataset.
2. Rectification: Equirectangular frames are rectified and cropped to 1024×1024 perspective stereo pairs.
3. Disparity estimation: FoundationStereo computes dense disparity at 784×784 resolution (resized by scale=0.765625 of the 1024×1024 input).

Camera Parameters

The rectified stereo pairs are generated at 1024×1024 with the following pinhole camera model:

Depth is derived as: depth = fx * baseline / disparity.

Since disparity is computed at 784×784 resolution (scale factor 784/1024 = 0.765625 of the 1024×1024 input), use the 784×784 camera parameters when converting disparity to depth:

import numpy as np
hfov = 60  # degrees
baseline = 0.063  # meters
imw = 784
fx = imw * 0.5 / np.tan(0.5 * np.radians(hfov))  # 678.8 px
depth = fx * baseline / disparity

Citation

If you use this dataset, please consider cite:

@article{wen2026fastfoundationstereo,
  title={Fast-FoundationStereo: Real-Time Zero-Shot Stereo Matching},
  author={Bowen Wen and Shaurya Dewan and Stan Birchfield},
  journal={CVPR},
  year={2026}
}
@article{wen2025foundationstereo,
  title={FoundationStereo: Zero-Shot Stereo Matching},
  author={Wen, Bowen and Trepte, Matthew and Aribido, Joseph and Kautz, Jan and Birchfield, Stan and Wan, Yao},
  journal={CVPR},
  year={2025}
}
@inproceedings{jin2025stereo4d,
  title={{Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos}},
  author={Jin, Linyi and Tucker, Richard and Li, Zhengqi and Fouhey, David and Snavely, Noah and Holynski, Aleksander},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2025},
}