UnrealStereo4K-Q Dataset

This repository contains a quarter-resolution (Q) version of the UnrealStereo4K dataset originally presented in the paper SMD-Nets: Stereo Mixture Density Networks (CVPR 2021) by Fabio Tosi, Yiyi Liao, Carolin Schmitt and Andreas Geiger.

Overview

UnrealStereo4K-Q is a downsampled version (quarter resolution) of the original high-resolution stereo dataset. It maintains the core structure but with reduced resolution, making it more manageable for experiments that don't require full 4K resolution.

The original dataset can be found at: https://github.com/fabiotosi92/SMD-Nets

Dataset Structure

The stereo dataset is organized into 9 separate zip files, one for each scene:

00000.zip, 00001.zip, 00002.zip, 00003.zip, 00004.zip, 00005.zip, 00006.zip, 00007.zip, 00008.zip

Each zip file contains the following structure:

scene_number/
├── Image0/         # Left stereo images (JPG format)
├── Image1/         # Right stereo images (JPG format)
├── Disp0/          # Left disparity maps (NPY format)
└── Disp1/          # Right disparity maps (NPY format)

Dataset Details

• Resolution: 960×540 (1/4 of the original 3840×2160)
• Image Format: JPG
• Disparity Format: NPY (NumPy arrays)
• Scenes: 9 different environments (8 for training/testing, 1 for out-of-domain testing)

Qualitative Examples

Below are some examples of the stereo images (left view) and their corresponding disparity maps:

Important Notes

1. Disparity values: The disparity values are scaled to match the quarter resolution. For example, a disparity of 40 pixels in the original dataset would be 10 pixels in this version.

2. Simplified structure: This version includes only the stereo images and disparity maps, without the extrinsics that were available in the original dataset.

3. Scene 00008: Contains 200 stereo pairs only, used as out-of-domain test set in the original paper.

Citation

If you use this dataset in your research, please cite the original paper:

@INPROCEEDINGS{Tosi2021CVPR,
  author = {Fabio Tosi and Yiyi Liao and Carolin Schmitt and Andreas Geiger},
  title = {SMD-Nets: Stereo Mixture Density Networks},
  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2021}
} 

Usage

To load the disparity maps:

import numpy as np

# Load disparity map
disp = np.load("00000/Disp0/00000.npy")

To load and display an image:

import cv2
import matplotlib.pyplot as plt

# Load image
img = cv2.imread("00000/Image0/00000.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# Display
plt.figure(figsize=(10, 6))
plt.imshow(img)
plt.axis('off')
plt.show()

License

This dataset maintains the same license as the original UnrealStereo4K dataset. Please refer to the original repository for details.