This repository implements and compares multiple stereo matching techniques to generate disparity maps from stereo image pairs. The methods include:
- Traditional Window-Based Matching (Block Matching)
- OpenCV Stereo Semi-Global Block Matching (StereoSGBM)
- Pyramid Stereo Matching Network (PSMNet)
The project also evaluates the Peak Signal-to-Noise Ratio (PSNR) between generated disparity maps and ground-truth disparity maps to assess the accuracy of the methods.
- Traditional Block Matching: A window-based matching approach implemented from scratch.
- StereoSGBM: Advanced stereo matching using OpenCV's Semi-Global Block Matching algorithm.
- PSMNet: A deep learning-based stereo matching method using a pre-trained Pyramid Stereo Matching Network.
- Disparity Map Evaluation: Quantitative evaluation using PSNR.
The project requires the following libraries:
- Python 3.x
- OpenCV
- NumPy
- tqdm
- PIL (Python Imaging Library)
- PyTorch (for PSMNet)
- CUDA (for GPU-based PSMNet operations)
- Matches a small window of pixels between left and right images.
- Implementation:
- Sum of Absolute Differences (SAD) as the cost metric.
- Adjustable window size and maximum disparity.
- Uses OpenCV's
StereoSGBM_createfor semi-global matching. - Parameters are tuned for better accuracy.
- A deep learning-based stereo matching method.
- Pre-trained on the Scene Flow dataset.
- Requires resizing images for limited GPU memory.
Below is a comparison of reconstructed disparity maps for the "Art" stereo image pair:
| Ground Truth | Window-Based Matching | StereoSGBM | PSMNet |
|---|---|---|---|
 |
 |
 |
 |
Below is a comparison of reconstructed disparity maps for the "Dolls" stereo image pair:
| Ground Truth | Window-Based Matching | StereoSGBM | PSMNet |
|---|---|---|---|
 |
 |
 |
 |
Below is a comparison of reconstructed disparity maps for the "Reindeer" stereo image pair:
| Ground Truth | Window-Based Matching | StereoSGBM | PSMNet |
|---|---|---|---|
 |
 |
 |
 |
| Method | Art | Dolls | Reindeer |
|---|---|---|---|
| Window-Based Matching | 12.41 | 12.15 | 10.84 |
| StereoSGBM | 12.67 | 16.15 | 15.58 |
| PSMNet (Resized 70%) | 15.79 | 15.26 | 15.95 |
- SGBM: Hirschmüller, Heiko. "Accurate and efficient stereo processing by semi-global matching and mutual information." CVPR 2005.
- PSMNet: Chang, Jia-Ren, and Yong-Sheng Chen. "Pyramid stereo matching network." CVPR 2018.
- Scene Flow Dataset: Mayer, N., et al. "A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation." CVPR 2016.