Stereo Matching by Self-supervision of Multiscopic Vision

This is the official implementation code for SMVmatching -- "Stereo Matching by Self-supervision of Multiscopic Vision". For technical details, please refer to :

Stereo Matching by Self-supervision of Multiscopic Vision
Weihao Yuan, Yazhan Zhang, Bingkun Wu, Siyu Zhu, Ping Tan, Michael Yu Wang, Qifeng Chen
IROS2021
[Paper] [Project Page]

Bibtex

If you find this code useful, please consider citing:

@inproceedings{yuan2021stereo,
  title={Stereo Matching by Self-supervision of Multiscopic Vision},
  author={Yuan, Weihao and Zhang, Yazhan and Wu, Bingkun and Zhu, Siyu and Tan, Ping and Wang, Michael Yu and Chen, Qifeng},
  booktitle={Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={},
  year={2021},
  organization={IEEE}
}

Contents

1. Environment Setup
2. Dataset
3. Training

Environment setup

Dependencies:

• Python2.7
• PyTorch (0.4.0+)
• torchvision (0.2.0+)

Datasets:

The synthetic datasets are available at here. It should be organized as follows for the training:

dataset/
    TRAIN/
        SyntheData/
            000001/
                view0.png
                view1.png
                view2.png
    TEST/
    EVAL/

Training:

. run.sh

Pretrained models

Results on KITTI 2015:

Model	Dataset	D1-all	D1-noc	D1-occ
Model1	Train set	4.23%	4.09%	-
Model2	Train set	3.82%	3.69%	10.87%
Model2	Online Test set	4.43%	4.20%	-

Model1: model without probability uncertainty

Model2: model with probability uncertainty

---

Acknowledgements

Thanks to Jia-Ren Chang et al. for opening source of their excellent work PSMNet.

License

Licensed under an MIT license.