Strong Stereo Features for Self-Supervised Practical Stereo Matching

Abstract

This paper proposes a new approach to the problem of self-supervised dense stereo correspondence. Disparity estimation is an important problem of computer vision but in many situations, correspondence-based reconstruction cannot be accompanied by a ground truth to train supervised methods. This paper proposes to train a siamese feature encoder in a self-supervised permutation framework and then build a cost volume which is fed to a classical stereo algorithm to compute the disparity. In the absence of ground truth disparity, rather than handcrafting features, we suggest a novel and straightforward way to leverage input images to train for features. A key aspect of this method is that all the trainable weights are located inside a feature computation step, which is followed by strong non-trainable constraints that enforce bidirectional correspondence through cross- attention. Validated on real and synthetic datasets and compared to various methods, our proposed approach yields competitive results. Given its high performance, simplicity, and direct integration with current stereo algorithms, we expect this method to further the adoption of deep methods in real life stereo applications

BibTeX

@inproceedings{brousseau2023strong,
  title={Strong Stereo Features for Self-Supervised Practical Stereo Matching},
  author={Brousseau, Pierre-André and Roy, Sébastien},
  booktitle={British Machine Vision Conference (BMVC), 2023},
  year={2023}
}

Documents

PDF Poster

Code

https://gitlab.com/labv3d/strong-stereo-features

Video

---