Robust Stereo Matching with an Unfixed and Adaptive Disparity Search Range

26th International Conference on Pattern Recognition (ICPR2022) 2022.03.09,

Jiazhi Liu, Feng Liu.


Stereo matching is an essential basis for various ap- plications in computer vision, but currently, most stereo matching methods have poor generalization performance and require a fixed disparity search range. In this work, we propose to adopt the operation “sub-pixel three-around-maximum”, which supports an unfixed disparity search range, to replace the currently popu- lar operation soft argmax. We also propose to directly supervise the feature extractor by three loss functions. Just depending on the feature extractor, we can obtain an accurate semi-dense disparity map before cost aggregation, which can help to remove adaptively the redundancy part of the predefined disparity search range. An adaptive disparity search range for each stereo pair can save much time and memory. The proposed architecture achieves state-of-the-art cross-domain generalization performance in the datasets KITTI2012&2015, and experimental results demonstrate that our method supports unfixed and adaptive disparity search range.