End-to-End Stereo Matching Network with Local Adaptive Awareness
Chenggang Guo, Dongyi Chen, Zhiqi Huang · 2020
End-to-end training of deep neural networks have been extensively utilized for the problem of stereo matching. Recent efforts mainly focus on the matching cost computation and the cost aggregation with additional guidance. In this paper, we propose a novel end-to-end stereo matching network that explores local adaptive awareness in the disparity upsampling stage and the loss function. An improved hierarchical upsampling module is presented to adaptively learns a disparity refinement function. A shared convolution structure is adopted in the upsampling module to adaptively discover local semantic correlations between image intensity pixels and disparity pixels. The same shared convolution structure is utilized in the proposed loss function to explicitly exploit local patterns among disparity pixels. Experiments on several recent large scale stereo matching benchmarks validate the effectiveness of our proposed network architecture. Comparisons with state-of-the-art networks also show that our network has a comparable accuracy and promising running speed.