EHSS: An Efficient Hybrid-supervised Symmetric Stereo Matching Network
Dapeng Zhang, Peng Zhi, Binbin Yong, Jinqiang Wang, Yufeng Hou, Lan Guo, Qingguo Zhou, Rui Zhou · 2023
Disparity estimation is a crucial task for 3D reconstruction and autonomous driving, and has seen significant improvements since the application of deep learning. Prior research typically involves building cost volumes for left images and using 2D or 3D convolutional kernels to regress disparity maps for the left images. However, most of those previous architectures are trained with left disparity labels, which limits their ability to exploit occluded regions and textureless regions, thus reducing their overall robustness for full scenes. To address this issue, we introduce EHSS, an efficient hybrid-supervised symmetric stereo matching network that includes a symmetric branch to predict the disparity map for the left and right image, which helps to alleviate the influence of occlusion and eliminate noise from textureless regions. Additionally, we train our network with mixed supervision, including both supervised and self-supervised learning, to enhance the network's robustness in unfamiliar scenes. Our experimental evidence demonstrates that compared to other methods, our approach significantly improves precision, particularly in unfamiliar scenes.