A Crop-Based Multi-Branch Network for Matching Cost Computation

Stereo matching is a challenging problem in computer vision. An excellent matching cost computation method is useful for enhancing stereo matching performance. Traditional matching cost computation is lack of robustness. In this paper, we propose a crop-based multi-branch convolution neural network (CBMBNet) for robust matching cost computation. We employ ResNeXt block for feature extraction and introduce a new crop-based multi-branch network structure to enhance the accuracy of matching. Several post-processing techniques are used further to enhance disparity map equality. The experimental results show that the proposed CBMBNet can reduce error rates than MC-CNN-fst and MC-CNN-acrt approaches based on Middlebury stereo data set.