Hourglass stereo matching network

Abstract

Although the deep learning-based stereo matching networks have made significant progress, the ability to find corresponding relationships in ill-conditioned regions (weak textures, repeated textures, occluded regions, etc.) still needs to be improved. Aiming at the above problems, an hourglass stereo matching network is proposed, which is mainly composed of an hourglass structure. The first is the hourglass feature extraction module, which uses the information of the global context to obtain more detailed features; then the obtained feature information is aggregated together to construct a cost volume. In the three-dimensional convolution module, multiple hourglass structures are used to refine disparity and use intermediate supervision to standardize the cost volume. It fuses the information again and makes better use of global context information. Finally obtain the disparity map through disparity regression. Through the verification test on the Scene Flow and KITTI dataset, it shows that the proposed method maintains better performance while reducing parameters, significantly reduces the error in ill-conditioned regions, and achieves competitive results.