To address the reduced accuracy of stereo matching networks near object boundaries and disparity discontinuities, a double cost–volume stereo matching network with entropy-difference-guided fusion is proposed. The proposed network was built based on RAFT-Stereo. It employs a pretrained backbone to extract multi-scale features and uses deformable attention for cross-scale feature fusion. A shallow image-guided branch was used to generate pixel-wise constraint information to limit the magnitude of sampling offsets and alleviate cross-structure sampling. Based on the extracted features, a group-wise correlation cost–volume and a normalized correlation cost–volume were constructed. Both cost–volumes were regularized by 3D Hourglass networks, and a structure-consistent intra-scale aggregation module was introduced during the regularization of the group-wise correlation cost–volume. The two aggregated results were then fused by the entropy-difference-guided fusion module to obtain the final cost–volume. The experimental results show the effectiveness of the proposed network in the Scene Flow, KITTI, and ETH3D datasets, achieving an endpoint error of 0.45 px and a >3 px error rate of 2.41% on the Scene Flow dataset.
Yang et al. (Mon,) studied this question.