Channel and Spatial Reconstruction for Multi-View Patchmatch Stereo

During the process of multi-view 3D reconstruction, there may be redundant features that can affect the reconstruction results. To address this issue, we have applied an improved S-shaped feature pyramid network to capture connections at different scales. We have also incorporated a spatial redundancy removal module (SRU) and a channel redundancy removal module (CRU) to eliminate redundant information during feature extraction. Furthermore, we have developed a new patchmatch-based network structure using an improved cascade structure. By training on the publicly available DTU dataset, we have achieved good results in terms of completeness and accuracy, the results are better than the benchmark network and our approach has good performers in 3D reconstruction using deep learning methods. The results of the point cloud which we reconstructed from the depth map have also shown significant improvement.