Feature distribution normalization network for multi-view stereo

As a key technique in 3D reconstruction, research in multi-view stereo (MVS) has made significant progress with the development of deep learning. However, MVS faces challenges due to inconsistent feature distributions across different viewpoints. This phenomenon has been recognized as a significant bottleneck in MVS. To overcome this limitation, we propose an innovative approach called feature distribution normalization network (FDN–MVS). To mitigate errors arising from inconsistent relative poses, we leverage homography transformations between multiple views and propose a distribution residual refinement (DRR) module. Expanding the scale, residual connection, and coordinate system fixation results in more meaningful feature distribution learning. In order to further narrow the disparities among different views, we suggest feature normalization, constraining the computation of features within a common distribution. Experimental results demonstrate the effective of our method. Compared to the commonly used benchmark algorithm CasMVSNet, we reduced the error by 13.52% on the DTU dataset and improved the F-score by 18.77% on the Tanks and Temples dataset. Code is available at https://github.com/ZYangChen/FDN-MVS.