Focusing on Cross Views Improves Reconstruction in Unsupervised Multi-View Stereo

The purpose of Multi-view Stereo(MVS) is to reproduce three-dimensional scenes through two-dimensional images. Unsupervised Multi-view Stereo has achieved remarkable progress both in terms of reconstruction accuracy and completeness. However, previous methods are mainly rely on the prerequisite that the corresponding pixels between different views share similar characteristics. These methods have certain limitations in the actual scene, especially for some special non-Lambert surface and specular area. In the process of 3D reconstruction of the above-mentioned scenes, the depth estimation work we want to adopt will be greatly affected. In order to improve the situation, we carry out relevant adjustment and optimization work on the self-supervised signal. We consider focusing on cross-views to improve the stability and anti-disturbance of self-supervised signals. Therefore, we design a novel self-supervised network to improve the quality of generated point clouds during reconstruction. In addition, we also introduce a new loss function to capture and characterize the relationship between cross-views to achieve the purpose of improving the completeness and accuracy of depth estimation in boundary regions. We carry out experiments and tests on DTU dataset and Tanks&Temples dataset respectively. The experimental results prove that our proposed method has good reconstruction ability and good generalization effect.