GLOBAL BALANCED NETWORKS FOR MULTI-VIEW STEREO

Existing deep learning-based multi-view stereo (MVS) methods have achieved satisfactory results in terms of accuracy of reconstruction results, but since these methods introduce significant ambiguity in low-texture or even no-texture, or highly reflective regions, this leads to poor depth prediction in low-texture smooth regions, which directly affects the completeness of the reconstruction results. Therefore, how to strike a balance between accuracy and integrity is currently a major challenge for MVS. Based on this problem, we propose a global balanced network(GBNet) for multi-view stereo which aims to maximize completeness with guaranteed accuracy to achieve relative balance. Specifically, our network consists of three parts. The first part is our proposed a feature distillation modular, which mainly serves to reduce feature redundancy and enhance the expressiveness of important distinguishing features. The second part is our hydra attention fusion module, which maximizes the capture of global contextual features to build compact cost-volume. In addition, we propose a non-standard 3D convolution module for aggregating the depth features. Experiments conducted on the DTU and tank temple datasets show that our method outperforms many current state-of-the-art methods in terms of completeness and exhibits a certain degree of generalization performance.