Multi-view stereo network with point attention

Abstract

In recent years, learning-based multi-view stereo (MVS) reconstruction has gained superiority when compared with traditional methods. In this paper, we introduce a novel point-attention network, with an attention mechanism, based on the point cloud structure. During the reconstruction process, our method with an attention mechanism can guide the network to pay more attention to complex areas such as thin structures and low-texture surfaces. We first infer a coarse depth map using a modified classical MVS deep framework and convert it into the corresponding point cloud. Then, we add the high-frequency features and different-resolution features of the raw images to the point cloud. Finally, our network guides the weight distribution of points in different dimensions through the attention mechanism and computes the depth displacement of each point iteratively as the depth residual, which is added to the coarse depth prediction to obtain the final high-resolution depth map. Experimental results show that our proposed point-attention architecture can achieve a significant improvement in some scenes without reasonable geometrical assumptions on the DTU dataset and the Tanks and Temples dataset, suggesting that our method has a strong generalization ability.