SA-MVSNet: Self-attention-based multi-view stereo network for 3D reconstruction of images with weak texture

Abstract

Multi-view stereo (MVS) reconstruction is a key task of image-based 3D reconstruction, and deep learning-based methods can achieve better results than traditional algorithms. However, most of the current deep learning-based MVS methods use convolutional neural networks (CNNs) to extract image features, which cannot achieve the aggregation of long-distance context information and capture robust global information. In addition, in the process of fusing depth maps into point clouds, the confidence filters will filter out the depth values with low confidence in weak texture areas. These problems will lead to the low completeness of 3D reconstruction of weak texture and texture-less areas. To address the above problems, this paper proposes SA-MVSNet based on the PatchmatchNet with a self-attentive mechanism. First, we design a coarse-to-fine network framework to advance depth map estimation. In the feature extraction network, a module with a pyramid structure based on Swin Transformer Block is used to replace the original Feature Pyramid Network (FPN), and the self-correlation between weak texture areas is enhanced by applying a global self-attention mechanism. Then, we also propose a self-attention-based adaptive propagation module (SA-AP), which applies a self-attention calculation within depth value propagation window to obtain the relative weight values of current pixel and others, and then adaptively samples the depth values of neighbors on the same surface for propagation. Experiments show that SA-MVSNet has significantly improved the completeness of 3D reconstruction for the images with weak texture on DTU (provided by Danish Technical University), BlendedMVS, and Tanks and Temple datasets. Our code is available at https://github.com/miaowang525/SA-MVSNet.