A Self-Attention based Network for Low Resolution Multi-View Stereo

We present SA-MVSNet, a novel two-stage multi-view stereo network equipped with self-attention mechanism, which can improve the quality of low-resolution image 3D reconstruction. SA-MVSNet consists of two stages, and the lower resolution depth maps predicted in the first stage provide a priori information for the second stage. To increase the utilization of image information, a pyramid scheme was used to fuse the feature maps at different resolutions. Moreover, we introduce an improved self-attention module in the first stage to improve reconstruction accuracy by learning the long-term dependence information of feature maps. The experiments on the DTU dataset show a promising result in both completeness and accuracy metrics of the 3D scene reconstructed by the proposed method.