关键词: 深度学习, 多视图立体匹配, 路径聚合, ＳＥ＋模块

Abstract: Research on learning-based multi-view 3D reconstruction has achieved significant progress,issues like incomplete feature extraction and poor correlation between cost volume still exist,resulting in low reconstruction accuracy.To address these problems,a cascaded network called PAA-MVSNet is designed.Firstly,inspired by the idea of PANet,a “top-down” feature extraction network is designed to extract features at different scales.These features are then concatenated to reduce feature loss.This module outputs three different-scale feature layers,which are fed into the subsequent network to adaptively adjust the depth values and generate higher-precision depth maps.Additionally,to tackle the poor correlation between cost volume,an improved SE attention module is incorporated to optimize the regularization process of the 3D cost volume.This module enhances reconstruction accuracy and completeness.Experimental results on the DTU dataset and Tanks and Temples dataset demonstrate that PAA-MVSNet achieves generalization