Multi-View Stereo Matching Method Based on Recursive Neural Networks
Jiajia Liu, Guoliang Jiang · Journal of Image Processing Theory and Applications · 2023
This article proposes a recursive layered network reconstruction method with pixel-wise attention cost aggregation to address the problems of textureless areas and poor reconstruction results at scene edges in multi-view stereo matching methods. First, multi-scale features of multiple images are extracted through downsampling and transformed into a cost volume using three-dimensional differentiable homography. Then, a pixel-wise attention aggregation module is added to the cost volume aggregation stage to reweight different pixels and generate a new cost volume. Next, a network with recursive layers is used to regularize the cost volume, replacing the traditional 3D CNN network, and an initial depth map is generated. Finally, the filtered and refined depth maps are merged to generate a three-dimensional dense point cloud. Experimental results show that the proposed network model improves completeness, accuracy, and overall quality by 0.377, 0.363, and 0.370, respectively, compared to other network models, and produces more complete point cloud reconstructions in weak texture areas and scene edges.