Multi-View Stereo of UAV Remote Sensing Images Based on Adaptive Propagation with Multi-Region Refinement

Based on computer vision and image processing techniques, 3D reconstruction of targets such as the ground surface or buildings from unmanned aerial vehicle (UAV) images can be realized, in which the dense reconstruction for weak texture regions is very important. In this paper, we propose a novel multi-view stereo method for UAV remote sensing images based on adaptive propagation and multi-region refinement, named APMRR-MVS. Firstly, in the propagation step, we propose an adaptive propagation strategy based on a checkerboard grid, which expands the distal sampling region by continuously selecting pixels with better hypotheses, that can explore the distal end more flexibly and improve the quality of sampling hypotheses for the pixels within the same view. Secondly, in the refinement step, we propose a multi-region refinement strategy, which can improve the efficiency of exploring the solution space by arranging several regions independently and reduce the possibility of the target pixel hypothesis being trapped in a local optimum. It is demonstrated on DTU and Blended MVS datasets that our method has better performance in the reconstruction of weakly textured regions, in addition to preserving specific texture details to a greater extent and reducing the erroneous estimation of spatial points.