Multi-View 3D Reconstruction Algorithm Based on SAM Plane Prior

Traditional multi-view stereo matching (MVS) algorithms often encounter mismatch problems when dealing with weak texture areas, and relying solely on photometric consis-tency and geometric consistency constraints cannot guarantee accurate depth estimation. To address the mismatch problem in weak texture areas, this paper introduces a SAM (Segment Anything Model) instance segmentation model, which uses the SAM model to perform instance segmentation on weak texture areas to generate plane prior information, and then further incorporates the segmented prior information into the matching cost to assist depth estimation. In addition, an adaptive depth propagation method that combines SAM instance segmentation with local consistency is proposed. This method uses the SAM instance segmentation results to guide the propagation of depth information, effectively improving the accuracy and robustness of depth estimation. Through comparative experiments with the current mainstream multi-view 3D reconstruction algorithms, the results show that the method that combines SAM segmentation prior information has significant advantages in depth estimation and point cloud reconstruction. Finally, the algorithm is tested on a large-scale aerial photography dataset of oil pipelines, and the reconstructed 3D map of the real terrain has high integrity and accuracy.