Accurate and Robust Image Correspondence for Structure-From-Motion and its Application to Multi-View Stereo

Abstract

In this paper, we propose a robust and accurate image correspondence method by combining SuperPoint + SuperGlue (SP+SG) and Local feature matching with TRansformers (LoFTR). The proposed method finds corresponding points on regions with rich texture by SP+SG and those with poor texture by LoFTR since SP+SG exhibits high localization accuracy of image correspondence and LoFTR exhibits high robustness against poor texture regions. The proposed method can be used for image correspondence in SfM to not only improve the estimation accuracy of camera parameters in SfM, but also to improve the reconstruction accuracy and expand the reconstruction area in MVS. Through experiments on the ETH3D dataset, we demonstrate that the proposed method achieves more accurate 3D reconstruction than conventional methods, and also show the impact of image correspondence accuracy in SfM on multi-view 3D reconstruction.