3d Line Segment Reconstruction In Structured Scenes Via Coplanar Line Segment Clustering

This paper presents a new algorithm aiming for 3D Line Segment (LS) reconstruction in structured scenes that are comprised of a set of planes. Due to location imprecision of image LSs, it often produces many erroneous reconstructions when reconstructing 3D LSs by triangulating corresponding LSs from two images. We propose to solve this problem by first recovering space planes and then back-projecting image LSs onto the recovered space planes to get reliable 3D LSs. Given LS matches identified from two images, we estimate a set of planar homographies and use them to cluster the LS matches into groups such that LS matches in each group are related by the same homography induced by a space plane. In each LS match group, the corresponding space plane can be recovered from the 3D LSs obtained by triangulating all the LS correspondences. To reduce the incidence of incorrect LS match grouping, we formulate to solve the LS match grouping problem into solving a multi-label optimization problem. The advantages of the proposed algorithm over others in this area are that it can generate more complete and detailed 3D models of scenes using much fewer images and can recover the space planes where the reconstructed 3D LSs lie, which is beneficial for upper level applications, like scene understanding and building facade extraction.