Urban image stitching using planar perspective guidance

Abstract

Image stitching methods with spatially-varying homographies have been proposed to overcome partial misalignments caused by global perspective projection; however, local warp operators often fracture the coherence of linear structures, resulting in an inconsistent perspective. In this paper, we propose an image stitching method that warps a source image to a target image by local projective warps using planar perspective guidance. We first detect line structures that converge into three vanishing points, yielding line-cluster probability functions for each vanishing point. Then we estimate local homographies that account for planar perspective guidance from the joint probability of planar guidance, in addition to spatial coherence. This allows us to enhance linear perspective structures while warping multiple urban images with grid-like structures. Our results validate the effectiveness of our method over state-of-the-art projective warp methods in terms of planar perspective.