Seamless Stitching using Multi-Perspective Plane Sweep

Sing Bing Kang, Rick Szeliski, Matt Uyttendaele

MSR-TR-2004-48 |

A major problem in stitching a partially overlapping set of images taken from different viewpoints is the presence of parallax, which causes ghosting. We propose a new technique, Multi-Perspective Plane Sweep (MPPS), to handle this problem. Given a pair of images to stitch, we rectify them, find their area of intersection, and estimate a disparity map by plane sweeping different columns referenced at different virtual camera positions. This minimizes object distortion across the stitched area while producing a seamless composite. We show experimental results which validate our technique. We also introduce a variant, Stretching Stereo, that works almost as well, but is significantly faster.