Multi-homography Estimation and Inference Driven by Contour Alignment

Tao Cai, Yunde Jia, Huijun Di, Yuwei Wu

Abstract: Recently, multiple homography estimation is preferable for image stitching to handle the parallax problem, by estimating homographies from the feature correspondence in each local region. However, correspondence outliers and insufficient feature coverage will lead to unreliable local homography fitting. In this paper, we propose a novel method of multi-homography estimation and inference, driven by contour alignment. Our method uses explicit structural verification through contour alignment to eliminate incorrectly fitted homographies in some regions, and to select a better homography from other regions if current homography is rejected or with worse accuracy. With the guidance of the contour alignment, dense image alignment result is obtained by further inferring the local homography per superpixel. Quantitative and qualitative comparisons demonstrate the effectiveness of our method, especially for scenes with large parallax and viewpoint changes.