Two-Frame Wide Baseline Matching

• Authors: Jiangjian Xiao;Mubarak Shah
• Venue: ICCV '03 Proceedings of the Ninth IEEE International Conference on Computer Vision - Volume 2
• Year: 2003

This paper describes a novel approach to automatically recovercorresponding feature points and epipolar geometry over two widebaseline frames. Our contributions consist of several aspects:First, the use of an affine invariant feature, edge-corner, isintroduced to provide a robust and consistent matching primitives.Second, based on SVD decomposition of affine matrix, the affinematching space between two corners can be approximately dividedinto two independent spaces by rotation angle and scaling factor.Employing this property, a two-stage affine matching algorithm isdesigned to obtain robust matches over two frames. Third, using theepipolar geometry estimated by these matches, more correspondingfeature points are determined. Based on these robustcorrespondences, the fundamental matrix is refined, and a series ofvirtual views of the scene are synthesized. Finally, severalexperiments are presented to illustrate that a number of robustcorrespondences can be stably determined for two wide baselineimages under significant camera motions with illumination changes,occlusions, and self-similarities. After testing a number ofexamples and comparing with the existing methods, the experimentalresults strongly demonstrate that our matching method outperformsthe state-of-art algorithms for all of the test cases.