Robust Matching of Images Across Large Viewpoint Changes

C. Ancuti, C.O. Ancuti, and P. Bekaert (Belgium)


feature points, image matching, descriptor, homography


This paper introduces a new matching method that deals with images that are taken under large viewpoints changes. Our main motivation is based on the recent eval uations of the state of the art detectors/descriptors that dis closed that the most difficult matching task is to find valid corresponding feature points under large affine transforma tion. None of the existing detectors/descriptors can per form reliable for a viewangle that separates images higher than 40◦ . Our approach overcomes this drawback and consists in two main steps. Firstly, a few good matches, an chor points, are identified in the considered pair of im ages. Secondly, after these initial corresponding points have been established the affine transformation parameters are estimated using the neighbor information of the already matched keypoints. The estimated parameters are used in computation of the descriptors in the elliptical regions re lated by the inferred affine geometry. The method has the advantage that does not need ad ditional refinements (e.g Ransac, epipolar constraints). The experimental results demonstrate a substantial improve ment of our approach against the state of the art methods.

Important Links:

Go Back