A Framework to Improve Matching Results of Widely Separated Views

Abstract

Matching images is a crucial step in many computer vision applications. In this paper we present an alternative strategy built on the SIFT operator to solve the problem of wide-baseline matching. We first show how to add the color information to the SIFT descriptors of extracted keypoints. Practically, the SIFT descriptor vector is blended with the main parameters (contrast, correlation and energy) of the color co-occurrence histogram computed in the same image patch. Afterward, in order to better improve the matching results of images taken under large variations of the camera viewpoint angle, the valid matches obtained by the previous strategy are employed to estimate the geometry between patches of corresponding keypoints. This overcomes the lack of affine invariance of the existing operators (including SIFT), allowing to use a more appropriate region shape where descriptors will be calculated for better preciseness. In our experiments the proposed method shows a substantial improvement of the matching results compared with the results obtained by the original local operator.