An Evaluation of Affine Invariant-Based Classification for Image Matching

Abstract

This paper presents a detailed evaluation of a new approach that uses affine invariants for wide baseline image matching. Previously published work presented a new approach to classify tentative feature matches as inliers or outliers during wide baseline image matching. After typical feature matching algorithms are run and tentative matches are created, the approach is used to classify matches as inliers or outliers to a transformation model. The approach uses the affine invariant property that ratios of areas of shapes are constant under an affine transformation. Thus, by randomly sampling corresponding shapes in the image pair a histogram of ratios of areas can be generated. The matches that contribute to the maximum histogram value are then candidate inliers. This paper evaluates the robustness of the approach under varying degrees of incorrect matches, localization error and perspective rotation often encountered during wide baseline matching. The evaluation shows the affine invariant approach provides similar accuracy as RANSAC under a wide range of conditions while maintaining an order of magnitude increase in efficiency.