On the Improvement of Image Feature Matching under Perspective Transformations

This paper presents a novel methodology to perform consistent matching between visual features of a pair of images, particularly in the case of point-of-view changes between shots. Traditionally, such correspondences are determined by computing the similarity between descriptor vectors associated with each point which are obtained by invariant descriptors. Our methodology first obtains a coarse global registration among images, which constrains the correspondence space. Then, it analyzes the similarity among descriptors, thus reducing both the number and the severity of mismatches. The approach is sufficiently generic to be used with many feature descriptor methods. We present several experimental results that show significant increase both in accuracy and the number of successful matches.