Iterative Point Matching via multi-direction geometric serialization and reliable correspondence selection

Abstract

Point matching aims at finding the optimal matching between two sets of feature points. It is widely accomplished by graph matching methods which match nodes of graphs via minimizing energy functions. However, the obtained correspondences between feature points vary in their matching qualities. In this paper, we propose an innovative matching algorithm which iteratively improves the matching found by such methods. The intuition is that we may improve a given matching by identifying “reliable” correspondences, and re-matching the rest feature points without reliable correspondences. A critical issue here is how to identify reliable correspondences, which is addressed with two novel mechanisms, Multi-direction Geometric Serialization (MGS) and Reliable Correspondence Selection (RCS). Specifically, MGS provides representations of the spatial relations among feature points. With these representations, RCS determines whether a correspondence is reliable according to a reliability metric. By recursively applying MGS and RCS, and re-matching feature points without reliable correspondences, a new (intermediate) matching can be obtained. In this manner, our algorithm starts with a matching provided by a classical method, iteratively generates a number of intermediate matchings, and chooses the best one as the final matching. Experiments demonstrate that our algorithm significantly improves the matching precisions of classical graph matching methods.