Image matching based on tracking matching paths in the similarity space

Abstract

In this paper, we explore the feasibility of implementing surface continuity and ordering constraints by tracking connected local peak points in the space of matching similarity. We present a new image matching algorithm based on match paths directly extracted in the similarity space. The main advantage of these kinds of tokens over tokens extracted from the left and right images is that the geometric distortions caused by perspective effects and description inconsistency caused by independent extraction of tokens in the left and right images are automatically eliminated. Through the tracking process, the computation is greatly reduced. In addition, global information available to support a local match for resolving matching ambiguities is fully utilized in such a way that unrelated global information is excluded. Thus, the new image matching algorithm is reliable and efficient. By eliminating the interpolation process at the levels except the finest, the occluded regions and depth discontinuities can be well localized.