Coarse-to-Fine Matching of Shapes Using Disconnected Skeletons by Learning Class-Specific Boundary Deformations

Disconnected skeleton [1] is a very coarse yet a very stable skeleton-based representation scheme for generic shape recognition in which recognition is performed mainly based on the structure of disconnection points of extracted branches, without explicitly using information about boundary details [2, 3]. However, sometimes sensitivity to boundary details may be required in order to achieve the goal of recognition. In this study, we first present a simple way to enrich disconnected skeletons with radius functions. Next, we attempt to resolve the conflicting goals of stability and sensitivity by proposing a coarse-to-fine shape matching algorithm. As the first step, two shapes are matched based on the structure of their disconnected skeletons, and following to that, the computed matching cost is re-evaluated by taking into account the similarity of boundary details in the light of class-specific boundary deformations which are learned from a given set of examples.