A curvature salience descriptor for full and partial shape matching

Shape representation and matching is one of the fundamental problems in compute vision. In this paper, we propose a novel shape contour descriptor, called curvature salience descriptor (CSD), for full shape matching. The presented descriptor utilizes only the most representative information to globally describe shape contour and is invariant to translation, rotation, scaling and partial occlusion of shapes. Shape matching is performed using the Dynamic Time Warping (DTW) to establish the point-to-point correspondence. To the solution of partial shape matching, we slightly modify the descriptor and develop a new matching framework that can establish the correspondence between the open query and candidate from coarse to fine. Different from previous methods, a coarse matching process is implemented to fast reject the false candidates before we establish the point-to-point correspondence. We conduct the experiments on the public datasets but also in the context of some specific applications and the results demonstrate that the proposed technique can achieve favorable performance compared to the existing methods.