Robust correspondence and retrieval of articulated shapes

We consider the problem of shape correspondence and retrieval. Although our focus is on articulated shapes, the methods developed are applicable to any shape specified as a contour, in the 2D case, or a surface mesh, in 3D. We propose separate methods for 2D and 3D shape correspondence and retrieval, but the basic idea for both is to characterize shapes using intrinsic measures, defined by geodesic distances between points, to achieve robustness against bending in articulated shapes. In 2D, we design a local, geodesic-based shape descriptor, inspired by the well-known shape context for image correspondence. For 3D shapes, we first transform them into the spectral domain based on geodesic affinities to normalize bending and other common geometric transformations and compute correspondence and retrieval in the new domain. Various techniques to ensure robustness of results and efficiency are proposed. We present numerous experimental results to demonstrate the effectiveness of our approaches.