Shape matching using curvature processes

Shape matching is a fundamental problem of vision in general and interpretation of deforming shapes in particular. The objective of matching in this instance is to recover the deformation and therefore generalizes the notion of correlation, which aims to only produce a numerical measure of the similarity between two shapes. To address shape matching, we introduce a new representation of a closed 2D shape a cyclic sequence of the extended circular images of the convex and concave segments of its contour. This representation is then used to establish correspondences between segments of the two contours using dynamic programming. Finally, we compute a recovery of the differences between two similar contours in terms of the action of curvature process. Computation of convex and concave segments of the contours, given in piecewise linear form, is accomplished using the analytic representation of a local B-spline fit. We show the result of our deformation recovery scheme appliedto dynamic cloud silhouette analysis using hand-traced input from real satellite images.