Computing correspondences in a sequence of non-rigid shapes

摘要

Image correspondences are computed using a new relaxation technique applied to edge features. Each image is represented by a network of nodes connected by branches. The nodes represent the locations of corners or other conspicuous points on the contours of the image. The branches may be real image contours or imaginary lines drawn to link the nodes. Nodes in one image are matched to nodes in the next image using a measure of the length and angular separation between connecting branches. An iterative relaxation scheme searches for consistency in the structure and relative motion of neighboring branches and nodes, and rewards the matching probability of each node accordingly. Around 15–20 iterations are needed for convergence. The matching process favors structural rigidity on a local scale, while allowing global image-to-image deformations to occur, as in the case of fluid motion. Examples are given as evidence that the algorithm can match patterns undergoing distortion, rotation, and scaling.