Correspondence Using Mahalanobis Distance, Dynamic Programming, and Relaxation

We present a new solution to the correspondence problem. It is based on dynamic programming, and can deal with image sequences. The overall potential characterizes the quality of matching of a point, and dynamic programming achieves consistency independently for each edge of the new image. The overall potential is deened on a pair of points as a function of two components : a potential deened for each new point as a function of its Mahalanobis distance to the old edges, and a potential associated to their respective edges on the basis of their relative distance and orientation. Finally, occlusions, artifacts and ambiguities are dealt with by applying relaxation to the diierent solutions found by dynamic programming, where the compatibility coeecients associated to the empty label are updated at each iteration.