Perceptual organization in motion correspondence

Abstract

Geometrical cues for motion correspondence are addressed. Ullman's minimal distance mapping solution to motion correspondence suffers from a preponderance of global matches with near optimal values. This presence of a number of local minima is problematic for neural-network-based optimization schemes. Perceptual organization is used to provide additional constraints on the correspondence process. These constraints, namely the proximity and co-curvilinearity constraints, act to increase the difference between the local minima obtained from the minimal distance mapping and the global optimum. This leads to substantially improved performance of the optimization scheme used for constraint satisfaction.