Cooperative optimization for energy minimization in computer vision: A case study of stereo matching

Abstract

This paper presents a cooperative optimization algorithm for energy minimization in a general form. Its operations are based on parallel, local iterative interactions. This algorithm has many important computational properties absent in existing optimization methods. Given an optimization problem instance, the computation always has a unique equilibrium and converges to it with an exponential rate regardless of initial conditions. There are sufficient conditions for identifying global optima and necessary conditions for trimming search spaces. To demonstrate its power, a case study of stereo matching from computer vision is provided. The proposed algorithm does not have the restrictions on energy functions imposed by graph cuts [1,2], a powerful specialized optimization technique, yet its performance was comparable with graph cuts in solving stereo matching using the common evaluation framework [3]. © Springer-Verlag 2004.