Comparison of two proximal splitting algorithms for solving multilabel disparity estimation problems

Disparity estimation constitutes an active research area in stereo vision, and in recent years, global estimation methods aiming at minimizing an energy function over the whole image have gained a lot of attention. To overcome the difficulties raised by the nonconvexity of the minimized criterion, convex relaxations have been proposed by several authors. In this paper, the global energy function is made convex by quantizing the disparity map and converting it into a set of binary fields. It is shown that the problem can then be efficiently solved by parallel proximal splitting approaches. A primal algorithm and a primal-dual one are proposed and compared based on numerical tests.