A stochastic cost function for stereo vision

Abstract

In this paper, we use random walks to infer information about the disparities in the image. While the random walks help to maintain sharp object boundaries, we introduce a correlation step that handles occlusions and slanted surfaces. The random walks are then used to build a stochastic cost function which serves to identify the most probable disparities. This also delivers valuable information about the reliability of the depths by exploiting their consistency statistically. Our method delivers very good results by means of local matching, but we also demonstrate that the obtained cost function is well suited for global optimization techniques. Moreover, our consistency maps deliver reliable statistical information about the confidence of disparity maps. In our paper we provide extensive evaluations with challenging images and show that our cost function based on random walks is very useful.