Segment image with depth in a bayesian framework

This paper presented the problem of unsupervised extraction of relative depths and occlusion boundaries of regions from a single image in a Bayesian framework, and presents a solution. The algorithm firstly segmented an image into several over-segmented regions, then from bottom to up inferring the atomic regions into large regions cued by regions features and the 3-D information with Swendsen-Wang cut sampling mechanism. The graph nodes are represented with a dynamic system, which called mixed Markov Field. The experiments showed that the proposed method working well.