Stereo Matching with Semi-limited Belief Propagation

In this paper, we present an effective scheme with semi-limited belief propagation to improve the performance of stereo matching. the mean shift segmentation is first introduced to decompose both reference image and target image of stereo pair into regions with similar disparity. We use the regions as matching primitives, and figure out disparity range for each segment by applying a fast local method. then a color-weighted method is applied to calculate matching cost. the unreliable cost is detected by a stability test method. It is helpful in formulating the data term of energy function. in the smoothness term, segment cue and occluded cue was incorporated into the prior assumption to control the way messages passes in the Markov network. then semi-limited belief propagation is used to minimize energy function. Experimental results using the Middlebury stereo test bed show the proposed method works well.