Reliability based graph-cuts algorithm for stereo

Energy functions that are used in stereo problem, are still far from modeling the real world. The energy terms in the models are not successful on all image regions and it is very difficult to adjust the weights of these terms automatically. An exact model for the stereo problem will be too complex and intractable. Therefore, we need some extra information to achieve better results while using basic energy models. Knowing where a method fails is a very important information and in this paper, we show that graph-cuts based dual-mesh framework can provide this information. The proposed method can mark the reliable and unreliable parts of the image regions. It then blocks the effect of unreliable parts over the reliable parts and it repairs unreliable parts iteratively by using the information from reliable parts.We tested our algorithm on Middlebury image dataset and on stereo X-ray images. We compare our algorithm with the other stereo algorithms and achieve better results than the base graph-cut algorithm as well as the graph-cut based dual-mesh algorithm.