Stereo disparity calculation in real-world scenes with Informative Image Partitioning

Despite their presence in almost all real-world navigational environments, finding stereo disparity in large, weakly textured regions is a difficult task. In this paper, we present a method which constructs a novel segmentation of an image into separate planar regions prior to disparity calculation. This segmentation allows us to fit planar models to disparity estimates calculated using a graph cut approach within each segment, producing smooth, accurate disparity maps in ordinarily difficult areas. Our method achieves good results using intensity features in a variety of indoor and outdoor navigational scenes. A comparison of our results with other benchmark approaches is presented, in which the advantages of our technique can be clearly seen.