Establishing Visual Correspondence from Multi-Resolution Graph Cuts for Stereo-Motion

This paper presents the design and implementation of a multi-resolution graph cuts (MRGC) for stereo-motion framework that produces dense disparity maps. Both stereo and motion are estimated simultaneously under the original graph cuts framework. Our framework extends the problem from one to five dimensions, creating a large in- crease in complexity. Using three different multi-resolution graph cut algorithms, LDNR, EL and SAC, we reduce the number of pixels m and the number of labels n that limit the alpha - beta swap algorithm (with complexity O(mn 2) required from the definition of our semi-metric smoothness function. This results in a reduction of computation time and the ability to handle larger images and larger label sets. The choice of the three MRGC algorithms to use in computation deter- mines the appropriate level of accuracy and computation time desired.