Solving stereo correspondence through minimizing energy function with higher-order cliques

Stereo correspondence is one of the most active research areas in computer vision. Energy minimization is widely used for early vision problems, such as image restoration, segmentation and stereo correspondence. Pairwise clique is the most commonly used smoothness term of energy function, but it is unable to capture rich statistics of natural scene. Energy function considering higher-order clique potentials can characterizes richer statistics of natural scene than pairwise clique, but it is difficult to model higher-order clique potentials and the computation for minimization is much heavier. We introduce an reduced P
ⁿ
Potts model which can characterize higher-order clique potentials and was first used for image segmentation. Specifically, we present two new models which map the P
ⁿ
Potts model to alpha-expansion move and alpha-beta swap move. Furthermore, we propose a new graph construction method for them which has fewer extra nodes than before. Those models can be easily applied to other vision problems. The experiment shows that the results considering Pn Potts model are more accurate than those without.