Stereo Correspondence using Multi-label QPBO Method

Stereo correspondence has been one of the most intense areas of research in computer vision. Graph-based energy minimization has proven to be a powerful framework for incorporating global constraints with local matching costs. In this work, dense stereo correspondence is cast as multi-label energy minimization problem, which is then solved using MQPBO method [1]. We believe the ability of MQPBO to compute the approximate solution with all labels at once has the potential to avoid local optima that may occur in iterative methods. Furthermore, by using QPBO-based technique, the class of energy functions that can be minimized is not limited to being submodular thus our approach is independent of matching costs and smoothness priors. We have tested our postulate with different combinations of cost functions and our experiment has shown encouraging results in various settings.