ABSTRACT
Stereo matching, and geometric computer vision in general, were among the last areas of computer vision to benefit from machine learning. Recently, the pendulum has swung in the opposite direction and most new methods are almost entirely data driven. In this talk, I will present supervised learning approaches that address binocular and multi-view stereo matching by leveraging data with ground truth, as well as conventional wisdom in the form of constraints that have been proven effective over long periods of time with an emphasis on generalization.