Non-isotropic regularization of the correspondence space in stereo-vision

The correspondence problem in stereo vision is notoriously difficult. In many approaches a noisy solution is extracted from the correspondence space. Various sophisticated regularization techniques are applied then on this noisy solution. We study here the possibility to denoise the correspondence/correlation space before extracting the solution, by a non-linear and non-isotropic scheme. We show that this method preserves edges (depth discontinuities) well and overcomes some of the problems encountered in previous approaches.