Probabilistic Signal Models to Regularise Dynamic Programming Stereo

Ill-posedness of the binocular stereo problem stems from partial occlusions and homogeneous textures of a 3D surface. We consider the symmetric dynamic programming stereo regularised with respect to partial occlusions. The regularisation is based on Markovian models of epipolar profiles and stereo signals that allow for measuring similarity of stereo images with due account of binocular and monocular visibility of the surface points. Experiments show that the probabilistic regularisation yields mostly accurate elevation maps but fails in excessively occluded or shaded areas.