A correlation-based model prior for stereo

All non-trivial stereo problems need model priors to deal with ambiguities and noise perturbations. To meet requirements of increasingly demanding tasks such as modeling for rendering, a proper model prior should impose preference on the true scene structure, while avoiding artificial bias such as fronto-parallel. We introduce a geometric model prior based on a novel technique we call kernel correlation. Maximizing kernel correlation is shown to be equal to distance minimization in the M-estimator sense. As a model prior, kernel correlation is demonstrated to have good properties that can result in renderable, very smooth and accurate depth map. The results are evaluated both qualitatively by view synthesis and quantitatively by error analysis.