Selective image diffusion: application to disparity estimation

Inverse problems encountered in image processing and computer vision are often ill-posed. Whether set in a Bayesian or energy-based context, such problems require prior assumptions expressed through an a priori probability or a regularization term, respectively. In some cases, the prior term exhibits partial dependence on the observations (e.g., images) that is often ignored to simplify modeling and computations. We review methods that take this dependence into account and we propose a new formulation of the prior term that blends some other simple approaches. Similarly to others, we apply a linear transformation to the prior term but, in addition, we require that the eigenvalues of the transformation have specific properties. These properties are chosen so that diffusion is allowed only along the direction perpendicular to the local image gradient. If the gradient magnitude is small, isotropic diffusion is performed. We apply this formulation to stereoscopic disparity estimation and we show several experimental results; improvements over a standard approach are clear.