Towards a stochastic depth maps estimation for textureless and quite specular surfaces

Abstract

The human brain is constantly solving enormous and challenging optimization problems in vision. Due to the formidable meta-heuristics engine our brain equipped with, in addition to the widespread associative inputs from all other senses that act as the perfect initial guesses for a heuristic algorithm, the produced solutions are guaranteed to be optimal. By the same token, we address the problem of computing the depth and normal maps of a given scene under a natural but unknown illumination utilizing particle swarm optimization (PSO) to maximize a sophisticated photo-consistency function. For each output pixel, the swarm is initialized with good guesses starting with SIFT features as well as the optimal solution (depth, normal) found previously during the optimization. This leads to significantly better accuracy and robustness to textureless or quite specular surfaces.