Stochastic stereo matching over scale

摘要

A stochastic optimization approach to stereo matching is presented. Unlike conventional correlation matching and feature matching, the method provides a dense array of disparities, eliminating the need for interpolation. First, the stereo-matching problem is defined in terms of finding a disparity map that satisfies two competing constraints: (1) matched points should have similar image intensity, and (2) the disparity map should vary as slowly as possible. These constraints are interpreted as specifying the potential energy of a system of oscillators. Ground states are approximated by a new variant of simulated annealing, which has two important features. First, the microcanonical ensemble is simulated using a new algorithm that is more efficient and more easily implemented than the familiar Metropolis algorithm (which simulates the canonical ensemble). Secondly, it uses a hierarchical, coarse-to-fine control structure employing Gaussian or Laplacian pyramids of the stereo images. In this way, quickly computed results at low resolutions are used to initialize the system at higher resolutions.