An Intensity-Based, Coarse-to-Fine Approach to Reliably Measure Binocular Disparity

Abstract

Disparity estimation is an ill-conditioned problem: at certain locations, the disparity vectors cannot be estimated solely using local intensity patterns. According to the strategy they use to solve this problem, the algorithms to estimate disparity vectors can be divided into two classes: the feature-based and intensity-based algorithms. In this paper, a new strategy is proposed. First, the disparity vectors are estimated, and then the unreliable and unfeasible vectors are eliminated. The method bears some resemblance with methods of both classes. In principle, the new approach is intensity-based, but it will be proved that because of the elimination process it determines the features on a binocular basis: the binocular raw primal sketch is constructed. A computational technique is developed for this new strategy. This technique is based on the simplex algorithm and its associated sensitivity analysis. Methods to reduce the required storage space and computation time are discussed. Numerical results are given for a random dot stereogram and an artificially generated stereo pair. Qualitative results are given for natural stereo pairs.