Improving depth estimation using colour information in stereo vision

Abstract

This paper presents a multi-scale stereo correspondence algorithm that incorporates different modules corresponding to every stage. A correspondence method is proposed based on minimizing an energy function using simulated annealing. Multiresolution techniques are also introduced to improve both response time and the quality of results. A technique of calculating the distance between two colours is defined for comparing colour features in HLS model. This measure not only highlights the difference in the H component but also considers the differences in S and L values. A metrics for evaluating the quality of a disparity map is defined. Experimental results with real images are presented to illustrate the approach taken.