Performace Evaluation of Similarity Metrics for Stereo Corresponce Problem

Abstract

Lately, many promising stereo matching algorithms mostly relying on global energy minimization framework have been proposed which model disparity surface as a MRF and deploy various optimization techniques to produce a MAP approximation of MRF model corresponding to minimum energy level of the matching cost function. However, most of these methods mainly concentrate on how to minimize the cost function rather than attempting to improve its definition. We believe that regarding all advances made recently in cost function optimization area and also deeming results of some new experiments proving that even calculating the globally optimal solutions of energy minimization using re-weighted belief propagation approach does not solve most of the problems associated with common BP or GC based methods, the next logical step for further improvement is to devise cost functions which are better trimmed to cope with stereo matching major inherent problems. In this paper, we illustrate the importance of well-defined cost functions by incorporating six similarity measures, namely, Manhattan (LI), Euclidean (L2), Canberra, Bray-Curtis, Squared Chord and Square Chi-Squared into matching cost function and comparing the obtainable results. To accomplish this, we apply the test data and ground truth supplied on Middlebury college stereo webpage. Finally, we analyze the obtained results and discuss its impact on the design of enhanced matching cost functions.