Probabilistic scene analysis for robust stereo correspondence

Abstract

Most area-based approaches for stereo correspondence are leading to a large set of non-correct matches in the generated disparity-map. These are mainly caused by low textured areas, half occlusions, discontinuities in depth and the occurrence of repetitive patterns in the observed scene. This paper proposes a novel framework where non salient regions inside the stereo pair are identified previously to the matching, whereat the decision about the involvement of particular areas in the correspondence analysis is realized based on the fusion of separate confidence maps. They describe the possibility for a correct matching based on different criteria. © 2009 Springer Berlin Heidelberg.