Robust Fast Belief Propagation for Real-time Stereo Matching

There are growing needs in computer vision applications for stereo matching, requiring not only accuracy and robustness but also fast processing speed. Global matching algorithms such as belief propagation (BP) shows remarkably robust results with presence of occlusions, textureless region, image noise. In this paper, a novel approach that combines merits of global matching and robust matching cost is proposed. We modeled illumination difference as biased intensity model. Applying window-based matching cost which are insensitive to intensity bias, erroneous matching results under different illumination can be prevented. Moreover, adoption of memory-efficient fast belief propagation enables high speed processing with aid of parallel computing architecture. Experimental result demonstrates that our method is robust under various illumination difference.