high performance computing on graphics processing units: hgpu.org

State of the art local stereo correspondence algorithms that adapt their supports to image content allow to infer very accurate disparity maps often comparable to algorithms based on global disparity optimization methods. However, despite their effectiveness, accurate local approaches based on this methodology are also computationally expensive and several simplifications aimed at reducing their computational load have been proposed. Unfortunately, compared to the original approaches, the effectiveness of most of these simplified techniques is significantly reduced. In this paper, we consider an efficient and accurate algorithm referred to as Fast Bilateral Stereo (FBS) that enables to efficiently obtain results comparable to state of the art local approaches describing its mapping on GPUs with CUDA. Experimental results on two NVIDIA GPUs show that our CUDA implementation delivers, on standard stereo pairs, accurate and dense disparity maps in near real-time achieving speedup greater than 100X with respect to the equivalent CPU-based implementation.