high performance computing on graphics processing units: hgpu.org

We present a high-speed dense stereo algorithm that achieves both good quality results and very high disparity estimation throughput on the graphics processing unit (GPU). The key idea is to make use of directional center-biased support windows to strike a good quality balance between homogeneous areas and depth discontinuities, and it is instantiated by a truncated separable Laplacian kernel approximation. To boost its real-time speed and tackle the computation precision problem on GPUs, a number of critical optimization schemes have been proposed. Evaluation using the benchmark Middlebury stereo database shows that our approach is better than previous GPU-based local stereo methods and it even outperforms some methods using global optimization techniques. Our optimized implementation completely running on an Nvidia GeForce 7900 graphics card achieves over 668 million disparity estimations per second (Mde/s) including all the overhead, which is about 2.3 to 13.4 times faster than the GPU-based solutions ever reported.