Efficient GPU-based Graph Cuts for Stereo Matching
Computer Science Department, University of California, Los Angeles
IEEE Embedded Vision Workshop (EVW2013 in conjunction with CVPR2013), 2013
@article{choi2013efficient,
title={Efficient GPU-based Graph Cuts for Stereo Matching},
author={Choi, Young-kyu and Park, In Kyu},
year={2013}
}
Although graph cuts (GC) is popularly used in many computer vision problems, slow execution time due to its high complexity hinders wide usage. Manycore solution using Graphics Processing Unit (GPU) may solve this problem. However, conventional GC implementation does not fully exploit GPU’s computing power. To address this issue, a new GC algorithm which is suitable for GPU environment is presented in this paper. First, we present a novel graph construction method that accelerates the convergence speed of GC. Next, a repetitive block-based push and relabel method is used to increase the data transfer efficiency. Finally, we propose a low-overhead global relabeling algorithm to increase the GPU occupancy ratio. The experiments on Middlebury stereo dataset shows that 5.2X speedup can be achieved over the baseline implementation, with identical GPU platform and parameters.
May 29, 2013 by hgpu
