CUDA-based GPU Implementation of Hierarchical Belief Propagation for Fast Stereo Matching
Gwangju Institute of Science and Technology (GIST), 261 Cheomdan-gwagiro, Buk-gu, Gwangju, 500-712, Korea
International Conference on Embedded Systems and Intelligent Technology (ICESIT) 2011, pp. 209-211, Feb. 2011
@article{jung2011cuda,
title={CUDA-based GPU Implementation of Hierarchical Belief Propagation for Fast Stereo Matching},
author={Jung, J.I. and Ho, Y.S.},
booktitle={International Conference on Embedded Systems and Intelligent Technology (ICESIT) 2011},
year={2011}
}
Stereo matching based on the Markov random field model has a global optimization problem. Solutions of the problem can be inferred by the belief propagation (BP) algorithm. The BP algorithm effectively estimates global solutions, but it takes a very long time to calculate messages. In this paper, we implement the hierarchical BP algorithm on a graphics processing unit (GPU) using compute unified device architecture. We perform memory uploads to the global memory of the GPU, and compute messages of all pixels in parallel. We analyze the performance according to block and image sizes. Experimental results show that our implementation is faster than the conventional loopy BP algorithm and the hierarchical BP algorithm implemented on CPU.
September 6, 2011 by hgpu
