The stereo matching problem has been under attention of numerous researchers for many years, due to the wide range of applications in computer vision. The MEMOCODE 2013 design contest was aimed to develop a very fast and efficient stereo matching method to infer the depth information for each pixel of a pair stereo image via Belief Propagation on Markov Random Field graph. We first explored different platforms, i.e. FPGA, GPU, multi-core CPU, to find the best one for solving the contest problem. Based on our investigation, we select multi-core system and design and implement a solution on this platform. We use a series of techniques and optimization methods to improve the running time of the stereo matching problem while preserving the error compared to the given reference solution. Our method outperform other team's solutions in both classes: absolute-performance and cost-adjusted-performance.
展开▼
