Two-step Belief Propagation for Stereo Matching
Belief Propagation is a powerful inference mechanism for stereo matching. However like other global discrete optimization algorithms, such as Graph Cuts, it also has a severe limitation that increasing the number of discretized disparities quickly makes it computationally infeasible, especially for scenes with large disparity range. This limitation makes the BP difficult in practice to get finer matching results (sub-pixel disparity level) and often yield staircase effect for slanted or curved surface. To tackle this problem, we propose a simple 2-step BP algorithm in this work. We use BP to get some initial disparity matching(pixel level) in the first step, then BP is used again to refine the initial matching to sub-pixel level in the second step. This method makes Pixel matching and sub-pixel matching all cast in a global minimization framework. Experiments show that our 2-step algorithm can efficiently get finer matching results to substantially reduce the staircase effect
stereo matching Belief Propagation
Wang Xiaoming Zhang Chunming
department of Mathematics and Physics Shandong Jiaotong University Jinan, China
国际会议
太原
英文
499-502
2011-02-26（万方平台首次上网日期，不代表论文的发表时间）