Spmvp: Spatial Patchmatch Stereo With Virtual Pixel Aggregation
NEURAL INFORMATION PROCESSING (ICONIP 2017), PT III(2017)
摘要
Stereo matching is one of the critical problems in the field of computer vision and it has been widely applied to 3D Reconstruction, Image Refocusing and etc. Recently proposed PatchMatch (PM) stereo algorithm effectively overcomes the limitation of integer-value within the support window but it is still inferior in twofold: (1) view propagation of PM stereo algorithm generally yields underwhelming particle propagation; (2) it still suffers from a coarse performance in textureless regions. To mitigate these weaknesses, a Spatial-PM stereo algorithm without view propagation is proposed for improving the original one at first. Then a virtual pixel based cost aggregation framework with two sped-up strategies is proposed for tackling the problem of textureless mismatching. Jointing the two incremental improvements, we name the novel one as Spatial PatchMatch Stereo with Virtual Pixel Aggregation (SPMVP). Experiments show that SPMVP achieves superior results than other four challenging PM based stereo algorithms both in integer & subpixel level accuracy on all 31 Middlebury stereo pairs; and also performs better on Microsoft i2i stereo videos.
更多查看译文
关键词
PatchMatch stereo,Particle propagation,Virtual pixel,Sped-up strategy,Subpixel accuracy
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络