Chinese Journal of Computers   Full Text
Title	Real-Time Post-Processing for Online Video Segmentation
Authors	ZHONG Fan1) QIN Xue-Ying1),2) CHEN Jia-Zhou1) MO Ming-Zhen1) PENG Qun-Sheng1)
Address	1)(State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou 310027) 2)(Department of Computer Science, Shandong University, Jinan 250101)
Year	2009
Issue	No.2(261—267)
Abstract &   Background	Abstract Applications of online video segmentation usually need to do post-processing in order to remove mis-segmentation and suppress flicking. Traditional matting-based methods are too slow, while simply blur the (foreground/background) boundary not only cause over-blur but also can’t remove mis-segmentation. This paper proposes a novel post-processing method. For each pixel around the boundary, a local color model is first estimated through a new fast clustering algorithm, which is designed specially for color clustering. Mis-segmentation is then removed by re-estimating the alpha value for each pixel according to its local color model. In order to improve the consistence of result, an adaptive edge model is applied as a smooth constraint. The edge model can adjust the center and width of the transition region according to the local context of each pixel, and this way prevents the boundary from being over-blurred. The proposed method is very fast, and can meet the requirement of online video segmentation very well. Keywords alpha value; video segmentation; matting; fast clustering; edge model Background High-quality online video segmentation is very difficult due to the scene complexity and limited computational resource. Recent works almost all regard it as a MRF-based optimization problem, and then solve with max-flow/min-cut, which yields a binary segmentation that is global optimal but less satisfying nearby the foreground/background boundary, so post-processing must be applied in order to suppress flicking. Previous works do post-processing by either feathering or matting, both are not designed specially for online video segmentation and perform not very well. This paper proposed a post-processing method that fully considering the requirement of online video segmentation. The proposed method can effectively remove the mis-segmentation nearby the boundary and adaptively smooth the boundary without causing over-blur. In addition, it is very fast and brings only a little computational burden for real-time applications. This paper is supported by the National High Technology Research and Development Program (863 Program) of China under grant No.2007AA01Z326; the National Basic Research Program (973 Program) of China under grant No.2002CB312101.