提出了一种交互式的时空一致性视频抠取方法,能够有效地从视频序列中提取出移动的前景目标.只需在视频体上简单地勾画时空的前景和背景线条,然后构造一个拉普拉斯方程驱动线条在三维的视频体上进行扩散,得到一个粗糙的抠取结果.随后,采用一种新颖的保持时空一致性的抠取技术,利用局部的统计信息和邻域信息,通过少数的迭代即可收敛到全局最优的抠取结果.最后,最优化一个新的全局代价函数在整个三维体上重建前景颜色,如实地保留了抠图结果的时空连贯性.算法的每一步计算都可转化为线性方程组进行求解,因此对于千万像素级的视频数据,也能快速得到高质量的抠取结果.通过对复杂的视频序列进行测试,展示了高质量的抠图结果和算法的高效性.

This paper presents an interactive matting approach for efficiently extracting alpha mattes and foreground objects from video sequences. Beginning from user-specified strokes across space and time, the paper formulates their expansions in the video volume as a Laplacian equation, resulting in a coarse alpha matte. It then employs a novel spacetime alpha matting technique that makes use of local statistics and neighboring information and converges to a globally optimal alpha matte in few iterations. Finally, the paper derives a new global cost function to reconstruct the foreground color of the whole video volume, which faithfully preserves the spatio-temporal coherence. The computation in each step can be reformulated as solving a set of linear equations, allowing users to quickly extract high-quality alpha matte and foreground objects, even for data sets with ten million pixels. Experimental results on complex natural video sequences demonstrate the high quality and efficiency of the proposed approach.

Supported by the National High-Tech Research and Development Plan of China under Grant No.2006AA01Z314 (国家高技术研究发展计划(863)