Efficient video cutout based on adaptive multilevel banded method

Although many efficient image cutout approaches have been proposed, it is still a challenge to cut out foreground objects from a video clip because of the costs of time and memory consumption and the difficulty in maintaining temporal coherence for the cutout results. In this paper, we propose a novel and efficient video cutout approach by extending a multilevel banded image segmentation method to video cutout. In our method, the segmentation results are propagated frame by frame at the lowest resolution level based on optical flow, and are refined in image pyramids level by level through banded graph cuts. We construct an adaptive narrow band to ensure that the segmentation result falls exactly into the narrow band during propagation. Experimental results show that the video cutout approach based on adaptive multilevel bands dramatically reduces time and memory costs, and achieves ideal segmentation results.