Video object cut in complex scenes

Bo Li,Baozong Yuan,Jiangfeng Chen,Zhenjiang Miao

DOI: https://doi.org/10.1109/ICOSP.2008.4697391

2008-01-01

Abstract:Extracting the moving interesting objects accurately in complex scenes is one of the most challenging problems in computer vision. In this paper, we introduce a novel approach to cutout interesting moving objects with userpsilas interaction, which partitions pre-segmentation regions into foreground and background while preserving temporal coherence. First frame of video is divided into several pre-segmentation regions, which are used as nodes of directed graph. After graph is constructed, user draws several simple lines as seeds in background and foreground in image. Graph cuts algorithm is used to divide the image into two parts: background and foreground by the seed pixels. When next frame comes in, Bhattacharyya distance and spatial adjacent condition are employed to measure the similarity between pre-segmentation regions in current frame and regions of prior segmentation result in prior frame. Then we modify the weight of graph link, cut object out. If the result is goods, next frame is dealt with. Otherwise, there exit some errors, user can draw the seed line again to segment interesting object accurately. Experiments show that our approach achieves nice performance.