This paper prroposes a novel semiautomatic method for semantic video object extraction. The proposed method aims at improving the semiautomatic segmentation by integrating various semantic properties of the video. Object region in the first frame is extraded by a user input polygon. Then

a tracking scheme based on backward block-marching is used to automatically extract object region in the rest of frames for general objects. Translational rigid objects and slowly changing objects are specially considered in the tracking. Finally

a mask refinement algorithm based on peer group is applied to extract accurate object boundary and insure reliaible tracking. Accurate and consistent results are obtained in the experiments by the proposed method.