Abstract: With the advent of MPEG 4 and MPEG 7, the segmentation of SVO （semantic video object） becomes the common task of video encoding and retrieval And object shape retrieval is in demand in MPEG 7 But current algorithms can not satisfy these requirements, especially on current MPEG 1/2 video data Based on the proposed algorithm WaterShed, a modified method is developed to extract SVO semi automatically, which can get better results when applied to MPEG 1 video stream The major contributions are that the element count is introduced as a classification parameter and that the segmentation is carried out twice on the compressed area and decoded image In experiments, the modified algorithm is more robust and quicker than the proposed one It can extract VO （video object） with velvet contour, which can be used in the shape retrieval and the important object retrieval