A novel framework for semi-automatic video object segmentation

Abstract

A novel framework for semi-automatic video object segmentation is proposed to facilitate user interaction and improve the performance of the system. The proposed framework scans the video sequence more than once, featured as multi-pass scan. In each pass, the sub-shots detected by a self-supervisor are processed under a novel bi-directional auto-tracking algorithm that depends on not only temporal but also spatial information and is capable of dealing with occlusion/disocclusion. In order to merge the two results obtained by bi-directional tracking, an online learner is introduced. These new features make the proposed framework flexible, efficient and able to extract video objects with pixel accuracy. The results for several MPEG test sequences show that this scheme performs well in practice.