Multiple PiPs detection in unbounded video stream

摘要

Picture-in-picture detection aims to detect the video clips that are embedded as a sub-window into a background video with similar semantic meaning, which plays an important role in various multimedia applications such as video copy detection and semantic relationship mining. While some PiP detection algorithms have been given in existing work, less effort has been made on the detection of multiple PiP clips in unbounded video streams. In this paper, we propose a successive-group prior of both PiP and non-PiP clips based on the statistics of single PiP detection accuracy. Using this prior, a new algorithm, which involves a series of geometrical and temporal constraints for handling both the spatial and temporal localizations, is proposed for detecting multiple PiP clips. The extensive experimental results show that the proposed algorithm can efficiently localize the spatial positions and precisely detect the starting and ending time stamps of multiple PiP clips.