Abstract
Detection of shot boundary plays important roles in many video applications. Herein, a novel method on shot boundary detection from compressed video is proposed. Firstly, we extract several local indicators from macroblocks, and these features are used in determining candidate cuts via rule-based decision making into five sub-spaces. Then, global indicators of frame similarity between start and end frames of candidate cuts are examined, using fast phase correlation on cropped DC images. Gradual transitions like fade and dissolve as well as combined shot cuts are also identified in compressed domain. Experimental results on the test data from TRECVID 2007 have demonstrated the effectiveness and robustness of our proposed methodology. Moreover, our submissions can achieve nearly 5 times faster than real-time video play (25 frames/s) due to the nature of its compressed-domain processing, achieving additional advantages in terms of processing speed and computing costs.