THU and ICRC at TRECVID 2007

摘要: Shot boundary detection The shot system in 2007 is basically the same as that of last year. We make three major modifications this First, CUT detector and GT use block based RGB color histogram with different parameters instead ones. Secondly, we add a motion module to remove false alarms caused by camera or large object movements. Finally, post-processing on SIFT feature after both detector. evaluation results show all these bring performance improvements system. brief introduction each run shown following table: Run_id Description Thu01 Baseline system: RGB4_48 for detector, no sift post-processing, only using development set 2005 training Thu02 Same algorithm thu01, but RGB16_48 Thu03 thu02, Thu04 thu03, Motion Thu05 thu04, Thu06 thu05, processing Thu09 thu11 Thu13 Thu14 trained data from 2003-2006 High-level extraction try novel approach, Multi-Label Multi-Feature learning (MLMF learning) learn joint-concept distribution regional level an intermediate representation. Besides, improve our Video diver indexing designing new features, comparing algorithms exploring fusion algorithms. Based efforts improving feature, algorithms, achieve top HFE