Refining and Expanding WordNet for Video Retrieval

Recent research in video retrieval has focused on automated, high-level feature indexing on shots or frames. One of the most important applicationof such indexing is to support precise video retrieval. We report on extensions ofthis semantic indexing on news video retrieval. First, we utilize extensive queryanalysis to relate various high-level features and query terms by matching the tex-tual description and context in a time-dependent manner. Second, we introducea framework to effectively fuse the relation weights with the detectors’ confi-dence scores. This results in individual high level features that are weighted on aper-query basis. Tests on the TRECVID 2005 dataset shows that the above twoenhancements yield significant improvement in performance over a correspond-ing state-of-the-art video retrieval baseline.