SINAI at TRECVID 2007

This paper describes the first participation of the SINAI group of the University of Jaén in TRECVID 2007. We have only participated in the automatic search task. Our approach is a very simple system made up of three main modules: the text-based retrieval subsystem, the image-based retrieval subsystem and the fusion module. We have submitted several runs exploring fusion of both textual and visual lists. Also the effect of text expansion in the topics has been of our concern: • F C 1 SINAI 1: Baseline run using only ASR/MT text features and topics without text expansion. • F C 2 SINAI 2: Baseline run using only ASR/MT text features and topics with text expansion. • F C 2 SINAI 3: Baseline run using only visual test data. • F C 2 SINAI 4: Run using text and visual features mixing the text shots recovered from text-based retrieval subsystem and image shots recovered from image-based retrieval subsystem, following the RoundRobin[7] algorithm. • F C 2 SINAI 5: Run using text and visual features mixing the lists from both subsystems, including the 75 percent of best text shots recovered by the textbased retrieval subsystem and the 25 percent of best image shots recovered by the image-based retrieval subsystem, with topics without text expansion. • F C 2 SINAI 6: Like the previous one but using topics with text expansion. With these experiments we have tried to establish a baseline study of automatic search task of TRECVID, in order to improve the system in the future. The results of the runs submitted indicate that using visual data does not seem to improve the overall results compared to the text-only baseline. On the other hand, the expansion of the text topics with synonyms does not improve the baseline result either.