EURECOM at TrecVid 2011: The Light Semantic Indexing Task

This year EURECOM participated in the TRECVID light Semantic Indexing (SIN) Task for the submission of four different runs for 50 concepts. Our submission builds on the runs submitted last year at the 2011 SIN task with the first two runs following the same pattern as those of last year. The details of 2011 system can be found in [8]. One of our run adds uploaders bias to the pool of visual features while another run is prepared in collaboration with ECNU. Our basic run adds visual features based on larger vectors to the pool of features of last year’s base run. Larger dictionaries provide a finer representation of the visual/clustering space and increase the precision of the retrieval task. Like in last year’s submission we add two global descriptors to visual features with one capturing temporal statistics along each shot and the other capturing salient details or gist of a keyframe. Then we add textual metadata based information that has been provided with the 2012 video database to the visual features. We further benefit from the metadata by including uploaders bias to increase scores of videos uploaded by same users. The runs are composed as follows: