LIG and LIRIS at TRECVID 2008: High Level Feature Extraction and Collaborative Annotation

This paper describes our participations of LIG and LIRIS to the TRECVID 2008 High Level Features detection task. We evaluated several fusion strategies and especially rank fusion. Results show that including as many low-level and intermediate features as possible is the best strategy, that SIFT features are very important, that the way in which the fusion from the various low-level and intermediate features does matter, that the type of mean (arithmetic, geometric and harmonic) does matter. LIG and LIRIS best runs respectively have a Mean Inferred Average Precision of 0.0833 and 0.0598; both above TRECVID 2008 HLF detection task median performance. LIG and LIRIS also co-organized the TRECVID 2008 collaborative annotation. 40 teams did 1235428 annotations. The development collection was annotated at least once at 100\%, at least twice at 37.6\%, at least three times at 3.99\% and at least four times at 0.06\%. Thanks to the active learning and active cleaning used approach, the annotations that were done multiple times were those for which the risk of error was maximum.