LIG at MediaEval 2012 affect task: use of a generic method

This paper describes the LIG participation to the MediaEval 2012 Affect Task on violent scenes’ detection in Hollywood movies. We submitted four runs at the shot level: hierarchical fusion of descriptors and classifier combinations (LIG-4), the same with conceptual feedback (LIG-3), and the same two with reranking (LIG-2 and LIG-1). Our reference run obtained a performance slightly above the median with the official MAP@100 metric. The temporal re-ranking brings a significant improvement on the overall (classical) MAP but has almost no effect on the MAP@100. The conceptual feedback does not improve the overall MAP but it improves the precision in the head of the returned list (MAP@100 or P@100).