Multi-modal classifier fusion for video shot content retrieval

In this paper we present a new chromosome to solve the problem of classifier fusion using genetic algorithm. Experiments are conducted in the context of TRECVID. In particular we focus on the feature extraction task that consists in retrieving video shots expressing one of predefined semantic concepts. Three modalities (visual, textual and motion) and two features per modality are used to describe the content of a video shot. Thus, we require fusion techniques to efficiently manage all these heterogeneous sources of information. A first step achieves the classification per feature and concept, then a genetic algorithm is used to efficiently fuse the output of all classifiers. For this purpose, a dynamic binary tree is proposed to model the novel chromosome for hierarchical fusion.