A Generalized Discriminative Muitiple Instance Learning for Multimedia Semantic Concept Detection

In the paper we present a generalized discriminative multiple instance learning algorithm (GD-MIL) for multimedia semantic concept detection. It combines the capability of the MIL for automatically weighting the instances in the bag according to their relevance to the positive and negative classes, the expressive power of generative models, and the advantage of discriminative training. We evaluate the GD-MIL on the development set of TRECVID 2005 for high-level feature extraction task. The significant improvement is observed using the GD-MIL over the benchmark. The mean of AP´s over 10 concepts using the GD-MIL is 4.18% on the validation set and 3.94 % on the evaluation set. As the comparison, they are 2.12% and 2.63% for the benchmark, correspondingly.