A cooperative learning strategy for interactive video search

The goal of this paper is to develop a learning strategy for interactive video search that can effectively mitigate the burden on users without decreasing search performance. Taking SVM as underlying learner, a cooperative training strategy is proposed for learning a ranking function, in which semi-supervised learning procedure is started with a combination of a few positive training seeds and a relative large set of unlabeled data. The main merit of the proposed framework is its ability to mine automatically training samples from previous answer set and to refine gradually ranking model during cooperative training phase. In addition, as an extension of the proposed framework, multiple modalities can be potentially combined for effectively learning user´s query intention. Following the guideline of TRECVID´ 06 video search task, we validate the effectiveness of our proposed method.