ThesisSemi-supervised learning of multimedia concepts
The objective of this thesis is to develop multimedia indexing techniques which allow to classify and organize various streams of multmedia information. The research will more precisely focus on the combination of several analysis modules, for text, audio and video, in order to build efficient semantic classifiers. Among the possible tracks for investigation, we plan to consider approaches such as active learning, co-training, which seems potentially very useful for this problem. An important point is to optimize the use of manual annotations, so as to minimize the amount of human effort to build and adapt those classifiers, while maximizing the quality and performance of those classifiers.