Collaborative learning for image and video annotation

Classical machine learning methods, such as Support Vector Machines, by taking each concept detection as an independent classification problem, can not achieve a sound performance for image and video annotation due to the overfitting problems. Thus, some prior knowledge is required to assist the learning of independent concept detectors, e.g. some concepts look much more alike to each other. In this paper, we assume that visually similar concepts should share resembled detectors. Based on the assumption, Collaborative Learning is proposed, to incorporate cross-concept collaborations into the joint learning of similar detectors over related concepts. Besides the collaborations, different concepts should also perform discriminations for classifying each other. To benefit from different trade-offs between collaboration and discrimination, we propose Multi-Granularity Boosting strategy, where each granularity corresponds to a specific balance between collaboration and discrimination for Collaborative Learning. The ultimate concept detector is an additive model that combines classifiers under different collaboration granularities together. Evaluations on both image and video annotation benchmark demonstrate that our method achieves a superior performance over independent annotation.