Partition sampling for active video database annotation

Annotating a video-database requires an intensive human effort that is time consuming and error prone. However this task is mandatory to bridge the gap between low-level video features and the semantic content. We propose a partition sampling active learning method to minimize human effort in labeling. Formally, active learning is a process where new unlabeled samples are iteratively selected and presented to teachers. The major problem is then to find the best selection function that maximizes the knowledge gain acquired from new samples. In contrast with existing active learning approaches, we focus on the selection of multiple samples. We propose to select samples such that their contribution to the knowledge gain is complementary and optimal. Hence, at each iteration we ensure to maximize the knowledge gain. Our method offers many advantages; among them the possibility to share the annotation effort among several teachers.