Partition sampling: an active learning selection strategy for large database annotation

Annotating a video database requires an intensive, time consuming and error prone human effort. However, this is a mandatory task to efficiently analyze multimedia contents. We propose an new selection strategy for active learning methods to minimize human effort in labeling a large database of video sequences. Formally, active learning is a process where new unlabeled samples are iteratively selected, presented to users for annotation and added to the training set. The major problem is then to find the best selection function to quickly reach high classification accuracy. We will show that existing active learning approaches using selective sampling do not maintain their performances when the number of selected samples per iteration increases. The presented selection strategy attempt to provide a solution to this problem. In practice, selecting many samples offers many advantages when dealing with a large amount of data; among them the possibility to share the annotation effort between several users. Finally we attempt to tackle the more realistic and challenging task of multiple label annotation. This would reduce to greater extend the human effort for labeling.