Active sampling via tracking

To learn an object detector labeled training data is required. Since unlabeled training data is often given as an image sequence we propose a tracking-based approach to minimize the manual effort when learning an object detector. The main idea is to apply a tracker within an active on-line learning framework for selecting and labeling unlabeled samples. For that purpose the current classifier is evaluated on a test image and the obtained detection result is verified by the tracker. In this way the most valuable samples can be estimated and used for updating the classifier. Thus, the number of needed samples can be reduced and an incrementally better detector is obtained. To enable efficient learning (i.e., to have real-time performance) and to assure robust tracking results, we apply on-line boosting for both, learning and tracking. If the tracker can be initialized automatically no user interaction is needed and we have an autonomous learning/labeling system. In the experiments the approach is evaluated in detail for learning a face detector. In addition, to show the generality, also results for completely different objects are presented.