Interactive Learning a Person Detector: Fewer Clicks - Less Frustration
Peter M. Roth, Helmut Gräbner, Christian Leistner, Martin Winter, Horst Bischof · 2008
To train a general person detector a huge amount of training samples is required to cope with the variability in the persons ’ appearance and all possible backgrounds. Since this data is often not available we propose an interactive learning system, that enables an efficient training of a scene specific person detector. For that purpose we apply a two stage approach. First, a general detector is trained autonomously from labeled data. Later on this detector is improved and adapted to a specific scene by user interaction. Thus, only highly valuable samples are selected and only a small number of updates is necessary to adapt to a specific scene. In particular, for learning we apply off-line boosting in the first stage and on-line boosting in the second stage. Since we use the same underlying representation for both methods, we can efficiently re-train an existing classifier. In the experiments we applied the proposed approach for different scenarios and showed that the detection results (recall and accuracy) can be significantly improved by hand-labeling only a few novel samples.