Reactive learning strategy for AsymBoost based face detectors

The face detection problem is certainly one of the most studied problems in the field of computer vision. It finds indeed application in the human-computer interaction field, automotive, etc. but especially in video surveillance and security systems. In the last years, AdaBoost-based systems showed good performance in both detection rate and computation time allowing its exploitation in realtime face detectors. Although effective, the natural asymmetry, brought by the problem of separating objects from the rest of the world, highlighted the limits of such an algorithm. To overcome this limit the AsymBoost version has been introduced to better distinguish the patterns of the two classes. In this paper, we further optimize the learning strategy by extending the AsymBoost cascade algorithm by introducing a reactive control of the asymmetry at both cascade and classifiers learning stages. The results will point out how the proposed strategy cuts the false negatives by keeping low the false positives.