Skip to Main Content
In this paper, we briefly review AdaBoost and expand on the discrete version by building weak classifiers from a pair of biased classifiers which enable the weak classifier to abstain from classifying some samples. We show that this approach turns into a 3-bin real AdaBoost approach where the bin sizes and positions are set by the bias parameters selected by the user and dynamically change with every iteration which make it different from the traditional real AdaBoost. We apply this method to face detection more specifically the Viola-Jones approach to detecting faces with Haar-like features and empirically show that our method can help improving the generalization ability by reducing the testing error of the final classifier. We benchmark the results on the MIT+CMU database.