Boosting with Multiple Classifier Families

This paper demonstrates the importance of creating an even playing field between weak classifiers and classifier families in the RealBoost boosting algorithm. Classifier families are constructed based on Haar-like features in various color spaces, which are then trained simultaneously in RealBoost to create a strong classifier rule. It is shown that the usual method for minimising error at each RealBoost round may express a bias against some weak classifier families. A particular bias toward overfitting features is found. An initial method for achieving parity between families of weak classifiers is applied to improve classification. Classification results for various groups of classifier families are shown on pedestrian and sign detection tasks. Particular attention is given to the effect of recently proposed model improvements, including response binning and smoothed response binning. The final system yields significantly lower error rates on classification tasks, and demonstrates the value of color information within the context of the improved methods.