Generalized Multiclass AdaBoost and Its Applications to Multimedia Classification

AdaBoost has received considerable attention in the vision and multimedia research community in recent years. It is originally designed for two-class classification problems. To handle multiple classes, many AdaBoost extensions have been developed primarily based on various schemes for reducing multiclass classification to multiple two-class problems. From a statistical prospective, AdaBoost can be viewed as a forward stepwise additive model using an exponential loss function. In this paper, we derive a generalized form of AdaBoost for multiclass classification based on a multiclass exponential loss function. To prove its effectiveness, we benchmarked a number of multimedia problems of different nature. Experimental results show that the new boosting algorithm outperforms other multiclass alternatives. In addition, the generalized boosting algorithm can be used to either boost a multiclass classifier, or build a multiclass classifier from a binary one.