Abstract

The application of semi-supervised learning algorithms to large scale vision problems suffers from the bad scaling behavior of most methods. Based on the Expectation Regularization principle, we propose a novel semi-supervised boosting method, called SERBoost that can be applied to large scale vision problems. The complexity is mainly domi- nated by the base learners. The algorithm provides a margin regularizer for the boosting cost function and shows a principled way of utilizing prior knowledge. We demonstrate the performance of SERBoost on the Pas- cal VOC2006 set and compare it to other supervised and semi-supervised methods, where SERBoost shows improvements both in terms of classifi- cation accuracy and computational speed.