A novel bootstrapping method for positive datasets in cascades of boosted ensembles

Abstract

We present a novel method for efficiently training a face detector using large positive datasets in a cascade of boosted ensembles. We extend the successful Viola-Jones [1] framework which achieved low false acceptance rates through bootstrapping negative samples with the capability to also bootstrap large positive datasets thereby capturing more in-class variation of the target object. We achieve this form of bootstrapping by way of an additional embedded cascade within each layer and term the new structure as the Bootstrapped Dual-Cascaded (BDC) framework. We demonstrate its ability to easily and efficiently train a classifier on large and complex face datasets which exhibit acute in-class variation.