Towards the Optimal Training of Cascades of Boosted Ensembles

Abstract

Cascades of boosted ensembles have become a popular technique forface detection following their introduction by Viola and Jones. Researchershave sought to improve upon the original approach by incorporatingnew techniques such as alternative boosting methods, feature sets,etc. We explore several avenues that have not yet received adequateattention: global cascade learning, optimal ensemble construction,stronger weak hypotheses, and feature filtering. We describe a probabilisticmodel for cascade performance and its use in a fully-automated trainingalgorithm.