Classifier cascades for support vector machines
Ignas Kukenys, Brendan McCane · 2008
Support vector machines (SVMs) are a binary classification technique with a growing popularity in the field of machine learning. While SVMs have shown to deliver good classification performance, in itpsilas original formulation the technique can be computationally complex and therefore slow at run-time. In this paper we review and compare two approximation techniques that address the speed problem by approximating the decision function of the SVM with a chosen number of vectors (often referred to as reduced set vectors, RSV). We construct cascades of such approximations and use them for object detection, measuring their ability to early reject non-objects and their average time taken. We then suggest a hybrid approach which combines the two techniques and further improves the performance of the SVM cascade.