A new boostrapping strategy for the adaboost-based face detector
Tat-Jun Chin, David Suter · Adelaide Research & Scholarship (AR&S) (University of Adelaide) · 2005
False alarms occur in the face detection process when non-face sub-images are classified as face images, and are a major problem in face detection. Usually practitioners perform “bootstrapping”to refine the face detector to account for the false alarms (or false positives). For AdaBoost-based face detection (specifically, the method propounded by Viola and Jones in [5]), two approaches of bootstrapping are used– incremental retraining of the entire face detector with the false positives as examples, or construct additional classifier stages to eliminate the false positives. While incremental retraining is costly in terms of training time and effort, additional stages will ultimately slow down the face detector, thus defeating the purpose of constructing the detector in a cascade manner for speed-ups. This technical report presents a novel bootstrapping strategy that although is similar in nature to the second approach, has some modifications with the aim of improving the efficiency of the bootstrapping procedure.