On output-processing in face detectors

A face detector will classify many thousands of windows within an image, generating multiple positive classifications in the proximity of a face. Consequently, practical detectors must perform output-processing to give a single hypothesis per face. This problem is considered in this paper, in order to facilitate comparisons of detectors and provide practitioners with guidelines for constructing these systems. This is achieved first by reviewing previous approaches and introducing a generic framework to encompass these, then through a systematic exploration of several of the choices involved in constructing such a system.