It Takes Two to Tango: Cascading Off-the-Shelf Face Detectors

Siqi Yang, Arnold Wiliem, Brian C. Lovell

Abstract: Recent face detection methods have achieved high detection rates in unconstrained environments. However, as they still generate excessive false positives, any method for reducing false positives is highly desirable. This work aims to massively reduce false positives of existing face detection methods whilst maintaining the true detection rate. In addition, the proposed method also aims to sidestep the detector retraining task which generally requires enormous effort. To this end, we propose a two-stage framework which cascades two off-the-shelf face detectors. Not all face detectors can be cascaded and achieve good performance. Thus, we study three properties that allow us to determine the best pair of detectors. These three properties are: (1) correlation of true positives; (2) diversity of false positives and (3) detector runtime. Experimental results on recent large benchmark datasets such as FDDB and WIDER FACE support our findings that the false positives of a face detector could be potentially reduced by 90% whilst still maintaining high true positive detection rate. In addition, with a slight decrease in true positives, we found a pair of face detector that achieves significantly lower false positives, while being five times faster than the current state-of-the-art detector.