Finding hard faces with better proposals and classifier

Abstract

Recent studies witnessed that deep CNNs significantly improve the performance of face detection in the wild. However, detecting faces with small scales, large pose variations, and occlusions is still challenging. In this paper, to detect challenging faces, we present a boosted faster RCNN (F-RCN) version with an enhanced region proposal network (eRPN) module and newly introduced hard example mining strategies. The eRPN module generates better proposals than traditional RPN by integarating semantic information into the input feature maps. Two hard example mining strategies, i.e., online hard proposal mining (OHPM) and offline hard image mining (OHIM), are proposed to train better classifier. The OHPM can effectively sample quality and diversity of hard positive examples, which is important for detecting hard faces like tiny faces. The OHIM further boosts the classifier to detect hard faces via an auxiliary fine-tuning on a small proportion of training data. Experimental results on the FDDB, WIDER FACE, Pascal Faces, and AFW datasets show that our method significantly improves the faster-RCNN face detector and achieves performance superior or comparable to the state-of-the-art face detectors.