Try Everything: Detecting Occluded Faces by Cascading Outrageous Proposal Generation and Deep Convolutional Neural Network

The performance of face detection in heavy occlusion conditions (e.g., occluded by masks) will drop dramatically since facial cue is missing or polluted by occluders. To address this issue, we propose a novel method by cascaded outrageous proposal generation and deep convolutional neural network (DCNN) via two major modules. First, the proposal generation module assembles many fast face detectors with low threshold to generate massive candidate proposals in a high recall rate. Then, the verification module, which is trained with a ResNet network, jointly performs face classification and regression tasks to identifies a proposal to be real face or not and refine its location respectively. Experiments on OFD-UCI and our collected masked face dataset MAFA demonstrate that our approach effectively outperforms the three representative face detectors by at least 3.3% and 11.6% in detecting crowd occluded faces and masked faces respectively.