Single-Stage Joint Face Detection and Alignment

Abstract

In practice, there are huge demands to localize faces in images and videos under unconstrained pose variation, illumination change, severe occlusion and low resolution, which pose a great challenge to existing face detectors. This challenge report presents a single-stage joint face detection and alignment method. In detail, we employ feature pyramid network, single-stage detection, context modelling, multi-task learning and cascade regression to construct a practical face detector. On the Wider Face Hard validation subset, our single model achieves state-of-the-art performance (92.0% AP) compared with both academic and commercial face detectors for detecting unconstrained faces in cluttered scenes. In the Wider Face AND PERSON CHALLENGE 2019, our ensemble model achieves 56.66% average AP (runner-up) in the face detection track. To facilitate further research on the topic, the training code and models have been provided publicly available.