Our FAN which integrates our reproduced one-shot RetinaNet and anchor-level attention significantly out-performs state-of-art detectors on the popular face detection benchmarks including WiderFace [34] and MAFA [6], especially in the occluded cases like MAFA.