YOLO-FD: YOLO for Face Detection

Face detection is a fundamental step for any face analysis approach. However, it remains as an unsolved problem in computer vision, specially, when it comes to the variability and distractions of in-the-wild environments. Moreover, a face detector must be accurate and fast to be used in surveillance/biometrics scenarios. In order to overcome these limitations, this paper proposes a customized version of the state-of-the-art object detector, YOLOv3, for face detection. The modifications aim at building a real-time, accurate model capable of detecting faces as small as 16 pixels in 34 FPS. Furthermore, this model was evaluated on three of the most difficult benchmarks for face detection, Wider Faces, UCCS and UFDD, showing a good score balance across them. Also, the comparison with the state-of-the-art shown that it was possible to achieve the second best FPS and the fifth best score on Wider Faces. Finally, the model will be available in https://github.com/luanps/yolofd.