Detection of Facial Directions and Features With YOLO-Based Deep-Learning Technology for Pre-Processing of Gaze Estimation

In this work, the pre-processing technology for gaze estimation is studied when the person is located at a distance of about three meters, and the proposed design applies the YOLO-based deep-learning model to detect two facial landmarks and six facial directions. By combining the appearance and geometric-features based schemes, the proposed design provides the pre-processing information for gaze estimation without the calibration process. By experiments, the input size of YOLOv3-tiny based model is set to 608x608 pixels. With the testing mode, the proposed method performs AP (average precision) to be 99% for detection of the six facial directions and two facial features. Compared with the previous design, the proposed method can detect more facial directions, and it also provides the simplified facial-landmark function for the further gaze estimation.