Multi-view Face Detection and Landmark Localization Based on MTCNN

摘要

As the basic tasks of face application technology, face detection and facial landmark detection are two important research directions in the fields of computer vision. In this paper, we employ the multi-task cascaded convolutional networks (MTCNN) to realize the multi-view face detection and landmark localization in complex environments. Firstly, a MTCNN-based frontal face detector is trained for frontal face detection and landmark localization. The detector can achieve high accuracy on FDDB benchmark for face detection and AFLW benchmark for facial landmark detection. Secondly, we construct a non-frontal face dataset including 10026 images and train a non-frontal face detection model to solve the detection problem of missing large-angle faces and improve the detection accuracy of non-frontal face. Finally, the frontal face detector and the non-frontal face detector are combined for multi-task and multi-view face detection. The experimental results have shown the effectiveness of the proposed method.