Data Augmented Network for Facial Landmark Detection

Abstract

Facial landmark detection objects to locate some predefined points on human face images. Compared with previous method, the accuracy of facial landmark detection has a great improvement. However, there are still many problems need to be solved in the field of facial landmark detection, such as the impact of environment (extreme pose, occlusion conditions and blur) on detection accuracy and the balance between model size with accuracy. In order to solve the impact of these problems, we introduce an approach for this task through data augmentation and cascade regression. By using AugNet, based on image disentangled representations and generative adversarial networks, we could map a human face image into content space and style space. Then rebuild a set of images that the locations of facial landmarks could be controlled. In this paper, we suppose that facial landmarks, edges and other semantic information in human face are charged by content space, whereas textures and colors mainly refer to style space. At the same time, we propose a multi-stage cascaded facial landmark detection, known as GLNet, to tradeoff between the model size and accuracy. GLNet localizes facial landmarks with coarse-to-fine search. GLNet would be optimized with augmented dataset that generated by AugNet. Besides, we conduct several experiments in benchmark datasets 300W, WFLW and AFLW. Extensive experimental results demonstrate that our approach outperforms competing methods. Specifically, our proposed facial landmark detector with tiny model size and high precision is suitable for real-time applications.