Accurate Facial Landmark Detector via Multi-scale Transformer

Facial landmark detection is an essential prerequisite for many face applications, which has attracted much attention and made remarkable progress in recent years. However, some problems still need to be solved urgently, including improving the accuracy of facial landmark detectors in complex scenes, encoding long-range relationships between keypoints and facial components, and optimizing the robustness of methods in unconstrained environments. To address these problems, we propose a novel facial landmark detector via multi-scale transformer (MTLD), which contains three modules: Multi-scale Transformer, Joint Regression, and Structure Loss. The proposed Multi-scale Transformer focuses on capturing long-range information and cross-scale representations from multi-scale feature maps. The Joint Regression takes advantage of both coordinate and heatmap regression, which could boost the inference speed without sacrificing model accuracy. Furthermore, in order to explore the structural dependency between facial landmarks, we design the Structure Loss to fully utilize the geometric information in face images. We evaluate the proposed method through extensive experiments on four benchmark datasets. The results demonstrate that our method outperforms state-of-the-art approaches both in accuracy and efficiency.