Towards Occlusion Robust Facial Landmark Detector

Facial landmark detection plays an essential role in many face analysis tasks, such as face recognition, facial action unit detection, etc. However, when occlusion occurs, the existing methods might not perform well. This paper studies a new framework for facial landmark detection referred to as the Occlusion Robust Facial Landmark Detector (OR-FLD), which attempts to solve the occlusion problem. Specifically, the proposed OR-FLD consists of three main modules: Feature Random Dropping Module (FRD), Feature Activation Diversity Module (FAD), and Landmark Refinement Attention Module (LRA). The FRD forces the model to focus on the geometric relationships between different facial components by randomly dropping some appearance information and then automatically recovering them. The FAD tries to enhance the model representation and make the extracted features discriminative and robust to occlusions. Besides, the recovered features contribute differently to the final results. Therefore, we propose a simple and effective attention mechanism to refine the landmark location and make the model more robust to occlusions. We extensively evaluate OR-FLD on four facial landmark detection benchmarks (300W, WFLW, COFW, and AFLW). The results show that our method outperforms the state-of-the-art by a significant margin, especially on the datasets with large occlusion.