Real-Time Face and Facial Landmark Joint Detection Based on End-to-End Deep Network

Facial landmark detection (FLD) is an important task in computer vision, involving the extraction of keypoints from facial images. Traditional methods typically employ a two-stage approach: first detecting faces, then predicting facial landmarks. However, computing deep features for accurate face and landmark detection is time intensive, and the features from each stage are not shared. This makes these methods suboptimal for real-time applications, especially on edge devices. In this article, we present a novel end-to-end deep network for joint face and FLD. Our approach builds upon the YOLO framework with minimal modifications, primarily involving the adjustment of multitarget labels for face detection and the addition of a separate head for landmark localization. Furthermore, we enhance the model using structural reparameterization, channel shuffling, and implicit modules. Experimental evaluations on the 300 W dataset demonstrate that our proposed method achieves high accuracy while maintaining real-time processing speeds, surpassing several state-of-the-art (SOTA) methods. Additional testing on challenging datasets such as Caltech Occluded Faces in the Wild (COFW) and AFLW2000-3D further highlights the robustness of our model in diverse conditions. Our model and source code will be made publicly available.