The High Precision Real-Time Facial Landmark Detection Technique Based on ShufflenetV2.

Although in the past few decades, many methods such as heatmap, 3D morphable model (3DMM), and generative adversarial network (GAN), have been used to assist facial landmarks extraction, there is a lack of research on balancing the models' size and accuracy. Therefore, this paper proposes a landmark detection model based on the ShufflenetV2 module and the Wingloss function. The model achieves 86% accuracy on the WFLW extended testing set when its size is only 789.9KB (x0.5), and its speed can reach 60 FPS on a single Intel i5-8250U CPU. We also use pruning to further compress the model. After pruning, the size of the model is reduced by about 30%, the accuracy is reduced by 4%, and the speed is increased by 17%.