Small and accurate heatmap-based face alignment via distillation strategy and cascaded architecture

Despite face alignment has made significant progress, it is still challenging to get a small and accurate face landmark detection model. In this paper, we focus on compressing heatmap-based face alignment algorithms using a novel proposed distillation strategy. We find that the activated areas of heatmaps generated from the well-trained teacher net can capture more local shape information of the facial parts than the ground-truth ones generated from the standard Gaussian distribution. To simultaneously transfer such shape information and correct the mis-predicted heatmaps generated from the teacher model, we first modify the heatmaps of teacher model by replacing the mis-predicted heatmaps with the ground-truth ones as the labels for the student net. To further improve the accuracy of the student net, we investigate the correlation between the extracted features and the predicted heatmaps, and divide the landmarks into two categories: simple and hard. A cascaded architecture is designed which firstly detects the simple points based on the extracted features, and then predicts the hard points resorting to the heatmaps of simple ones. Finally, a face alignment model with 3.64M parameters is obtained, which is about 6x smaller than the cumbersome model, and outperforms the state-of-the-art algorithms on both AFLW2000-3D and 300W-LP. The model and code are released in https://github.com/snow-rgb/Fast-Face-Alignment.