Towards Efficient Masked-Face Alignment Via Cascaded Regression

Practical and efficient face alignment has been highly required and widely focused in recent years, especially under the trend of edge computation and real-time operation. And it is a critical need to deal with masked faces in the context of COVID-19 epidemic. In this paper, we propose a novel cascaded facial landmark detector towards efficient masked face alignment, which we call QCN (Quantized Cascaded Network). QCN consists of three stages: alignment, estimation and refinement. The alignment stage help to pre-align the faces to alleviate extreme poses. And the next two stages localize facial landmarks in a coarse-to-fine manner. Thanks to the Network Architecture Search and Quantization techniques, the networks of QCN are designed as efficient as possible. Specifically, QCN occupies 1.75 Mb storage and runs in 84.18 MFLOPs only. Despite costs little computations, the proposed method yields 62.62% AUC (@0.08) on test set of JD-landmark-mask, which achieves 2nd place in the Grand Challenge of 106-point Facial Landmark Localization in ICME2021.