Lightweight Face Pose Estimation Multi-task Model Based On Rotation Matrix
Guanghui Ge, Youjia Fu · 2022
A lightweight multi-task sequence model is proposed to accommodate real-time computation on embedded devices. The model uses face pose estimation as the main task and face alignment as the secondary task and uses a 3×2 rotation matrix instead of the Euler angles to improve the model accuracy. The model uses a Sequence-sharing Multi-task Learning Model based on the MobileNet-V2 backbone network for structural de-branching, significantly reducing the number of model parameters. Experimental results show that, without loss of accuracy, the optimized model parametric is reduced to 0.5MB, it can run on mobile devices at 144 frames per sec, and its face alignment task has a 27.63% accuracy improvement compared with similar lightweight models.