PoLaNet: Head Pose Estimation and Face Landmarks Detection Network

In this paper, We proposed an improved efficient structure named PoLaNet for head pose estimation and landmarks detection from a single image while maintaining continuous representation. which transforms the Head Pose Estimation (HPE) problem into a problem of directly predict continuous 6D rotation matrix parameters belongs 3D Special Orthogonal Group (SO(3)), then use 3D Morphable Models (3DMM) to predict the landmarks and combined the two loss to enhance the accuracy. The model uses the light MobileNet-V2 as the backbone, and the improved combining loss function. Experiments show that compared with the original SOTA methods, our model parameters were reduced by about 83.3%, and the accuracy improved about 8%.