Continuity Rotation Representation for Head Pose Estimation without Keypoints
Xiong Zhao, Sarina Sulaiman, Liang Chen, Min Dong, Yunfeng Duo, Hao Song · 2023
In this paper, we present an improved end-to-end head pose estimation method in an unconstrained environment, which transforms the Head Pose Estimation(HPE) problem into a problem of directly predicting continuous 6D rotation matrix parameters belongs 3D Special Orthogonal Group(SO(3)). The method uses RepVGGplus-L2pse as the backbone, followed by one FC layer to output the results, model be trained on 300W-LP. The improved Root Mean Square Error of Geodesic Distance(RSME_GD) is used as the loss function to enhance the accuracy. The experiments on the two public face datasets AFLW-2000 and BIWI show that the results measured by Mean Absolute Error of Vectors (MAEV) are improved by 19.68% and 13.98% respectively compared with the original SOTA method.