Head Pose Estimation Based on 5D Rotation Representation

Head pose estimation (HPE) is a crucial problem in computer vision, as it significantly enhances the performance of face-related tasks involving a frontal view. However, recent applications demand head analysis across the entire 360° range, which poses significant challenges. This paper introduces an end-to-end method for an HPE task. We propose a continuous 5D rotation representation to address the challenge of discontinuous rotation called 5DResNet, which enables robust and efficient direct regression of the head pose. The proposed method adopted the inverse of the stereographic projection (ISP) with the Gram-Schmidt mapping to orthogonalization procedure in the network. This approach allows our model to learn the full-range angles, exceeding the abilities of most previous techniques that confine pose estimation to a limited angle range to achieve acceptable results. Furthermore, we present an ablation study to gain a deeper understanding of the factors influencing the performance of our method. Our proposed approach demonstrates notable competition over other state-of-the-art methods in comprehensive experiments conducted on the publicly available Carnegie Mellon University (CMU) dataset, which achieved error rates of 5.97° and 6.64° for narrow and full-range angles, respectively.