The researchers propose a 6D rotation matrix representation, which helps learn the full rotation appearance.
Researchers from Otto von Guericke University proposed a method "for unconstrained end-to-end head pose estimation" with a continuous 6D rotation matrix representation for efficient direct regression.
As opposed to previous approaches that restrict the pose prediction to a narrow-angle, 6DRepNet can learn the full rotation appearance. The authors also proposed a geodesic distance-based loss to penalize the neural network with respect to the manifold geometry.
The researchers say their method outperforms other state-of-the-art methods by up to 20%.
Find out more about the model and get the repository here.