Hierarchical Training For Distributed Deep Learning Based On Multimedia Data Over Band-Limited Networks

This paper proposes a method for head pose estimation from a single image. We employ a multi-stage regression strategy. To overcome the discontinuity of Euler angles and quaternions and avoid the additional constraints required to directly regress the rotation matrix, we apply a continuous 6D representation to the head pose estimation problem. Each stage of the network regresses two 1?3 vectors, which are then transformed into a 3?3 rotation matrix by this continuous 6D representation. To better perceive the difference in rotation angles, we adopt the Riemann distance to measure the closeness between the network-estimated rotation matrix and the ground truth rotation matrix corresponding to the head pose. Experiments show that our method achieves the state-of-the-art on BIWI dataset and performs favorably on AFLW2000 dataset.