Hempel Thorsten ; Abdelrahman Ahmed A. ; Al-Hamadi Ayoub

Abstract

Estimating the head pose of a person is a crucial problem for numerousapplications that is yet mainly addressed as a subtask of frontal poseprediction. We present a novel method for unconstrained end-to-end head poseestimation to tackle the challenging task of full range of orientation headpose prediction. We address the issue of ambiguous rotation labels byintroducing the rotation matrix formalism for our ground truth data and proposea continuous 6D rotation matrix representation for efficient and robust directregression. This allows to efficiently learn full rotation appearance and toovercome the limitations of the current state-of-the-art. Together with newaccumulated training data that provides full head pose rotation data and ageodesic loss approach for stable learning, we design an advanced model that isable to predict an extended range of head orientations. An extensive evaluationon public datasets demonstrates that our method significantly outperforms otherstate-of-the-art methods in an efficient and robust manner, while its advancedprediction range allows the expansion of the application area. We open-sourceour training and testing code along with our trained models:https://github.com/thohemp/6DRepNet360.