Zhou Huayi ; Jiang Fei ; Xiong Lili ; Lu Hongtao

Abstract

Most recent head pose estimation (HPE) methods are dominated by the Eulerangle representation. To avoid its inherent ambiguity problem of rotationlabels, alternative quaternion-based and vector-based representations areintroduced. However, they both are not visually intuitive, and often derivedfrom equivocal Euler angle labels. In this paper, we present a novelsingle-stage keypoint-based method via an {\it intuitive} and {\itunconstrained} 2D cube representation for joint head detection and poseestimation. The 2D cube is an orthogonal projection of the 3D regularhexahedron label roughly surrounding one head, and itself contains the headlocation. It can reflect the head orientation straightforwardly andunambiguously in any rotation angle. Unlike the general 6-DoF object poseestimation, our 2D cube ignores the 3-DoF of head size but retains the 3-DoF ofhead pose. Based on the prior of equal side length, we can effortlessly obtainthe closed-form solution of Euler angles from predicted 2D head cube instead ofapplying the error-prone PnP algorithm. In experiments, our proposed methodachieves comparable results with other representative methods on the publicAFLW2000 and BIWI datasets. Besides, a novel test on the CMU panoptic datasetshows that our method can be seamlessly adapted to the unconstrained full-viewHPE task without modification.