ARHPE : Asymmetric Relation-aware Representation Learning for Head Pose Estimation in Industrial Human-Computer Interaction

Head pose estimation (HPE) has many wide industrial applications such as human-computer interaction, online education and automatic manufacturing. This study addresses two key problems in HPE based on the deep learning and attention mechanism. 1) How to bridge the gap between the better prediction performance of networks and incorrect labeled pose images in the HPE datasets 2) How to take full advantages of the adjacent poses information around the centered pose image To tackle the first problem, we reconstruct all the incorrected labels as a two-dimensional Lorentz distribution. Instead of directly adopting the angle values as hard labels, we assign part of the probability values (soft labels) to adjacent labels for learning the discriminative feature representations. To address the second problem, we reveal the asymmetric relation nature of HPE datasets, namely, the yaw direction and pitch direction are assigned different weights by introducing the ratio of half with at half-maximum of Lorentz distribution. Compared to traditional end-to-end frameworks, the proposed one can leverage the asymmetric relation cues for predicting the head poses angle in the incorrected label scenarios. Extensive experiments on two public datasets and our infrared dataset demonstrate that the proposed ARHPE network significantly outperforms other state-of-the-art approaches.