Comparing Head and AR Glasses Pose Estimation

Comparing Head and AR Glasses Pose Estimation
Ahmet Firintepe, Oussema Dhaouadi, Alain Pagani, Didier Stricker
IEEE Adjunct Proceeding of. IEEE International Symposium on Mixed and Augmented Reality (ISMAR-21) IEEE 2021 .

Abstract:

In this paper, we compare AR glasses and head pose estimation performance. We train different pose estimation approaches for head pose estimation with the generated head pose labels to compare them to their AR glasses estimation accuracy. These include the state-of-art GlassPoseRN and P2P networks, as well as our novel CapsPose algorithm. We show that estimating the AR glasses pose is more accurate than the head pose in general. In a first analysis, we show the general regression performance of the models when the AR glasses and faces are both known to the network during training. We then analyze the driver generalization performance, where all glasses are known, but part of the drivers are unknown to the Neural Networks. There, the estimation of AR glasses pose again exceeds the head pose. Only in our third analysis, head pose estimation performs better than AR glasses pose estimation. In this case, a new glasses model is added, which was unknown to the Neural Network yet. In addition, we introduce a novel pose estimation network called CapsPose, which is the first network deploying Capsule Networks for 6-DoF pose estimation. We outperform the current state-of-theart method GlassPoseRN on the HMDPose dataset by reducing the error by 46% for orientation and 51% for translation.