Authors:
Roland Stenger
1
;
Sebastian Löns
2
;
Feline Hamami
2
;
Nele Brügge
3
;
Tobias Bäumer
2
and
Sebastian Fudickar
1
Affiliations:
1
MOVE Junior Research Group, Institute of Medical Informatics, University of Lübeck, 23562 Lübeck, Germany
;
2
Institute for Systems Motor Science, University of Lübeck, 23562 Lübeck, Germany
;
3
German Research Center for Artificial Intelligence, 23562 Lübeck, Germany
Keyword(s):
Domain Randomization, Deep Learning, Dystonia, Head Pose Estimation, Synthesized Avatars.
Abstract:
We present an extended head pose estimation algorithm, which is trained exclusively on synthesized human avatars. Having five degrees of freedom to describe such head poses, this task can be regarded as being more complex than predicting the absolute rotation only with three degrees of freedom, which is commonly known as head pose estimation. Due to the lack of labeled data sets containing such complex head poses, we created a data set, consisting of renderings of avatars. With this extension, we take a step towards an algorithm that can make a qualitative assessment of cervical dystonia. Its symptomatic consists of an involuntary twisted head posture, which can be described by those five degrees of freedom. We trained an EfficientNetB2 and evaluated the results with the mean absolute error (MAE). Such estimation is possible, but the performance works differently well for the five degrees of freedom, with an MAE between 1.71° and 6.55°. By visually randomizing the domain of the avata
rs, the gap between real subject photos and the simulated ones might tend to be smaller and enables our algorithm being used on real photos in the future, while being trained on renderings only.
(More)