Improved Strategies for HPE Employing Learning-by-Synthesis Approaches
Abstract
The first contribution of this paper is the presentation of a synthetic video database where the groundtruth of 2D facial landmarks and 3D head poses is available to be used for training and evaluating Head Pose Estimation (HPE) methods. The database is publicly available and contains videos of users performing guided and natural movements. The second and main contribution is the submission of a hybrid method for HPE based on Pose from Ortography and Scaling by Iterations (POSIT). The 2D landmark detection is performed using Random Cascaded-Regression Copse (R-CR-C). For the training stage we use, state of the art labeled databases. Learning-by-synthesis approach has been also used to augment the size of the database employing the synthetic database. HPE accuracy is tested by using two literature 3D head models. The tracking method proposed has been compared with state of the art methods using Supervised Descent Regressors (SDR) in terms of accuracy, achieving an improvement of 60%.
Cite
Text
Larumbe et al. "Improved Strategies for HPE Employing Learning-by-Synthesis Approaches." IEEE/CVF International Conference on Computer Vision Workshops, 2017. doi:10.1109/ICCVW.2017.182Markdown
[Larumbe et al. "Improved Strategies for HPE Employing Learning-by-Synthesis Approaches." IEEE/CVF International Conference on Computer Vision Workshops, 2017.](https://mlanthology.org/iccvw/2017/larumbe2017iccvw-improved/) doi:10.1109/ICCVW.2017.182BibTeX
@inproceedings{larumbe2017iccvw-improved,
title = {{Improved Strategies for HPE Employing Learning-by-Synthesis Approaches}},
author = {Larumbe, Andoni and Ariz, Mikel and Bengoechea, Jose Javier and Segura, Ruben and Cabeza, Rafael and Villanueva, Arantxa},
booktitle = {IEEE/CVF International Conference on Computer Vision Workshops},
year = {2017},
pages = {1545-1554},
doi = {10.1109/ICCVW.2017.182},
url = {https://mlanthology.org/iccvw/2017/larumbe2017iccvw-improved/}
}