Gaussian Process Domain Experts for Model Adaptation in Facial Behavior Analysis

Stefanos Eleftheriadis, Ognjen Rudovic, Marc P. Deisenroth, Maja Pantic; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2016, pp. 18-26

Abstract


We present a novel approach for domain adaptation, that is based upon the framework of Gaussian processes (GPs). We introduce domain-specific GPs as local experts for facial expression classification. The adaptation is facilitated in probabilistic fashion by conditioning the target expert on multiple source experts. Contrary to the existing approaches, we also learn a target expert from available target data solely. Then, a single classifier is obtained by combining the predictions from multiple experts based on their confidence. Learning of the model is efficient and requires no retraining of the source classifiers. We evaluate the proposed approach on two datasets for multi-class (MultiPIE) and multi-label (DISFA) facial expression classification. In our experiments we perform adaptation of two contextual factors: 'where' (view) and 'who' (subject). We show that the proposed approach consistently outperforms generic classifiers and the state-of-the-art methods on domain adaptation.

Related Material


[pdf]
[bibtex]
@InProceedings{Eleftheriadis_2016_CVPR_Workshops,
author = {Eleftheriadis, Stefanos and Rudovic, Ognjen and Deisenroth, Marc P. and Pantic, Maja},
title = {Gaussian Process Domain Experts for Model Adaptation in Facial Behavior Analysis},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2016}
}