Barycentric Representation and Metric Learning for Facial Expression Recognition
Résumé
— In this paper, we tackle the problem of dynamic facial expression recognition. An affine-invariant facial shape representation based on barycentric coordinates is proposed and related to the Grassmannian representation. Unlike the latter, the barycentric representation allows us to work directly on Euclidean space and apply a metric learning algorithm to find a suitable metric that is discriminative enough to compare facial shapes under different expressions. Finally, we exploit the learned metric in a machinery combining a Dynamic Time Warping (DTW) phase and a pairwise proximity function SVM classifier for a rate-invariant classification of the facial sequences. Experiments on the AFEW dataset show the effectiveness of our approach while exploiting only geometric features.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...