Résumé: Dynamic facial expression recognition is important for human-computer interaction. Learning the temporal dynamic representation of facial expressions and exploring the complementary information between features of different views (e.g., anatomy-based views and neural network-based views) are two main challenges. In this paper, we first propose an adaptive graph attention (AGAT) network for temporal dynamic modeling. Specifically, a non-fully connected graph with a GLOBAL node is constructed to model local and global temporal dynamics simultaneously. The graph attention mechanism is adopted to learn an adaptive graph structure and dynamically attend to salient moments of facial expressions. For multi-view dynamic facial expression recognition, we propose a cross-view attention (CVA) module to mutually enhance the features of different views through channel attention at each moment. Based on the CVA module, we extend the AGAT network to a cross-view adaptive graph attention (CV-AGAT) network to simultaneously model the temporal dynamics of the mutually enhanced features for robust dynamic facial expression recognition. Extensive experiments on two dynamic facial expression recognition datasets demonstrate that the proposed AGAT and CV-AGAT networks achieve state-of-the-art results. The ablation study and visualization analysis further demonstrate the effectiveness of the proposed models.

Langue: Anglais