Figure 2.

Overview of our proposed DAN. The method is composed of three sub-networks. The backbone features are first extracted and clustered by a Feature Clustering Network (FCN), where an affinity loss is applied to increase the inter-class margin and to reduce the intra-class variance. Next, a Multi-head Attention Network (MAN) is built to attend to multiple facial regions concurrently by a series of Spatial Attention (SA) and Channel Attention (CA) units. Finally, an Attention Fusion Network (AFN) regulates the attention maps by enforcing variance among the attention feature vectors and outputs a class confidence.