A Local-Global Metric Learning Method for Facial Expression Animation
Pengcheng Gao, Bin Huang, Jiayi Lyu, Haifeng Ma, Jian Xue · 2022
Facial expression animation plays an important role in character animation. The Expression Blendshape Model (EBM) provides a simple representation of various expressions through a linear combination of base blendshapes with expression coefficients. However, it is challenging to distinguish subtle expression changes. In this paper, we propose a method that combines local features and global features to regress the expression coefficients. Furthermore, local metric leaning (LML) and global metric learning (GML) are proposed to enhance the recognizability of cross-individual expression features. Specifically, the LML increases the feature distance of each blendshape that appears or disappears from the perspective of local representation, resulting in better capture of local appearance changes, while the GML raises feature distance between neutral and emotional expression in the high dimensional feature space from the global perspective. Experimental results and feature visualizations on the FEAFA dataset show the effectiveness of local and global metric learning.