Facial Animation Retargeting by Unsupervised Learning of Graph Convolutional Networks

This paper proposes an unsupervised framework for retargeting human facial animations to different characters. Our method uses a branching structure of two parallel autoencoders and a variant of generative adversarial networks. The two autoencoder branches, composed of graph convolutional networks, share a common latent space through which the retargeting between different mesh structures can be performed. The shared latent codes are obtained by graph pooling operators, and the character face is reconstructed from the latent codes by the unpooling operators. The graph pooling and unpooling operators are designed based on multiple landmarks in optical-based facial motion capture systems. The GAN-based unsupervised learning method requires no paired training animation data between source and target characters. Our experimental results demonstrated that the proposed framework provides a reasonable estimation of a target facial expression that mimics a source character.