MeshGAN: Non-linear 3D Morphable Models of Faces

• DOI: 10.48550/arxiv.1903.10384
• Publication Date: 2019-01-01
• Authors: Shiyang Cheng; Michael M. Bronstein; Yuxiang Zhou; Irene Kotsia; Maja Pantić; Stefanos Zafeiriou
• Keywords: High fidelity

Generative Adversarial Networks (GANs) are currently the method of choice for generating visual data. Certain GAN architectures and training methods have demonstrated exceptional performance in realistic synthetic images (in particular, human faces). However, 3D object, GANs still fall short success they had with images. One reasons is due to fact that so far been applied as convolutional discrete volumetric representations objects. In this paper, we propose first intrinsic architecture operating directly on meshes (named MeshGAN). Both quantitative qualitative results provided show MeshGAN can be used generate high-fidelity face rich identities expressions.