Decomposing a 3D face shape into different attribute components is usually beneficial to many applications, such as 3D face generation and attribute transfer. In this paper, we propose a novel method to learn independent latent representations of 3D face shapes to decompose a given 3D face shape into identity and expression components. We assume that the identity describes the intrinsic geometry of a face while the expression captures the extrinsic one, and thus they are independent of each other. Based on this assumption, we encode a 3D face shape into its identity and expression representations by a variational inference framework, that is equipped with Graph Convolutional Networks (GCN). Furthermore, we introduce a binary discriminator to enforce the latent representations of identity and expression to be distribution independent by adversarial learning. Both qualitative and quantitative experimental results show that the proposed approach can achieve state-of-the-art results in 3D face shape decomposition. Extensive applications on 3D facial expression transfer, 3D face recognition, and 3D face generation further demonstrate that the proposed method can achieve visually better transferred expressions, purer identity representations, and more diverse 3D face shapes, compared with existing state-of-the-art methods.

将 3D 人脸形状分解为不同的属性组件通常对许多应用程序都有好处，例如 3D 人脸生成和属性转移。在本文中，我们提出了一种新方法来学习 3D 人脸形状的独立潜在表示，以将给定的 3D 人脸形状分解为身份和表情组件。我们假设身份描述了人脸的内在几何形状，而表情捕捉了外在的几何形状，因此它们彼此独立。基于此假设，我们通过配备图卷积网络 (GCN) 的变分推理框架将 3D 人脸形状编码为其身份和表情表示。此外，我们引入了一个二元鉴别器来强制身份和表达的潜在表示通过对抗性学习独立于分布。定性和定量实验结果都表明，所提出的方法可以在 3D 人脸形状分解中达到最先进的结果。在 3D 面部表情转移、3D 面部识别和 3D 面部生成方面的广泛应用进一步表明，与现有的最新技术相比，所提出的方法可以实现视觉上更好的转移表情、更纯粹的身份表示和更多样化的 3D 面部形状方法。