Date: August 2020.
Source: Computer Vision – ECCV 2020, Lecture Notes in Computer Science, vol. 12374, Springer. DOI: 10.1007/978-3-030-58526-6_25.
Research Summary: Generating a realistic 3D face requires three separate pieces of information: its shape, its surface texture, and its surface normals, fine details like pores and wrinkles that determine how light reflects off skin. Earlier GAN-based methods generated these three separately, using disconnected processes that ignored how naturally correlated they are in a real face. This paper, known by its authors as TBGAN, introduces a trunk-branch generative adversarial network that generates all three together, learning their natural correlations rather than treating them as independent. This produces more realistic, photorealistic 3D faces than methods that generate each piece separately. An extended version of the model can also condition face generation on a specific facial expression.
3dMD’s Role: All training scans are registered to the LSFM template before use. The model’s core training data is approximately 10,000 neutral 3D facial scans from 3dMD’s MeIn3D dataset. Its expression-conditioned extension is trained on data from the 2017 “Are your facial expressions unique?” public exhibition at the Science Museum, London, approximately 5,000 distinct identities, the same event behind MimicME and TongueDB.
Article: Synthesizing Coupled 3D Face Modalities by Trunk-Branch Generative Adversarial Networks.
Authors: Baris Gecer, Alexander Lattas, Stylianos Ploumpis, Jiankang Deng, Athanasios Papaioannou, Stylianos Moschoglou, and Stefanos Zafeiriou. Imperial College London and FaceSoft.io.