Date: November 2017.
Source: ACM Transactions on Graphics (TOG), Volume 36, Issue 6, Article No. 194, Pages 1-17, https://doi.org/10.1145/3130800.3130813.
Research Summary: FLAME addresses the gap between high-end facial animation, which is realistic but labor-intensive, and low-end consumer depth-sensor models, which are fast but lack expressive range. FLAME (Faces Learned with an Articulated Model and Expressions) is a facial model built from thousands of accurately aligned 3D scans, designed to work with existing graphics software. It combines a linear shape space with an articulated jaw, neck, and eyeballs, pose-dependent corrective blendshapes, and global expression blendshapes, learned from real-world 3D/4D face scan sequences in the D3DFACS dataset along with additional 3D/4D scan sequences. In total, the model is trained from over 33,000 scans, and FLAME outperforms the FaceWarehouse and Basel Face Model on both static 3D scans and 4D sequences.
3dMD’s Role: FLAME’s facial pose and expression data were captured on dynamic-3D/4D 3dMD.u systems. Pose data, 10 subjects performing neck rotation and jaw motion, was captured on MPI’s own 60fps dynamic-3D/4D 3dMDtrio.u system, generating meshes of 45K vertices. Expression data came from two sources on the same system line: D3DFACS, a dynamic-3D/4D 3dMDface.u capture project at the University of Surrey, and additional sequences self-captured on MPI’s 3dMD system. Together, these sources produced over 69,000 registered frames, of which 21,000 were used for training. FLAME’s separate shape/identity space, 3,800 scans, draws on CAESAR, a 1998–2000 U.S. Air Force-led laser-scan anthropometric survey of roughly 4,400–5,000 civilian subjects, built for industrial sizing rather than facial or motion research.
Article: Learning a Model of Facial Shape and Expression from 4D Scans.
Authors: Tianye Li, Timo Bolkart, Michael J Black, Hao Li, Javier Romero, Max Planck Institute for Intelligent Systems, Tuebingen, Germany.