RAS: A Data-Driven Rigidity-Aware Skinning Model For 3D Facial Animation

We present a novel data-driven skinning model—rigidity-aware skinning (RAS) model, for simulating both active and passive 3D facial animation of different identities in real time. Our model builds upon a linear blend skinning (LBS) scheme, where the bone set and skinning weights are shared for diverse identities and learned from the data via a sparse and localized skinning decomposition algorithm. Our model characterizes the animated face into the active expression and the passive deformation: The former is represented by an LBS-based multi-linear model learned from the FaceWareHouse data set, and the latter is represented by a spatially varying as-rigid-as-possible deformation applied to the LBS-based multi-linear model, whose rigidity parameters are learned from the data by a novel rigidity estimation algorithm. Our RAS model is not only generic and expressive for faithfully modelling medium-scale facial deformation, but also compact and lightweight for generating vivid facial animation in real time. We validate the efficiency and effectiveness of our RAS model for real-time 3D facial animation and expression editing.