← ArchivePaper2005
Automatic Determination of Facial Muscle Activations from Sparse Motion Capture Marker Data
FEM-based facial muscle simulation with automatic activation estimation from sparse mocap markers, enabling physics-driven facial animation.
Abstract
We build an anatomically accurate model of facial musculature, passive tissue, and skeletal structure from volumetric data of a living subject, endowing the tissues with a nonlinear constitutive model and controllable anisotropic muscle activations based on fiber directions. To animate this model, we propose a method that automatically determines the muscle activations, head position, and jaw articulation that make a quasistatic finite element simulation track a sparse set of surface landmarks from motion capture marker data. The estimation is posed as a nonlinear least squares problem solved with a Gauss-Newton approach, where the Jacobian of the quasistatic configuration is computed efficiently. Because the search is performed over the space of physically attainable configurations parameterized by muscle activations, the method robustly handles outliers in the motion capture data and the resulting activations can be reused in dynamic simulations with contact and collision.
How to read this
- Category
- Method: anatomical FEM facial model with activation estimation from mocap
- Contributions
- Builds an anatomically accurate model of facial musculature, passive tissue, and skeleton from a subject's volumetric data, with a nonlinear constitutive model and anisotropic fiber-based muscle activations
- Automatically determines muscle activations, head position, and jaw articulation so a quasistatic FEM simulation tracks sparse mocap surface landmarks, solved as nonlinear least squares via Gauss-Newton
- Searches over physically attainable configurations, giving robustness to mocap outliers and activations reusable in dynamic contact/collision simulations
- Context
- Extends muscle-based facial animation (Waters' facial muscle model) and FEM musculoskeletal simulation (Teran et al.'s skeletal-muscle work) by inverting a physical face model from sparse marker data.Builds on: A Muscle Model for Animating Three-Dimensional Facial Expression · Creating and Simulating Skeletal Muscle from the Visible Human Data Set
- Correctness
- Accuracy depends on the per-subject anatomical model and the quasistatic assumption during tracking; estimating from sparse landmarks is inherently underdetermined, so the physical parameterization is what regularizes it, and results are tied to the captured individual.
- Clarity
- Dense, methods-heavy writing; a first pass conveys the build-model-then-invert-activations idea, with second and third passes needed for the constitutive model and the Gauss-Newton estimation.
- How to read it
- Focus on how the activation estimation is posed as least squares over physically attainable configurations and how the Jacobian is computed; multiple passes pay off for the FEM and solver if you work on physics-based faces.
Builds on
Related work
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →