Building Accurate Physics-based Face Models from Data

How to read this

Category

Method: data-driven physics-based face modeling

Contributions

• Builds a volumetric face model from MRI of a neutral face and registers 3D scans captured under varying gravity directions and expressions
• Solves an inverse physics problem to learn heterogeneous stiffness and prestrain from the training scans
• Demonstrates a model that generalizes to new 3D scans and predicts deformations more accurately than prior physics-based techniques

Context

Extends physics-based face modeling such as Phace (Ichim et al., 2017) and the authors' prior personalized anatomical body models (Kadlecek and Kavan, 2016), shifting from hand-set to data-learned material parameters.Builds on: Phace: Physics-based Face Modeling and Animation · Reconstructing Personalized Anatomical Models for Physics-based Body Animation

Correctness

Relies on MRI-derived volumetric geometry and scans under controlled gravity and expression conditions; the accuracy claim is relative to prior physics-based methods, and generalization is shown on new scans of presumably the same modeling regime, so per-subject capture cost and coverage are limits to keep in mind.

Clarity

Technically dense; a first pass conveys the inverse-physics idea, but the stiffness and prestrain estimation needs a careful second pass.

How to read it

Read for the inverse-physics formulation: on a second pass focus on how stiffness and prestrain are recovered from gravity-varied scans, and on the validation against prior physics-based baselines.