Active Learning for Accelerated Interactive Physical Facial Performance Design

Abstract

We present a practical neural computational approach for interactive design of animatronic facial performances. An offline quasi-static reference simulation accurately predicts hyperelastic skin deformations, driven by a coupled mechanical assembly. To achieve interactive digital pose design, we train a shallow fully connected neural network (KSNN) on input motor activations to solve the simulated mesh vertex positions. Our fully automatic synthetic training algorithm enables a first-of-its-kind Active Learning framework (GEN-LAL) for generative modeling of facial pose simulations. With adaptive selection we significantly reduce training time to within half that of the unmodified training approach for each new animatronic figure.