Learning Inverse Rig Mappings by Nonlinear Regression.

Area of Science:

• Computer Graphics
• Animation
• Machine Learning

Background:

• State-of-the-art computer animation often uses low-level pose representations (e.g., joint positions).
• Animators utilize high-level animation rigs for intuitive character control.
• A gap exists between these representations, hindering the adoption of advanced techniques in production.

Purpose of the Study:

• To develop a framework for inverse rig functions.
• To learn mappings from low-level pose data to animation rig controls.
• To facilitate the integration of advanced motion synthesis into animation pipelines.

Main Methods:

• Utilized nonlinear regression techniques to learn mappings.
• Trained models on example animation sequences provided by animators.
• Employed Gaussian process regression and feedforward neural networks for the mapping functions.

Main Results:

• Successfully demonstrated the framework on diverse character types (biped, quadruped, facial, deformable).
• The learned mapping enables estimation of rig controls from new motion data.
• The system allows animators to apply any motion synthesis algorithm to their rigs.

Conclusions:

• The framework effectively bridges low-level pose data and high-level animation rigs.
• This integration significantly improves 3D animation productivity.
• It preserves artistic flexibility while enabling the use of novel motion synthesis algorithms.