Graph-Driven Micro-Expression Rendering with Emotionally Diverse Expressions for Lifelike Digital Humans.

Area of Science:

• Computer Graphics and Animation
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Micro-expressions are vital for conveying nuanced emotions in digital humans.
• Existing rendering techniques struggle with temporal dynamics and action unit interdependencies, leading to rigid animations.
• There is a need for advanced methods to generate lifelike and emotionally diverse micro-expressions.

Purpose of the Study:

• To propose a novel graph-driven framework for micro-expression rendering.
• To enhance the emotional diversity and realism of digital human facial animations.
• To bridge the gap between micro-expression recognition and high-fidelity facial animation.

Main Methods:

• Utilized a 3D-ResNet-18 backbone for joint spatio-temporal feature extraction from facial videos.
• Modeled Action Units (AUs) as nodes in a symmetric graph, capturing dependencies with graph convolutional networks.
• Employed B-spline functions for interpolating AU activations into continuous motion curves for real-time animation (Unreal Engine).

Main Results:

• Achieved superior performance on the CASME II dataset with an F1-score of 77.93% and accuracy of 84.80% (5-fold cross-validation).
• Demonstrated improved temporal segmentation compared to existing baseline methods.
• Subjective evaluations confirmed enhanced perceptual clarity, naturalness, and realism in rendered digital humans.

Conclusions:

• The proposed graph-driven framework effectively renders emotionally diverse and lifelike micro-expressions.
• Modeling AU interdependencies using symmetric graphs significantly improves animation quality.
• This approach enables more expressive and realistic virtual interactions in digital human applications.