PSAvatar: A Point-Based Shape Model for Real-Time Head Avatar Animation With 3D Gaussian Splatting.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• 3D Computer Vision

Background:

• Current head avatar animation methods face challenges balancing speed and quality.
• 3D Morphable Models (3DMMs) struggle with non-facial structures (e.g., eyeglasses, hairstyles).
• Neural implicit models exhibit deformation inflexibility and rendering inefficiency.

Purpose of the Study:

• To introduce PSAvatar, a novel framework for creating animatable head avatars.
• To address the limitations of existing methods in modeling complex head geometry and achieving real-time animation.
• To leverage 3D Gaussian for high-fidelity rendering and fine detail representation.

Main Methods:

• Developed a Point-based Shape Model (PSM) using discrete geometric primitives (points) for enhanced representation flexibility.
• PSM samples points on and off the FLAME mesh to reconstruct complex geometries like hairstyles and eyeglasses.
• Integrated PSM with 3D Gaussian for fine detail representation and appearance modeling through analysis-by-synthesis alignment.

Main Results:

• PSAvatar successfully reconstructs high-fidelity head avatars from various subjects.
• The framework enables real-time animation of the created head avatars.
• Achieved improved modeling of non-facial structures compared to traditional 3DMMs.

Conclusions:

• PSAvatar offers a robust solution for high-fidelity, animatable head avatar creation.
• The combination of PSM and 3D Gaussian effectively handles head shape variations and complex details.
• The framework achieves real-time performance, advancing the state-of-the-art in digital human representation.