GaussianHead: High-Fidelity Head Avatars With Learnable Gaussian Derivation

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Modeling

Background:

• Generating lifelike 3D head avatars and animations for diverse subjects is a significant challenge in computer vision.
• Existing methods often struggle with capturing intricate head dynamics and detailed textures efficiently.
• The need for compact yet high-fidelity representations is crucial for practical applications.

Purpose of the Study:

• To introduce GaussianHead, a novel method for modeling active heads using anisotropic 3D Gaussians.
• To enable the creation of lifelike 3D head avatars and compelling animations.
• To achieve high-fidelity visual results with a compact model size and superior performance compared to state-of-the-art methods.

Main Methods:

• Modeling the active head using anisotropic 3D Gaussians.
• Integrating a motion deformation field and a single-resolution tri-plane for capturing dynamics and texture.
• Introducing a customized derivation scheme for precise position transformation and an inherited derivation strategy for expedited training.

Main Results:

• Achieved high-fidelity visual results with a remarkably compact model size (approximately 12 MB).
• Demonstrated superior performance over state-of-the-art alternatives in reconstruction, cross-identity reenactment, and novel view synthesis.
• Successfully generated multiple 'doppelgangers' through learnable parameters for precise position transformation.

Conclusions:

• GaussianHead effectively models active heads, enabling the creation of lifelike 3D avatars and animations.
• The method offers a significant advancement in terms of visual fidelity, model compactness, and performance.
• GaussianHead represents a promising direction for future research in 3D head modeling and animation.