StruGauAvatar: Learning Structured 3D Gaussians for Animatable Avatars From Monocular Videos.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural 3D avatar reconstruction from monocular videos is challenging but offers broad applications.
• Neural Radiance Fields (NeRF) have advanced high-fidelity avatar generation.
• 3D Gaussian Splatting (3D-GS) offers efficient rendering but struggles with pose generalization due to lack of structure.

Purpose of the Study:

• To enhance the pose generalization capabilities of animatable 3D avatars reconstructed from monocular videos.
• To address the limitations of existing 3D-GS methods in handling unseen poses.
• To develop a novel representation that combines geometric structure with efficient rendering.

Main Methods:

• Proposed a hybrid representation integrating Discrete MObius THeorem (DMTet) for coarse avatar geometry with 3D Gaussian Splatting.
• Bound most Gaussian points to mesh vertices, allowing some free expansion for data fitting.
• Developed a dual-space optimization framework to jointly optimize DMTet, Gaussian points, and skinning weights.

Main Results:

• The proposed method significantly improves the generalization ability of avatars to unseen poses compared to existing 3D-GS approaches.
• Achieved high-fidelity rendering while maintaining efficient training and rendering.
• Demonstrated robust performance across extensive experimental evaluations.

Conclusions:

• Integrating structured geometry (DMTet) with 3D Gaussian Splatting enhances avatar pose generalization.
• The dual-space optimization framework effectively constrains Gaussian point deformation for improved performance.
• This approach offers a promising direction for creating more versatile and robust animatable 3D avatars.