GeoAvatar: Adaptive Geometrical Gaussian Splatting for 3D Head Avatar

2025-07-24

AI Summary

Balancing identity fidelity and novel pose/expression animation remains challenging in 3D avatar generation, particularly due to heterogeneous geometric deviations across facial regions causing reconstruction artifacts. This paper proposes an adaptive geometric Gaussian splatting framework: first, unsupervised Gaussian segmentation coupled with adaptive pre-allocation decouples rigid (e.g., skull) and deformable (e.g., soft-tissue) Gaussian sets; second, a locally articulated deformation strategy is designed leveraging oral anatomical priors, while 3D Morphable Model (3DMM)-based binding regularization and adaptive offset regularization losses are introduced to constrain geometry and motion consistency. The method significantly improves reconstruction accuracy and animation naturalness, outperforming state-of-the-art approaches across multiple quantitative metrics. Additionally, we release DynamicFace—a high-fidelity, expressive facial video dataset featuring diverse identities, poses, and expressions—to support future research in dynamic 3D face modeling.

Technology Category

• Computer Vision: 3D Computer Vision
• Humans and AI: Game Design — Virtual Humans, NPCs and Autonomous Characters
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• User Modeling, Personalization and Recommendation: Accountability, Transparency, and Ethics for personalization
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

Despite recent progress in 3D head avatar generation, balancing identity preservation, i.e., reconstruction, with novel poses and expressions, i.e., animation, remains a challenge. Existing methods struggle to adapt Gaussians to varying geometrical deviations across facial regions, resulting in suboptimal quality. To address this, we propose GeoAvatar, a framework for adaptive geometrical Gaussian Splatting. GeoAvatar leverages Adaptive Pre-allocation Stage (APS), an unsupervised method that segments Gaussians into rigid and flexible sets for adaptive offset regularization. Then, based on mouth anatomy and dynamics, we introduce a novel mouth structure and the part-wise deformation strategy to enhance the animation fidelity of the mouth. Finally, we propose a regularization loss for precise rigging between Gaussians and 3DMM faces. Moreover, we release DynamicFace, a video dataset with highly expressive facial motions. Extensive experiments show the superiority of GeoAvatar compared to state-of-the-art methods in reconstruction and novel animation scenarios.

Problem

• Balancing identity preservation and novel animation in 3D head avatars
• Adapting Gaussians to varying facial region geometrical deviations
• Enhancing mouth animation fidelity through structure and deformation strategies

Innovation

• Adaptive Pre-allocation Stage for Gaussian segmentation
• Part-wise deformation strategy for mouth animation
• Regularization loss for precise Gaussian rigging