Core Concepts
GaussianAvatar proposes animatable 3D Gaussians for realistic human avatar modeling from a single video, optimizing motion and appearance jointly.
Abstract
The article introduces GaussianAvatar, an approach for creating realistic human avatars from a single video using animatable 3D Gaussians. It addresses challenges in accurate motion estimation and dynamic appearance modeling. The method involves explicit representation of humans with pose-dependent properties and joint optimization of motions and appearances. Experimental results demonstrate superior performance in appearance quality and rendering efficiency.
Directory:
Introduction
Creating customized human avatars from a single video is challenging due to underdetermined monocular observations.
Related Work
Neural rendering techniques for human reconstruction without predefined templates.
Method
Introduces animatable 3D Gaussians for dynamic appearance modeling.
Experiments
Evaluation on datasets like People-Snapshot, NeuMan, and DynVideo showcasing superior performance over baselines.
Conclusion and Discussion
Discusses limitations, potential social impact, acknowledgments, references.
Stats
GaussianAvatarは、単一のビデオからリアルな人間のアバターモデリングを提案します。
メソッドは、動きと外観を共同で最適化するためにanimatable 3Dガウスを使用します。