GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians
Models humans with animatable 3D Gaussians, learns pose-dependent appearance, and jointly refines motion and look for realistic monocular avatars.
Authors / Team
Liangxiao Hu · Researcher
Year
2024
Deep Dive
GaussianAvatar represents clothed humans with animatable 3D Gaussians anchored to a parametric body, using forward skinning to avoid inverse-skinning ambiguities common in NeRF avatars. A dynamic appearance network and an optimizable feature tensor capture pose-dependent detail such as wrinkles. Differentiable motion conditioning enables joint refinement of appearance and imperfect pose estimates from monocular video. Experiments report favorable quality and efficiency against prior avatar methods.
What we learn
- 01
Explicit Gaussians with differentiable skinning let you co-optimize motion and appearance errors in monocular setups.
- 02
Pose-conditioned appearance needs extra degrees of freedom to avoid overfitting on limited poses.
Verbatim quote
"We present GaussianAvatar, an efficient approach to creating realistic human avatars with dynamic 3D appearances from a single video."— source ↗
Tags
Links
Sources