MetaAvatar

Learning Animatable Clothed Human Models from Few Depth Images

MetaAvatar is meta-learned model that represents generalizable and controllable neural signed distance fields (SDFs) for clothed humans. It can be fast fine-tuned to represent unseen subjects given as few as 8 monocular depth images.

Inspired by recent advances in meta-learned initializations for neural implicit representations, we propose to use meta-learning to effectively incorporate cloth-deformation priors for clothed humans, thus enabling fast fine-tuning (few minutes) for generating neural avatars given only a few monocular depth images of unseen clothed humans and their corresponding SMPL fittings as inputs.

teaser



Overview

We follow the same pipeline of recently proposed SCANimate, which learns dynamic neural SDFs from dense full-body scans to represent subject/cloth-type specific avatars of clothed humans. Instead of learning subject/cloth-type specific models from scratch, we propose to meta-learn a prior model which can be fast fine-tuned to represent any subject/cloth-type specific neural avatars given only monocular depth frames.

method overview



Our key contribution is a meta-learned hypernetwork that can be fine-tuned to represent subject/cloth-type specific neural SDFs given only a few monocular depth frames. In practice, we decompose the training into two stages 1) meta-learn a static neural SDF 2) meta-learn a hypernetwork which predicts residuals to the parameters of the previously meta-learned static neural SDF.

training stage1

training stage2



Results

Comparison to baselines



Learning with reduced data



Learning with depth from raw scans



Publication


Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, Siyu Tang
MetaAvatar: Learning Animatable Clothed Human Models from Few Depth Images
In Proceedings of Advances in Neural Information Processing Systems. 2021

@InProceedings{MetaAvatar:NeurIPS:2021,
  title = {MetaAvatar: Learning Animatable Clothed Human Models from Few Depth Images},
  author = {Shaofei Wang and Marko Mihajlovic and Qianli Ma and Andreas Geiger and Siyu Tang},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2021}
}