Skip to content

Latest commit

 

History

History
64 lines (51 loc) · 4.93 KB

README.md

File metadata and controls

64 lines (51 loc) · 4.93 KB
Raw

I M Avatar: Implicit Morphable Head Avatars from Videos

Paper | Video Youtube | Video Download | Project Page

Official Repository for CVPR 2022 oral paper I M Avatar: Implicit Morphable Head Avatars from Videos.

Getting Started

  • Clone this repo: git clone --recursive [email protected]:zhengyuf/IMavatar.git
  • Create a conda environment conda env create -f environment.yml and activate conda activate IMavatar
  • We use libmise to extract 3D meshes, build libmise by running cd code; python setup.py install
  • Download FLAME model, choose FLAME 2020 and unzip it, copy 'generic_model.pkl' into ./code/flame/FLAME2020
  • When choosing your GPU, avoid RTX30xx since it seems unstable with broyden's method, see here if you want to know more. The results in the paper are obtained from a GeForce RTX2080Ti GPU. Quadro RTX6000 is also tested to converge well.

Preparing dataset

Download a preprocessed dataset from the ETH Zurich server (subject1 and subject2). You can run download_data.bash.

Or prepare your own dataset following intructions in ./preprocess/README.md.

Link the dataset folder to ./data/datasets. Link the experiment output folder to ./data/experiments.

Training

python scripts/exp_runner.py --conf ./confs/IMavatar_supervised.conf [--wandb_workspace IMavatar] [--is_continue]

Evaluation

Set the is_eval flag for evaluation, optionally set checkpoint (if not, the latest checkpoint will be used) and load_path

python scripts/exp_runner.py --conf ./confs/IMavatar_supervised.conf --is_eval [--checkpoint 60] [--load_path ...]

Pre-trained model

Download a pretrained model from the ETH Zurich server (subject1, subject2). See download_data.bash.

Additional features

The following features are not used in the main paper, but helpful for training.

  • Semantic-guided Training: set loss.gt_w_seg to True to use semantic segmentation during training. Using semantic maps leads to improved training stability, and better teeth reconstruction quality.
  • Ghost Bone: If FLAME global rotations in your dataset are not identity matrices, set deformer_network.ghostbone to True. This allow the shoulder and upper body to remain un-transformed.
  • Pose Optimization: When the FLAME parameters are noisy, I find it helpful to set optimize_camera to True. This optimizes both the FLAME pose parameters and the camera translation parameters. Similarly, set optimize_expression and optimize_latent_code to True to optimize input expression parameters and per-frame latent codes.

Warning

  • Our preprocessing script scales FLAME head meshes by 4 so that it would fit the unit sphere tighter. Remember to adjust camera positions accordingly if you are using your own preprocessing pipeline.
  • Multi-GPU training is not tested. We found a single GPU to be sufficient in terms of batch size.

You might find interesting

Acknowledgement

Yufeng Zheng and Xu Chen were supported by the Max Planck ETH Center for Learning Systems. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program grant agreement No 717054. This work was partly supported by the German Federal Ministry of Education and Research (BMBF): Tuebingen AI Center, FKZ: 01IS18039B. MJB has received research gift funds from Adobe, Intel, Nvidia, Meta/Facebook, and Amazon. MJB has financial interests in Amazon, Datagen Technologies, and Meshcapade GmbH. While MJB is a consultant for Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society.

Citation

If you find our code or paper useful, please cite as:

@inproceedings{zheng2022imavatar,
  title={{I} {M} {Avatar}: Implicit Morphable Head Avatars from Videos},
  author={Zheng, Yufeng and Abrevaya, Victoria Fernández and Bühler, Marcel C. and Chen, Xu and Black, Michael J. and Hilliges, Otmar},
  booktitle = {Computer Vision and Pattern Recognition (CVPR)},
  year = {2022}
}