ygx2011/4D-Facial-Avatars

Dynamic Neural Radiance Fields for Monocular 4D Facial Avater Reconstruction

NerFACE: Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction [CVPR 2021 Oral Presentation]

Guy Gafni¹, Justus Thies¹, Michael Zollhöfer², Matthias Nießner¹

¹ Technichal University of Munich, ²Facebook Reality Labs

ArXiv: PDF, abs

Project Page & Video: https://gafniguy.github.io/4D-Facial-Avatars/

If you find our work useful, please include the following citation:

@InProceedings{Gafni_2021_CVPR,
    author    = {Gafni, Guy and Thies, Justus and Zollh{\"o}fer, Michael and Nie{\ss}ner, Matthias},
    title     = {Dynamic Neural Radiance Fields for Monocular 4D Facial Avatar Reconstruction},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {8649-8658}
}

Dataset and License

Dataset sample available.

Code Structure The nerf code is heavily based on this repo by Krishna Murthy. Thank you!

Installation etc: Originally the project used torch 1.7.1, but this should also run with torch 1.9.0 (cuda 11). If you get any errors related to torchsearchsorted, ignore this module and don't bother installing it, and comment out its imports. Its functionality is impmlemented in pytorch. These two are interchangeable:

    #inds = torchsearchsorted.searchsorted(cdf, u, side="right")  # needs compilationo of torchsearchsorted
    inds = torch.searchsorted(cdf.detach(), u, right=True)  # native to pytorch 

The main training and testing scripts are train_transformed_rays.py and eval_transformed_rays.py, respectively. The training script expects a path to a config file, e.g.:

python train_transformed_rays.py --config ./config/dave/dave_dvp_lcode_fixed_bg_512_paper_model.yml

The eval script will also take a path to a model checkpoint and a folder to save the rendered images:

python eval_transformed_rays.py --config ./config/dave/dave_dvp_lcode_fixed_bg_512_paper_model_teaser.yml --checkpoint /path/to/checkpoint/checkpoint400000.ckpt --savedir ./renders/dave_rendered_frames

The config file must refer to a dataset to use in dataset.basedir. An example dataset has been provided above (see sample).

If you have your own video sequence including per frame tracking, you can see how I create the json's for training in the real_to_nerf.py file (main function). This does not include the code for tracking, which unfortunately I cannot publish.

If you want access to our video sequences to run your methods on, don't hesitate to contact us [guy.gafni at tum.de]

The material in this repository is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.

Code for the webpage is borrowed from the ScanRefer project.