Code repository for the paper:
ReFit: Recurrent Fitting Network for 3D Human Recovery
Yufu Wang, Kostas Daniilidis
ICCV 2023
[Project Page]
2024-01-05: Added an example for multiview refit.
2023-12-20: Added support for GART with a preprocessing script.
python3 -m venv refit
source refit/bin/activate
pip install -U pip
pip install -r requirements.txt
There are some few key data you need for the demo, training and evaluation.
- SMPL: Please download SMPL_NEUTRAL/FEMALE/MALE.pkl and put them under
data/smpl - ReFit weights: download each checkpoint folders and put them under
data/pretrain - Yolov7: for detection, download this Yolov7 checkpoint and put it under
data/pretrain. Be sure to also clone the yolov7 submodule withgit clone --recurse-submodules this-repo.
Additionally for training and evaluation, please follow the dataset preparation guide here.
We provide a few examples in data/examples. Running the following demo will detect each person in each image and then run ReFit for 3D reconstruction. The resulting rendering will be saved.
python demo.py
For multiview refit, run the following demo
python scripts/multiview_demo.py
To run on a folder of images (eg. frames from a video), and save the SMPL and camera pose for GART, run the following.
python scripts/process_folder.py
Training: config.yaml, config_bedlam.yaml and config_all.yaml control the dataset composition.
python train.py --cfg configs/config.yaml
Evaluation:
python scripts/eval.py
We benefit greatly the following repos, from which we adapted parts of our code.
If you find the model and code useful, please consider citing the following paper:
@Inproceedings{wang23refit,
Title = {ReFit: Recurrent Fitting Network for 3D Human Recovery},
Author = {Yufu Wang and Kostas Daniilidis},
Booktitle = {International Conference on Computer Vision},
Year = {2023}
}