This repository is reimplementation of "LGSC-for-FAS" repository, which is implementation of "Learning Generalized Spoof Cues for FaceAnti-spoofing (LGSC)"
Code of this repository uses "pytorch" and "pytorch_lightning"
There are 2 ways:
- Configure your own dataset, redefine
val_dataloader
&train_dataloader
functions inpl_model.py
- Specify in
configs/train_config.yml
next variables:train_df
- path to csv with train info. Each object should havetarget
(1 for live, 0 for spoof) andpath
(relative path, seesrc/dataset
to understand, how path is build), Also specifyval_df
(liketrain_df
),path_root
(absolute path to your data) anddefault_root_dir
(path to storepytorch_lightning
logs)
Note You can crop faces online with facenet-pytorch,
just specify parameters in configs/train_config.yml
Then run python src/train.py -c configs/train_config
- Specify parameters in
configs/infer_configs.yml
- Run
src/infer.py
@inproceedings{
feng2020deep,
title={Learning Generalized Spoof Cues for Face Anti-spoofing},
author={Haocheng Feng and Zhibin Hong and Haixiao Yue and Yang Chen and Keyao Wang and
Junyu Han and Jingtuo Liu and Errui Ding},
year={2020},
eprint={2005.03922},
archivePrefix={arXiv},
primaryClass={cs.CV}
}