adamhtoolwin / lgsc

LGSC-for-FAS pytorch

This repository is a reimplementation of a reimplementation (yes that's right) 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".

Train

There are 2 ways:

1. Configure your own dataset, redefine val_dataloader & train_dataloader functions in pl_model.py
2. Specify in configs/train_config.yml next variables: train_df - path to csv with train info. Each object should have target (1 for live, 0 for spoof) and path (relative path, see src/dataset to understand, how path is build), Also specify val_df (like train_df), path_root (absolute path to your data) and default_root_dir (path to store pytorch_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

Infer

1. Specify parameters in configs/infer_configs.yml
2. Run src/infer.py

Citation

@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}
}