VIS-VAR/LGSC-for-FAS
VIS-VAR/LGSC-for-FAS has 232 GitHub stars and 58 forks. Star changes are not measured yet. Last checked 2026-11-15 19:07 UTC. Last push 2020-06-18.
측정된 스타 증가
- in 24h
- 기록 수집 중
- in 7 days
- 기록 수집 중
- in 15 days
- 기록 수집 중
- in 1 month
- 기록 수집 중
서로 겹치는 기간의 순증가량이며 일별 합계가 아닙니다. 측정값이 없으면 비워 둡니다. 스타는 관심도를 나타내며 품질을 보장하지 않습니다.
소개
Learning Generalized Spoof Cues for FaceAnti-spoofing
학습README.md
OriginalLGSC-for-FAS
This repository contains code of "Learning Generalized Spoof Cues for FaceAnti-spoofing (LGSC)", which reformulate face anti-spoofing (FAS) as an anomaly detection problem and learn the discriminative spoof cues via a residual-learning framework.
Framework

FaceForensics Benchmark
Based on the LGSC, we achieved 1st Place on FaceForensics benchmark

Dataset and Preprocessing
We use two different compression rates: c23(medium compression) and c40(high compression) videos in FaceForensics++ dataset as our training data. As a preprocessing step we extract crop faces frame from raw videos and totally we get 0.96 million face frames, which contrains 0.16 million pristine frames and 0.8 million manipulated frames. Further, we balanced the proportion of positive and negative samples during training.
Technology Detail
Our deep network is implemented on the platform of PaddlePaddle(dygraph).
- Download and preproces the dataset
FaceForensics/data
├── manipulated_sequences
│ └── DeepFakeDetection
│ └── Deepfakes
│ └── Face2Face
│ └── FaceSwap
│ └── NeuralTextures
├── original_sequences
│ └── actors
│ └── youtube
│ └── c23
│ └── c40
│ └── images
├── train_add_train.txt
├── train_val_train.txt
└── faceforensics_benchmark.txt
- Download and convert imagenet pretrained model
python convert.py
./pretrained
├── resnet18-5c106cde.pth
└── resnet18-torch.pdparams
- Train
python train.py
- Test
python test.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}
}