m-sunrise / core

CORE: COnsistent REpresentation Learning for Face Forgery Detection, CVPRW 22

CORE: COnsistent REpresentation Learning for Face Forgery Detection, CVPRW 22

paper | video

Dependencies

conda create -n CORE python=3.7
conda activate CORE
conda install pytorch torchvision cudatoolkit=11.3 -c pytorch
pip install -r requirements

Data pre-processing

Some errors that crop real faces in manipulated videos occur whether select the biggest or highest detection probability face by MTCNN, especially when there are two characters in a video. We solved the problem by using the provided video mask in FF++.

Please run python prep_w_mask.py --src-root Your_source_directory --dst-root Your_destination_directory --fake-type Deepfakes[Face2Face, FaceSwap, NeuralTextures, DeepFakeDetection, real]

Quickstart

You can reproduced the in-dataset results in Tab.1 using the following three commands:

• Xception: python main.py --dataset celebdf --aug-name None
• Xception+: python main.py --dataset celebdf --aug-name RE
• Ours: python main.py --dataset celebdf --aug-name RE --consistency cos --consistency-rate 1.0

For the cross-dataset results in Tab.8, three commands are as follows:

• None: python main.py --dataset ff --aug-name None
• RaAug: python main.py --dataset ff --aug-name RaAug --consistency cos --consistency-rate 100.0
• DFDC_selim: python main.py --dataset ff --aug-name DFDC_selim --consistency cos --consistency-rate 100.0

Thanks

1. We used RFM data augmentations, https://github.com/crywang/RFM
2. We used DFDC_selim data augmentations, https://github.com/selimsef/dfdc_deepfake_challenge
3. We modify xception model to output feature, https://github.com/tstandley/Xception-PyTorch

Citation

coming soon