Home

Awesome

Official implementation for TALL and TALL++

[ICCV-2023] Thumbnail Layout for Deepfake Video Detection

[IJCV-2024] Learning Spatiotemporal Inconsistency via Thumbnail Layout for Face Deepfake Detection

Attention: The code for our improved IJCV extension (TALL++, https://arxiv.org/pdf/2403.10261.pdf ) will be made available in this repository.

Our implementation is based on Swin-Transformer.

Requirements

Data Preparation

Please refer to FaceForensic for how to prepare deepfake datasets such as FF++, Celeb-DF, and DFDC.

The data loader can load image sequences stored in txt files in the following format:

#example for train.txt
# path  |  start frame | end frame | label
original_faces_c23/928 1 300 0
original_faces_c23/712 1 300 0
original_faces_c23/582 1 300 0
original_faces_c23/602 1 300 0
deepfakes_faces_c23/143_140 1 300 1
deepfakes_faces_c23/408_424 1 300 1
deepfakes_faces_c23/766_761 1 300 1
deepfakes_faces_c23/964_174 1 300 1

Training:

[IMPORTANT] Edit main.py and change the default arg-parser values according to your convenience (especially the config paths)

CUDA_VISIBLE_DEVICES=0 python main.py --dataset ffpp \
 --input-size 112 --num_clips 8 --output_dir [your_output_dir] --opt adamw --lr 1.5e-5 --warmup-lr 1.5e-8 --min-lr 1.5e-7 \
 --epochs 60 --sched cosine --duration 4 --batch-size 4 --thumbnail_rows 2 --disable_scaleup --cutout True \
 --pretrained --warmup-epochs 10 --no-amp --model TALL_SWIN \
 --hpe_to_token 2>&1 | tee ./output/train_ffpp_`date +'%m_%d-%H_%M'`.log

Evaluation:

CUDA_VISIBLE_DEVICES=0 python test.py  --dataset ffpp \
 --input_size 112 --opt adamw --lr 1e-4 --epochs 30 --sched cosine --duration 4 --batch-size 4 --thumbnail_rows 2 --disable_scaleup \
 --pretrained --warmup-epochs 5 --no-amp --model TALL_SWIN  \
 --hpe_to_token --initial_checkpoint [model_checkpoint] --eval --num_crops 1 --num_clips 8 \
 2>&1 | tee ./output/test_ffpp_`date +'%m_%d-%H_%M'`.log

Citation

@inproceedings{xu2023tall,
  title={TALL: Thumbnail Layout for Deepfake Video Detection},
  author={Xu, Yuting and Liang, Jian and Jia, Gengyun and Yang, Ziming and Zhang, Yanhao and He, Ran},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={22658--22668},
  year={2023}
}

Contact