VLAD-VSA ACM MM 2021

VLAD-VSA: Cross-Domain Face Presentation Attack Detection with Vocabulary Separation and Adaptation.pdf

• VLAD aggregation instead of global pooling.
• Vocabulary adaptation, a general trick to improve VLAD optimization.
• Vocabulary separation for domain generalization.

These codes are mainly based on the implementation of SSDG, specifically thank Yunpei Jia

Dependencies

python 3.6

pytorch 0.4

torchvision 0.2

Preparation

1. Download the OULU-NPU, CASIA-FASD, Idiap Replay-Attack, and MSU-MFSD datasets. Download the pretrained model and put it in the new-created pretrained_model package

2. Detect, crop and resize faces to 256x256x3 using MTCNN

   To be specific, we process every frame of each video and then utilize the sample_frames function in the utils/utils.py to sample frames during training.

   Put the processed frames in the path root/data/dataset_name.

3. Data Label Generation.

   Move to the root/data_label and generate the data label list:

   python generate_label.py
   

Training

• VLAD aggregation

python experiment/.../train_vlad_baseline.py

• VLAD + Vocabulary adaptation

python experiment/.../train_vlad_baseline2.py

• VLAD + Vocabulary adaptation + Vocabulary separation

python experiment/.../train_vlad_baseline3.py

The file config.py contains all the hype-parameters used during training. The parameters can be tuned and better performance may got.

Testing

python experiment/.../dg_test.py

Citation

Please cite this paper if the code is helpful to your research.

@inproceedings{wang2021vlad,
  title={VLAD-VSA: Cross-Domain Face Presentation Attack Detection with Vocabulary Separation and Adaptation},
  author={Wang, Jiong and Zhao, Zhou and Jin, Weike and Duan, Xinyu and Lei, Zhen and Huai, Baoxing and Wu, Yiling and He, Xiaofei},
  booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
  pages={1497--1506},
  year={2021}
}