Awesome

FRT-PAD

This is an official pytorch implementation of 'Effective Presentation Attack Detection Driven by Face Related Task'. (Accepted by ECCV 2022)

Effective Presentation Attack Detection Driven by Face Related Task

• Paper Home Page

Requirements

• numpy>=1.17.0
• scipy>=1.5.2
• Pillow>=8.2.0
• pytorch>=1.7.1
• torchvision>=0.8.2
• tqdm>=4.59.0
• scikit-learn>= 0.24.2

Datasets

The proposed method is evaluated on four publicly-available datasets, i.e.

• MSU-MFSD
• OULU-NPU
• CASIA-FASD
• Idiap Replay-Attack

Usage

The proposed FRT-PAD method is trained through three steps:

• Data Preparation

  Generate the image list:

  python data_find.py \
  --data_path {Four paths of saved datasets}
  

  For example, python data_find.py --data_path ['msu_path', 'casia_path', 'idiap_path', 'oulu_path']

  And then you can get four lists containing images and corresponding labels in './label/' to establish cross-dataset.

• Pre-trained Model Preparation

  FRT-PAD method consists of CNN-based PA Detector, Face-Related Tasks and Cross-Modal Adapter. For CNN-based PA Detector (i.e. baseline), the pre-trained model is carried on ImageNet, and you can download the weights from resnet18. For Face-Related Tasks, we applied three different models.

In Face Recognition model, we use a Pre-trained ResNet-18, and you can download the weights from ms1mv3_arcface_r18_fp16/backbone.

In Face Expression Recognition model, we also use a pre-trained ResNet-18, and you can download the weights from SCN.

In Face Attribute Editing model, we only use its Discriminator, which can be downloaded from pretrained-celeba-128x128.

• Training and testing model

  python train_main.py \
  --train_data [om/ci]
  --test_data [ci/om]
  --downstream [FE/FR/FA]
  --graph_type [direct/dense]