STDN-PyTorch

This is an unofficial implementation of "On Disentangling Spoof Traces for Generic Face Anti-Spoofing" in PyTorch

[Paper] [Official Implementation in TensorFlow]

Requirements

• python 3.6+
• PyTorch 1.6.0
• easydict
• TensorFlow 2+

Data Preparation

• Create keypoints for both Live and Spoof faces using any of these repository:
  • https://github.com/cunjian/pytorch_face_landmark
  • https://github.com/cunjian/pytorch_face_landmark
• Create a csv with Column Names being 'rgb_path', 'keypoints', 'label'
• For labels, Spoof is 1 and Real is 0
• Equally Sample the dataset so that there is equal number of spoof and equal number of real.

Configurations

To change your configurations, open config.py and for setting up data paths, change in flags.data_config. You can change other parameters too based on your need.

How to train?

Once data preparation is done, and data path is created, run the following code:

python train.py

All your checkpoints will be saved under the directory ckpts. However, you can change the path from config.py. the file names will be in the following format:

model_epoch-number_val-accuracy_val-apcer_val-bpcer.pth

How to test?

Still in progress.

Help taken from

1. On Disentangling Spoof Traces for Generic Face Anti-Spoofing [1].
2. [Official Implementation in TensorFlow(https://github.com/yaojieliu/ECCV20-STDN)]

Citations

[1] Yaojie Liu, Joel Stehouwer, Xiaoming Liu.