Apply Central Difference Convolutional Network (CDCN) for Face Anti Spoofing

An adaptation from CDCN-Face-Anti-Spoofing.pytorch and face-anti-spoofing, add script to build custom data and fix bugs

Dependence

pip install -r requirements.txt

RTX3060 Laptop

Dataset

get original data from here

OR if you‘d like to train on custom data, use custom_data.py to add/create your custom dataset

before run custom_data.py, labeled images should be put in data/custom. Label format is like 0_xxxx.jpg

Train

tune params in config yaml file and run train.py

to see log visualization:

tensorboard --logdir=experiments/log --port=8008

Inference

single image inference by inference.py

Notes

difference from origin CDCN:

• The original version uses living face depth map estimated by PRNet and spoofing depth map is set to 0 In this version, spoofing depth map is also set to 0, but the living face depth map is set to 1 directly
• due to the difference about depth map, scores are calculated differently
• different training strategies

difference from other version:

• add custom data & param θ=0.7 as in origin paper；
• add pre-processing before inference
• save optimal weights instead of latest
• add api to load pretrained model, continue training checkpoints you saved in output dir
• add api for pre-generated depth maps. If you want to train with the depth map estimated by PRNet, put the depth map you generated from live face into data/train/depth and add a column "depth/xxxx" to the .csv corresponding to the depth map file. then set "depth_map_default" to 0 and train