Code for ChaLearn Face Anti-spoofing Attack Detection Challenge @ CVPR2019 by SeuTao

This is the source code for my solution to the ChaLearn Face Anti-spoofing Attack Detection Challenge hosted by ChaLearn.

Recent Update

2019.3.29: Final code is not ready, will update soon.

2019.3.10: code upload for the origanizers to reproduce.

Dependencies - TODO - for now use requirements.txt

• imgaug==0.2.6
• scikit-image==0.14.0
• scikit-learn==0.19.2
• tqdm==4.23.4
• torch==1.2.0
• torchvision
• mtcnn

Train single-modal Model

train model_A with color imgs， patch size 48：

CUDA_VISIBLE_DEVICES=0 python3 train_CyclicLR.py --model=model_A --image_mode=color --image_size=48 --dataset_path=<loaction of casia-surf-formatted dataset>

infer

CUDA_VISIBLE_DEVICES=0 python3 train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=color --image_size=48

Train multi-modal fusion model [TODO]

train model A fusion model with multi-modal imgs， patch size 48：

CUDA_VISIBLE_DEVICES=0 python train_Fusion_CyclicLR.py --model=model_A --image_size=48

infer

CUDA_VISIBLE_DEVICES=0 python train_Fusion_CyclicLR.py --mode=infer_test --model=model_A --image_size=48

For the origanizers to reproduce final two submissions [TODO]

unzip the models.zip in the root folder and infer all the trained models

infer

CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=color --image_size=32
CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=depth --image_size=32
CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=ir    --image_size=32

CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=color --image_size=48
CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=depth --image_size=48
CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=ir    --image_size=48

CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=color --image_size=64
CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=depth --image_size=64
CUDA_VISIBLE_DEVICES=0 python train_CyclicLR.py --mode=infer_test --model=model_A --image_mode=ir    --image_size=64

CUDA_VISIBLE_DEVICES=0 python train_Fusion_CyclicLR.py --mode=infer_test --model=baseline --image_size=32
CUDA_VISIBLE_DEVICES=0 python train_Fusion_CyclicLR.py --mode=infer_test --model=baseline --image_size=48
CUDA_VISIBLE_DEVICES=0 python train_Fusion_CyclicLR.py --mode=infer_test --model=baseline --image_size=64

run ensemble script submission.py to generate the final two submissions in phase2: (test_first.txt and test_second.txt)

python submission.py