Face Anti-Spoofer

Face anti-spoofing task solution using CASIA-SURF CeFA dataset, FeatherNets and Face Alignment in Full Pose Range.

• Requirements
• Setup
• Train
• Test
  • CASIA-SURF
  • Demo with Intel® RealSense™ camera
    • Running
    • Demo
• Submit

Model	Params, M	Computational complexity, MFLOPs	RGB	Depth	IR	Loss function	Optimal LR	Minimal ACER (CASIA-SURF val)	Snapshot
FeatherNet	0.35	79.99	✔️	❌	❌	Cross-entropy	3e-6	0.0242	Download
FeatherNet	0.35	79.99	✔️	✔️	❌	Cross-entropy	3e-6	0.0174	Download
FeatherNet	0.35	79.99	✔️	✔️	✔️	Cross-entropy	1e-7	0.0397	Download
FeatherNet	0.35	79.99	✔️	❌	❌	Focal loss	3e-6	0.0066	Download
MobileLiteNet	0.57	270.91	✔️	❌	❌	Cross-entropy	3e-7	0.1542	Download
MobileLiteNet	0.57	270.91	✔️	✔️	❌	Cross-entropy	3e-6	0.1019	Download
MobileLiteNet	0.57	270.91	✔️	✔️	✔️	Cross-entropy	3e-6	0	Download
MobileLiteNet	0.57	270.91	✔️	❌	❌	Focal loss	3e-7	0.1666	Download
ResNet18	13.95	883730	✔️	❌	❌	Cross-entropy	1e-3	0.0049	Download
ResNet18	13.95	883730	✔️	✔️	✔️	Cross-entropy	1e-3	0	Download
ResNet18	13.95	883730	✔️	❌	❌	Focal loss	1e-4	0.0021	Download
ResNet18 with dropout	13.95	883730	✔️	❌	❌	Cross-entropy	1e-3	0.0001	Download
ResNet18 with dropout	13.95	883730	✔️	✔️	✔️	Cross-entropy	1e-3	0	Download
ResNet18 with dropout	13.95	883730	✔️	❌	❌	Focal loss	1e-4	0.0034	Download

TensorBoard

Requirements

• Python 3.7.6
• PyTorch 1.4.0

Setup

1. Get CASIA-SURF dataset.
2. Move dataset folder to ./data/CASIA_SURF:

ln -s <your_path_to_CASIA> ./data/CASIA_SURF

3. Install requirements: pip install -r requirements.txt

Train

1. Tensorboard logs will be written to ./runs folder. To monitor them during training process, run:

tensorboard --logdir runs

2. Run training process:

python train.py --protocol PROTOCOL --config-path CONFIG_PATH --data_dir DATA_DIR
                [--epochs 10] [--checkpoint ''] [--train_batch_size 1]
                [--val_batch_size 1] [--eval_every 1] [--save_path checkpoints]
                [--num_classes NUM_CLASSES] [--save_every 1] [--num_workers 0]

Protocol must be either 1, 2 or 3. It determines CASIA-SURF benchmark sub-protocol of Protocol 4.

Test

CASIA-SURF

1. When you have the model, you can test it by running:

python test.py --protocol PROTOCOL --checkpoint CHECKPOINT --config-path CONFIG_PATH 
               [--data-dir DATA_DIR] [--num_classes NUM_CLASSES] [--batch_size BATCH_SIZE]
               [--visualize VISUALIZE] [--num_workers NUM_WORKERS] [--video_path VIDEO_PATH]

Protocol must be either 1, 2 or 3. It determines CASIA-SURF benchmark sub-protocol of Protocol 4.

Demo with Intel® RealSense™ camera

Running

python realsense_demo.py --video-path VIDEO_PATH --config-path CONFIG_PATH  [--num_classes NUM_CLASSES]

WARNING: Current evaluation for RealSense cameras was developed only for legacy devices which supported by pyrealsense library. Everything works fine for F200.

Demo

Submit

Submission is made for Face Anti-spoofing Detection Challenge at CVPR2020.

1. Run:

python submit.py --model1_path MODEL1_PATH --model2_path MODEL2_PATH --model3_path MODEL3_PATH 
                 [--num_classes 2] [--batch_size 1] [--output submission.txt]
                 [--num_workers 0]