FeatherNets for Face Anti-spoofing Attack Detection Challenge@CVPR2019[1]
The detail in our paper：FeatherNets: Convolutional Neural Networks as Light as Feather for Face Anti-spoofing
In the first phase,we only use depth data for training ,and after ensemble ACER reduce to 0.0. But in the test phase, when we only use depth data, the best ACER is 0.0016.This result is not very satisfactory. If the security is not very high, just using single-mode data is a very good choice. In order to achieve better results, we use IR data to jointly predict the final result.
| model name | ACER | TPR@FPR=10E-2 | TPR@FPR=10E-3 | FP | FN | epoch | params | FLOPs |
|---|---|---|---|---|---|---|---|---|
| FishNet150 | 0.00144 | 0.999668 | 0.998330 | 19 | 0 | 27 | 24.96M | 6452.72M |
| FishNet150 | 0.00181 | 1.0 | 0.9996 | 24 | 0 | 52 | 24.96M | 6452.72M |
| FishNet150 | 0.00496 | 0.998664 | 0.990648 | 48 | 8 | 16 | 24.96M | 6452.72M |
| MobileNet v2 | 0.00228 | 0.9996 | 0.9993 | 28 | 1 | 5 | 2.23M | 306.17M |
| MobileNet v2 | 0.00387 | 0.999433 | 0.997662 | 49 | 1 | 6 | 2.23M | 306.17M |
| MobileNet v2 | 0.00402 | 0.9996 | 0.992623 | 51 | 1 | 7 | 2.23M | 306.17M |
| MobileLiteNet54 | 0.00242 | 1.0 | 0.99846 | 32 | 0 | 41 | 0.57M | 270.91M |
| MobileLiteNet54-se | 0.00242 | 1.0 | 0.996994 | 32 | 0 | 69 | 0.57M | 270.91M |
| FeatherNetA | 0.00261 | 1.00 | 0.961590 | 19 | 7 | 51 | 0.35M | 79.99M |
| FeatherNetB | 0.00168 | 1.0 | 0.997662 | 20 | 1 | 48 | 0.35M | 83.05M |
| Ensembled all | 0.0000 | 1.0 | 1.0 | 0 | 0 | - | - | - |
2019.4.4: updata data/fileList.py
2019.3.10:code upload for the origanizers to reproduce.
2019.4.23:add our paper FeatherNets
conda env create -n env_name -f env.yml
├── data
│ ├── our_realsense
│ ├── Training
│ ├── Val
│ ├── Testing
Download and unzip our private Dataset into the ./data directory. Then run data/fileList.py to prepare the file list.
| Method | Settings |
|---|---|
| Random Flip | True |
| Random Crop | 8% ~ 100% |
| Aspect Ratio | 3/4 ~ 4/3 |
| Random PCA Lighting | 0.1 |
download fishnet150 pretrained model from FishNet150 repo(Model trained without tricks )
download mobilenetv2 pretrained model from MobileNet V2 repo,or download from here,link: https://pan.baidu.com/s/11Hz50zlMyp3gtR9Bhws-Dg password: gi46 move them to ./checkpoints/pre-trainedModels/
nohup python main.py --config="cfgs/fishnet150-32.yaml" --b 32 --lr 0.01 --every-decay 30 --fl-gamma 2 >> fishnet150-train.log &
nohup python main.py --config="cfgs/mobilenetv2.yaml" --b 32 --lr 0.01 --every-decay 40 --fl-gamma 2 >> mobilenetv2-bs32-train.log &
Commands to train the model:
python main.py --config="cfgs/MobileLiteNet54-32.yaml" --every-decay 60 -b 32 --lr 0.01 --fl-gamma 3 >>FNet54-bs32-train.log
python main.py --config="cfgs/MobileLiteNet54-se-64.yaml" --b 64 --lr 0.01 --every-decay 60 --fl-gamma 3 >> FNet54-se-bs64-train.log
python main.py --config="cfgs/FeatherNetA-32.yaml" --b 32 --lr 0.01 --every-decay 60 --fl-gamma 3 >> MobileLiteNetA-bs32-train.log
python main.py --config="cfgs/FeatherNetB-32.yaml" --b 32 --lr 0.01 --every-decay 60 --fl-gamma 3 >> MobileLiteNetB-bs32--train.log
example:
python main.py --config="cfgs/mobilenetv2.yaml" --resume ./checkpoints/mobilenetv2_bs32/_4_best.pth.tar --val True --val-save True
run EnsembledCode_val.ipynb
run EnsembledCode_test.ipynb
notice:Choose a few models with large differences in prediction results
You can download my artifacts folder which I used to generate my final submissions: Available Soon
[1] ChaLearn Face Anti-spoofing Attack Detection Challenge@CVPR2019,link
[2] Shifeng Zhang, Xiaobo Wang, Ajian Liu, Chenxu Zhao, Jun Wan, Sergio Escalera, Hailin Shi, Zezheng Wang, Stan Z. Li, " CASIA-SURF: A Dataset and Benchmark for Large-scale Multi-modal Face Anti-spoofing ", arXiv, 2018 PDF