- I recommend mxnet version for classification, which has pretrained models on ImageNet.here a link
- created by : lxy and shj
- Time: 2018/12/10 15:09
- project Face Anti Spoofing
- company:
- rversion: 0.1
- tools: python 2.7
- modified:
- description: The codes for training and testing
- tensorflow >= 1.5.0
- python >= 2.7.15
- opencv >= 3.4.0
- caffe
- imgaug
- The training datas are downloaded from internet,using the tool BaiduDownload
- We have created the dataset including 4 classes (Mobilephone:1 TV:2 telectroller:3 background:0).
Configuration parameters lies in Root/src/configs/config.py
- directory
- data is used to store training and testing data.
- log is used to store traing logs.
- models is used to store network parameters.
- src is used to store training and testing codes.
- train
- get image list : running Root/src/prepare_data/run_script.sh to generate traing and testing data list.
- image augmentation: running Root/src/utils/transform.py for image augmentation, applying for images and images with boxes and images with keypoints
- pack training images: running Root/src/prepare_data/run.sh to pack training data.
- to train on packed images: running Root/src/train/run.sh
- test
- test one image: python Root/src/test/demo.py --img-path1 test.jpg --gpu 0 --load-epoch 10 --cmd-type imgtest
- test a video: python Root/src/test/demo.py --file-in test.mp4 --gpu 0 --load-epoch 10 --cmd-type videotest
- test on a test dataset: python demo.py --file-in ../prepare_data/output/test.txt --out-file ./output/record.txt --base-dir .../test_imgs/ --load-epoch 25 --cmd-type filetest
- video demo for face anti-spoofing
- run Root/src/face_test/run.sh
| class | TPR | FPR | Precision |
|---|---|---|---|
| Mobilephone | 0.631 | 0.026 | 0.856 |
| TV | 0.954 | 0.120 | 0.700 |
| Teleconrtoller | 0.827 | 0.013 | 0.929 |
| background | 0.809 | 0.106 | 0.812 |