CVPR2020-SingleModal

Code for competition : Chalearn Single-modal Cross-ethnicity Face anti-spoofing Recognition Challenge@CVPR2020

we use RGB data for training.

commit

The date of last commmit is 03/01/2020 ,our result is comes from the last version.

Prerequisites

We use Anaconda3 with python 3.7 , we use :

opencv-python 4.2.0

pytorch 1.4.0

imutils 0.5.3

numpy 1.18.1

in the enviroment.

Data pre-progress

Edit cut_face.py ,set your data path.

global_action = "dev" #"train" or "dev"
data_root = "/Users/wdh/Downloads/CASIA-CeFA/%s"%global_action

Change global_action = "train" ,to cut negative data in train

python cut_face.py

Change global_action = "dev" ,to cut negative data in dev

python cut_face.py

Final data index tree

like this

├── data
    ├── dev
        ├── 003001
            ├── depth
            ├── depth_bak
            ├── ir
            ├── ir_bak
            ├── profile
            ├── profile_bak	
 ... ...

Train

Set your datapath and checkpoint path and coda ，in train_gray_rgb_4@1.py train_gray_rgb_4@2.py train_gray_rgb_4@3.py

./train.sh

Test

Set your global_models、model_names 、global_actions and data_root valid_gray.py. We recommend the last epoch(step) model witch usually perform better.

python valid_gray.py

you will get six txt file :

4\@1_dev_gray_res.txt 4\@1_test_gray_res.txt 4\@2_dev_gray_res.txt 4\@2_test_gray_res.txt 4\@3_dev_gray_res.txt 4\@3_test_gray_res.txt

​ then ：

./merge.sh

finally get the submission.txt