ibug.face_embedding

A collection of pre-trained face embedding models based on ArcFace [1] and RoI Tanh Polar transformer [2]. Both embedding predictor and training codes are included.

The embedding predictor supports the following backbones (iResNet is a variant of ResNet [3]):

• iResNet-18
• iResNet-50
• RTNet-50 [2] (pretrained model will be updated)

Supporting warping spaces:

• Cartesian (No projection)
• RoI Tanh Polar
• RoI Tanh
• RoI Tanh Circular

Prerequisites

• Tested with python 3.9.4, cuda 11.1.1, cudnn 8.1.0.77, pytorch 1.8.1, torchvision 0.9.1
• All other dependency can be installed using requirements.txt $pip install -r requirements.txt

Let $ROOT denotes this project's root directory where this README resides

How to Test

• ibug.face_detection is required for testing the embedding predictor
• Download the pretrained models, extract and place as following:

├── $ROOT
│   ├── ibug
│       └── face_embedding
│           └── weights
│               └── arcface_cartesian_iresnet18.pth
│               ...

• Run the testing script:

cd $ROOT
bash test.sh

How to Train

• Organize the training data as "ImageFolder" format that can be read by torchvision.datasets.ImageFolder. Example "ImageFolder" structure:

$data_root/Adam_Brody/xxx.png
$data_root/Adam_Brody/xxy.png
$data_root/Adam_Brody/[...]/xxz.png
...
$data_root/Brendan_Fraser/123.png
$data_root/Brendan_Fraser/nsdf3.png
$data_root/Brendan_Fraser/[...]/asd932_.png

$data_root stands for the dataset root directory. An example training dataset:

• Download the verification data and unzip it to a directory $ver_dir

• In ./train.sh, replace $data_root and $ver_dir with training and verification data paths, and also change $output_dir to the path of saving training data. Then run it:

bash train.sh

References

[1] Deng, Jiankang, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. "Arcface: Additive angular margin loss for deep face recognition." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4690-4699. 2019.

[2] Lin, Yiming, Jie Shen, Yujiang Wang, and Maja Pantic. "RoI Tanh-polar Transformer Network for Face Parsing in the Wild." arXiv preprint arXiv:2102.02717. 2021.

[3] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning for image recognition." In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778. 2016.