Skip to content

Latest commit

 

History

History
61 lines (42 loc) · 2.41 KB

README.md

File metadata and controls

61 lines (42 loc) · 2.41 KB
Raw

Huya Dopamine face analysis project

More Information Supervised Probabilistic Deep Face Embedding Learning(ICML2020)[paper]

Ying Huang*, Shangfeng Qiu*, Wenwei Zhang, Xianghui Luo and Jinzhuo Wang

We illustrate how the margin based loss methods work for face embedding learning in a perspective of probability and proposed two principles for designing reasonably new margin loss function . At the same time,we regard the open-set face recognition as a problem of information transmission and construct an auto-encoder architecture trained with a teacher student learning strategy named LASTE, which increased the generalization ability for the face embedding.

It actually boost the single model performance with open training dataset to more than 99%+ on MegaFace test easily. Please see our paper for more details.

Requirements

  • Python 3.6 is used. Basic requirements are listed in the 'requirements.txt'.
   pip install -r requirements.txt

Usage

Training Data Prepare

All face images are aligned by RetinaFace and cropped to 112x112. Then use data_processing/im2rec.py to pack face dataset in MXNet binary format to accelerate the training procedure.

Train:

  • place the training set in ~/huya_face/face_datasets
  • cd ~/huya_face/analysis/recognition
  • edit config.py
  • sh recognition_baseline.sh

Megaface Test:

  • prepare test dataset and official test tool in Megaface
  • cd ~/huya_face/analysis/recognition/Evaluation/Megaface
  • run run.sh

License and Citation

The usage of this software is released under the MIT License. There is no limitation for both acadmic and commercial usage.

@inproceedings{huang2020information,
  title     = {More Information Supervised Probabilistic Deep Face Embedding Learning},
  author    = {Ying Huang, Shangfeng Qiu, Wenwei Zhang, Xianghui Luo and Jinzhuo Wang},
  booktitle = {Proc. International Conference on Machine Learning (ICML)},
  year      = {2020},
}

Acknowledgement

  • The code structure and open training dataset are borrowed from InsightFace.

Contacts

Ying Huang: huanying@huya.com  

Shangfeng Qiu: qiushangfeng@huya.com