Models and Evaluation codes



MODELS


To facilitate the research on fairness of face recognition, one can directly download the models and feature vectors of RFW datasets.

Models:

•  CASIA _Softmax.zip: The MxNet models trained with CAISA-Webface. The ResNet-34 architecture is used with the guidance of Softmax loss.

•  CASIA _Arcface.zip: The MxNet models trained with CAISA-Webface. The ResNet-34 architecture is used with the guidance of Arcface loss.

• Balanced_Softmax.zip: The MxNet models trained with BUPT-Balancedface. The ResNet-34 architecture is used with the guidance of Softmax loss.

•  Global_Softmax.zip: The MxNet models trained with BUPT-Globalface. The ResNet-34 architecture is used with the guidance of Softmax loss.

•  MS1M_Arcface.zip: The MxNet models trained with MS1M_wo_RFW. The ResNet-50 architecture is used with the guidance of Arcface loss.



Features:

•  CASIA _Softmax_feature.zip: 512-dimensional CASIA (Softmax) feature of each image.

•  CASIA _Arcface_feature.zip: 512-dimensional CASIA (Arcface) feature of each image.

•  Balanced_Softmax_feature.zip: 512-dimensional Balanced(Softmax) feature of each image.

•  Global_Softmax_feature.zip: 512-dimensional Global(Softmax) feature of each image.

•  MS1M_Arcface_feature.zip: 512-dimensional MS1M (Arcface) feature of each image.

In order to download the source images for further study, one must apply for ethnicity aware databases.




CODES


We also have uploaded the evaluation code on RFW, you can directly download them from here and use them to obtain the accuracy on 6000 pairs and the ROC curve on all pairs.


•  Feature_extract.zip. The code for extracting feature of your mxnet models.

•  Accuracy_for_RFW.zip. The evaluation code to get accuracy under LFW-like protocol. LFW-like protocol facilitates easy and fast comparison between algorithms with 6K pairs of images.

•  ROC_for_RFW.zip. The evaluation code to get ROC curve on all pairs of 3K identities (about 14K positive vs. 50M negative pairs).



PUBLICATIONS


Please cite the following if you make use of the dataset.

[1] Mei Wang, Weihong Deng, Jiani Hu, Xunqiang Tao, Yaohai Huang. Racial Faces in the Wild: Reducing Racial Bias by Information Maximization Adaptation Network. ICCV2019.

Bibtex  | PDF

[2] Mei Wang, Weihong Deng. Mitigating Bias in Face Recognition using Skewness-Aware Reinforcement Learning. CVPR2020.

Bibtex  | PDF

[3] Mei Wang, Weihong Deng. Deep face recognition: A Survey. arXiv:1804.06655.

Bibtex  | PDF

Contact Us

Please contact Mei Wang and Weihong Deng for questions about the database.