Face Brightness

This repo is mainly sharing the code of face-skin-brightness metric and brightness-information-metric in the Face Recognition Accuracy Across Demographics: Shining a Light Into the Problem. Also it includes the code for calculating the imposter and genuine distributions, face feature extraction, and plots.

Face-skin-brightness metric

This metric is used to measure the brightness level of an face image in terms of the upper face skin, which provides more accurate brightness measurement than the commercial SDK.

Boundaries: [94.92973987, 115.86682531, 198.77210088, 220.54641581]

Implementation

Make sure the face-parsing package is downloaded. Note that this package should be placed at the same directory level as your project, otherwise the path in degree_separation.py should be changed.

.
|---face-parsing
|---project

Run brightness analyzing code

python depree_separation.py

Brightness-information-metric

This metric is used to measure the brightness variance of the face skin in order to reflect the information on the face. $$BIM=\sum_{i=0}^{N}|B_i-\bar{B}|P(B_i)$$

Implementation

python brightness_information.py

Imposter and genuine distributions

We analyzed the effect of over-and-under exposed on the performance of the state-of-the-art model ArcFace (paper link: ArcFace: Additive Angular Margin Loss for Deep Face Recognition). The distributions of African-American male are as below:

More detailed analyses can be found in Face Recognition Accuracy Across Demographics: Shining a Light Into the Problem.

Citation

If you find any of the tools useful in your research, please consider to cite this paper:

@inproceedings{wu2023face,
  title={Face recognition accuracy across demographics: Shining a light into the problem},
  author={Wu, Haiyu and Albiero, V{\'\i}tor and Krishnapriya, KS and King, Michael C and Bowyer, Kevin W},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1041--1050},
  year={2023}
}