Home

Awesome

Face Synthesis for Eyeglass-Robust Face Recognition

[ArXiv]

Intro

This repo releases the MeGlass dataset in original paper. MeGlass is an eyeglass dataset originaly designed for eyeglass face recognition evaluation. All the face images are selected and cleaned from MegaFace. Each identity has at least two face images with eyeglass and two face images without eyeglass. More details are presented in paper Face Synthesis for Eyeglass-Robust Face Recognition.

NameDataset typeLink
MeGlass_120x120.zipCroppedGoogle Drive or Baidu Yun, 335.8M
MeGlass_ori.zipOriginBaidu Yun, 13.3G

Dataset description

meta.txt contains the eyeglass labels of images. 1 means black-eyeglass, 0 means no-eyeglass.

MeGlass_120x120.zip consists of the cropped images of size 120x120.

MeGlass_ori.zip contains the original face images.

test directory contains four lists corresponding to the four protocols in paper.

DatasetIdentityImagesBlack-eyeglassNo-eyeglass
MeGlass1,71047,91714,83233,085
Testing set1,7106,8403,4203,420

Samples

<p align="center"> <img src="samples/samples.jpg", width="800px"> </p>

Dataset usages

To build this dataset, we use eyeglass classifier, powerful face recognition model and manual labor to keep right the person identity and black eyeglass attribute. Therefore, MeGlass dataset can be used for face recognition (identification and verification), eyeglass detection, removal, generation tasks and so on.

Identity parsing rule

Take one filename 10032527@N08_identity_4@2897031059_1.jpg for example, the string before the second @ makes one face image's identity. The naming rule is corresponding to the original MegaFace dataset.

Acknowledgement

The 3D face model fitting is based on Xiangyu Zhu's work.

Citation

If your research benefits from MeGlass, please cite it as

@article{guo2018face,
  title={Face Synthesis for Eyeglass-Robust Face Recognition},
  author={Guo, Jianzhu and Zhu, Xiangyu and Lei, Zhen and Li, Stan Z},
  journal={arXiv preprint arXiv:1806.01196},
  year={2018}
}