So I downloaded and extracted glint360k and now I found that lfw.bin from glint360k is not the same as what I got from ICCV 2019 workshop . I then inspect the images in the bin files and found that images from glint360's lfw.bin are aligned pretty badly comparing to the one from ICCV 2019. Here are some of the images:
lfw.bin from glint360k:



lfw.bin from ICCV2019 workshop:



We will investigate this issue seriously, thanks for your feedback.
This is an interesting finding. Even if our align is wrong, we still use glint360k to get state of the art for LFW, lol.
The bin files of glint360k were copied from ms1m-v2(emore) which was aligned by mtcnn. But the challenge dataset uses retinaface.
@nttstar I see, thanks! btw, any reason behind this? should I train with mtcnn cropped and aligned lfw as validation dataset, or the one from the challenge is better (for validation)?
It's better to use the retinaface aligned version.
Most helpful comment
We will investigate this issue seriously, thanks for your feedback.
This is an interesting finding. Even if our align is wrong, we still use glint360k to get state of the art for LFW, lol.