training set CASIA-WebFace: did you use clean_list or all images from CASIA? about sphereface

[wy1iu]

We use a clean list of CASIA, which removed all the identities appearing in LFW. Using all images will violate the training protocol for LFW and it is prohibited.

[qiqiguaitm]

I have done quit lots of experiments but not reproduced your results;
could you open your clean list if possible?
or just give some statistics of your data, i just want to figure out how different between yours and mine, and may this is the reason.

[qiqiguaitm]

thx

[wy1iu]

The repo already contains the code to produce our training list.

[qiqiguaitm]

ok,thx

[Gerkam]

How many photos and classes must be in washed casia web face?

[yxchng]

@wy1iu @qiqiguaitm So the casia webface used in this repo is different from https://groups.google.com/forum/#!topic/cmu-openface/Xue_D4_mxDQ?

[yxchng]

Based on this, https://github.com/happynear/FaceDatasets/tree/master/CASIA, it seems like there are identities that you have not excluded from your train set. Does that mean that your accuracy on LFW is not reliable?

[happynear]

@yxchng
As I experimented, it doesn't affect much on LFW because there are more than 10,000 identities in LFW. 17 is an ignorable number compared with 10,000.

But for MegaFace FaceScrub set1, there are 42 overlap identities between it and WebFace. Note that there are only 80 identities in FaceScrub set1, which means over half of the identities are already trained. The performance drops significantly on MegaFace if the overlap identities are removed.

[yxchng]

@happynear When you say doesn't affect much, do you mean I can reproduce the result of this paper with identities from https://github.com/happynear/FaceDatasets/blob/master/CASIA/webface_lfw_overlap.txt removed? Currently, I am unable to. Are you?