search

happynear / FaceVerification

An Experimental Implementation of Face Verification, 96.8% on LFW.
447 stars 234 forks source link

Is normalization of features necessary on Megaface evalution #76

Open YangBowenn opened 6 years ago

YangBowenn commented 6 years ago

I used sphereface to extract features and I get rank1: 49%, and I have read another issue similar to my question. But I am confusing that if I can directly normalize the features by (x-min)/(max-min) rather than use the normalize_layer for each image.

happynear commented 6 years ago

Normalization means L2 normalization. You may refer to my paper https://arxiv.org/abs/1704.06369 to understand why we should do L2 normalization.

And you may also understand why (x-min)/(max-min) would not work.