search

deepinsight / insightface

State-of-the-art 2D and 3D Face Analysis Project
https://insightface.ai
23.45k stars 5.42k forks source link

Has anyone tried training Arcface with only open license databases? #2414

Open jayavanth opened 1 year ago

jayavanth commented 1 year ago

I want train a model that can be used for commercial purposes. So I was wondering if training it with open license databases like WebFace42M and Glint360K would result in a model with good accuracy. Has anyone tried it? What databases did you end up using?

mlourencoeb commented 1 year ago

Hello,

Those datasets are not open license, they are used for non-commercial purposes only. We trained with our internal datasets of around ~50M images and got very good accuracy (Top 5 In NIST). We did some changes to ARCFace loss and additional augmentation strategies. So, yes, it is possible to obtain good models with this repo.

jayavanth commented 1 year ago

Thanks for clarifying @mlourencoeb, I read that these databases were open-source in the training instructions and just assumed they are good for commercial use

Do you know of any commercial-use databases I can use to train Arcface to get decent accuracy? Doesn't have to be SOTA numbers but maybe a lot better than Facenet or some older models

marlowinnovations commented 1 year ago

Hello,

Those datasets are not open license, they are used for non-commercial purposes only. We trained with our internal datasets of around ~50M images and got very good accuracy (Top 5 In NIST). We did some changes to ARCFace loss and additional augmentation strategies. So, yes, it is possible to obtain good models with this repo.

Can you share what changes you made to ArcFace loss and the augmentation you used?

Hassan-miqdad commented 1 year ago

Hello,

Those datasets are not open license, they are used for non-commercial purposes only. We trained with our internal datasets of around ~50M images and got very good accuracy (Top 5 In NIST). We did some changes to ARCFace loss and additional augmentation strategies. So, yes, it is possible to obtain good models with this repo.

Would you mind if you can mention for how many identities the 50M images are?

BOT-at-office commented 2 days ago

hello can anyone help , i want to train a custom model for commercial purposes, how many models do i have to change/replace? 2d106 & 3d68 and inswapper_128 anything else?