search

exadel-inc / CompreFace

Leading free and open-source face recognition system
https://exadel.com/accelerator-showcase/compreface/
Apache License 2.0
5.52k stars 754 forks source link

Question : How to use the SubCenter-ArcFace-r100 ? #931

Open Jean-Jerome-Casalonga-i2 opened 2 years ago

Jean-Jerome-Casalonga-i2 commented 2 years ago

Hi all, I'm not very happy with the results of the default installation of Compreface. I've got lots of false positives, with a very high score. I see that it's possible to use the SubCenter-ArcFace-r100 custom build, but I don't understand how to change my current install to use it.

Thanks a lot

neonleo commented 2 years ago

You have to reinstall via docker again. change to directory: C:\Users(you name)\Downloads\CompreFace_1.1.0\custom-builds\SubCenter-ArcFace-r100 and then perform the command "docker-compse up -d"

I have the same question, I try the default setup with docker. I found that the accuracy is very bad on my own image set, 1000 person, each person has 2 clear 800px x 800px frontal image for training, each person has about 10 real life photo capture by ip cam in 1080p quality. even though Similarity is 0.99 , but it matching wrong person. And now I am using Arcface build it is way better, I saw another thread said for Asian face, use ArcFace model is better.

Jean-Jerome-Casalonga-i2 commented 2 years ago

Thanks a lot. Yes, I do have lots of false positives on asian and african faces.

neonleo commented 2 years ago

Thanks a lot. Yes, I do have lots of false positives on asian and african faces.

ya, I am Asian too, and trying much other face recognition SDK fit for attendance system. you can also have a look SeetaFace6 that one of my pick list that resulting 98%+ correct rate.

Jean-Jerome-Casalonga-i2 commented 2 years ago

Thanks a lot for this information. First time I've heard of SeetaFace6. Will definitely take a look.

raunak-r commented 1 year ago

Hello all.. thought I'll also put my experience so far.

Used the base model first - but it gave a lot of false positives. Came across compreface because AWS Rekognition has a high cost.

Used Subcenter-arcface-r100 to remove false positives. But in our R&D sessions, we're seeing an average of 15-20 seconds per image processing time. Dev laptop spec is 16 GB RAM, i51235U, SSD disk

Can someone point me out to where I can get seetaface? Any maybe any other recommendation on compreface alternatives?

Jean-Jerome-Casalonga-i2 commented 1 year ago

https://github.com/seetafaceengine/SeetaFace6 But I couldn't manage to make it work. If you have more luck than me, I'd be happy to know what you do...

pospielov commented 1 year ago

You can try Mobilenet build instead of Subcenter-arcface-r100. It was trained on the same dataset, but it uses a smaller neural network, so it should run faster but with a little bit worse accuracy. For Subcenter-arcface-r100, I recommend using GPU to make it work faster.