davisking/dlib

[Bug]: can not recognition different people if they are face feature are similarity

xingyuezhiyun opened this issue · 12 comments

What Operating System(s) are you seeing this problem on?

Linux (x86-64)

dlib version

19.24

Python version

No response

Compiler

c11

Expected Behavior

No response

Current Behavior

can not recognition different people if they are face feature are similarity

Steps to Reproduce

as the example code

Anything else?

No response

Can you give more information please? A reproducible example with data is the minimum I'm afraid otherwise we can't fix anything.

We still need additional information.
How are you running the code? Are you detecting the faces, extracting the aligned chips to be 150×150 and then running the model, or doing something else?

Also, as far as I know, this model might underperform on non-Caucasian people.

We still need additional information. How are you running the code? Are you detecting the faces, extracting the aligned chips to be 150×150 and then running the model, or doing something else?

Also, as far as I know, this model might underperform on non-Caucasian people.

https://github.com/davisking/dlib/blob/master/examples/dnn_face_recognition_ex.cpp
this is the example what i used, they are Chinese, how can i do ?

May I ask if anyone can help me? @arrufat

May I ask if anyone can help me? @arrufat

I'm afraid there's not much I can do... maybe you can try jitter_image with several images and see if that makes it work. But I think it's kind of expected that the model underperforms with Asian people.

jilen commented

I'm afraid there's not much I can do... maybe you can try jitter_image with several images and see if that makes it work. But I think it's kind of expected that the model underperforms with Asian people.

I can confirm this model does not work very well with asian faces.

It says I (and another my colleague) and Daniel Wu are the same person.
Daniel Wu

@xingyuezhiyun You can try opencv dnn method, it performs a bitter better.

Yeah it's trained mostly on American celebrities so people who don't look like them don't work as well :|

There's no way. Let me try other libraries, or if there are Asian models available, please let me know. Thank you very much!

Maybe you could train the dlib model from scratch on http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html ?
This website shows other datasets too which might be less biased: https://datagen.tech/blog/face-datasets/
If you do end up with a better model, maybe consider releasing the weights.

Maybe you could train the dlib model from scratch on http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html ? This website shows other datasets too which might be less biased: https://datagen.tech/blog/face-datasets/ If you do end up with a better model, maybe consider releasing the weights.

I found hard to train a dlib model to used face dataset, I haven't found a complete training example

OK, Thank you ~