Many false positives with a set of ~50 faces

[hawkeye217]

Describe the bug I have trained about 50 different faces with at least 4-5 images per face. When using the web interface to identify a person from an untrained image, I see high "similarity" numbers (> 98%) that are falsely identified as the wrong person.

I noticed this page - is there a setting I can change to get better results and less false positives?

Thanks in advance!

Expected behavior A much lower similarity number for these false positives.

Screenshots If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. iOS]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]

Additional context Add any other context about the problem here.

[pospielov]

Hi, false positives shouldn't depend on the number of faces. But we find out that FaceNet(default) model works badly on Asians. If this is your case - try to use our custom build SubCenter-ArcFace-r100: https://github.com/exadel-inc/CompreFace/tree/master/custom-builds The threshold you mentioned in our documentation - you need to choose it yourself on your side of the application, but the similarities are not configurable, they depend on the model. We will look at this problem more precisely, how we can reduce false positive

[hawkeye217]

Thank you for the suggestion. Unfortunately the SubCenter-ArcFace-r100 model performance is worse on my dataset. Similarity values are very low with both untrained and trained images.