How to find the best similarity threshold on my private dataset?

[SuyueLiu]

Hi, thanks for your impressive work on face recognition.

I have trained an FR model on the MS1MV2 dataset with ArcFace and test on my private dataset, I noticed that in the verification part, you find the best accuracy by grid searching threshold value between 0 and 4 with 0.01 diff point. The threshold with the best accuracy on my private dataset is 1.3480 which is larger than 1.

The problem is when I use the cosine similarity (Between -1 to 1) to determine whether the two images are from the same person, how can I find the similarity threshold with the best accuracy or meet the TAR@FAR requirement?

[pancarnas]

Hi, this is the l2 norm not the cosine angle between 2 embedding

[nlpkiddo-2001]

Hi I have large corpus of sentence and keywords, After finding the cosine sim , How would I choose threshold to cosine sim to cluster the large corpus of text?

[pancarnas]

You can use the clustering algorithm on word/sentences embedding. But it is not related to this repo.