If my feature extractor changes from arcface to adaface, will it not work as well, and will I need to retrain the model with adaface?

[xlovej]

If my feature extractor changes from arcface to adaface, will it not work as well, and will I need to retrain the model with adaface?

[nikky4D]

I have this question as well. How does the dependency on ArcFace affect performance

[foivospar]

Hi, the method works with the FR network in a black-box setting, using only its output embeddings (thus, any ID features could have been used for conditioning). However, since each network's embeddings reside in different spaces, the provided model is not interchangeable with, for example, AdaFace, as it was trained on ArcFace. Re-training would be required when switching to a different FR network.

[lambdaterm]

Another possible solution is to try converting AdaFace embeddings to the ArcFace latent space. We have achieved some successful results by converting buffalo_l embeddings to antelopev2 without having to retrain the Arc2Face models. https://github.com/lambdaterm/biometricshack-2024