Hi,
I have two questions:
1) It's not entirely clear to me how to use the FFHQ dataset, as the images are not grouped by identity, so I don't see a way they can be used for training. Could you give some pointers on that?
2) The standard approach when doing face verification is to compute the distance of the two faces' embeddings and compare it using a threshold. As I suppose it's not allowed to use ground truths from the provided test sets, are external datasets such as LFW allowed in order to compute such threshold?
Thanks!
Posted by: nicolo.didomenico @ Oct. 11, 2023, 8:56 a.m.Dear Nicolò,
Thanks for reaching out, here the answers to your questions:
1) You are correct about FFHQ, as the images are not grouped by identity. For training, we consider CASIA-WebFace and FFHQ and you are free to combine these datasets in any manner you prefer. Even if not designed for face verification, FFHQ can be considered in the proposed challenge for many aspects such as training a model for feature extraction and apply domain adaptation, among many other possibilities.
2) Only databases indicated for training in the specific sub-task can be used to compute the threshold of your face recognition system.