Clarification on Noiseprint++ training data and procedure

[jovanavidenovic]

Hi!

First, I'd like to thank you and congratulate on the amazing work you did.

I have some doubts regarding the Noiseprint++ training and would really appreciate your help. You mention that data was obtained from platforms DPReview and Flickr. However, both platforms contain huge amounts of data, so it would be helpful if you could provide information on which exact datasets/galleries were used.

I would also like to ask about Noiseprint++ training procedure itself. In the original Noiseprint paper, the batches of 200 patches were split into 50 internally homogeneous groups (which are heterogeneous with respect to each other). Does Noiseprint++ also works in a similar approach?

Thank you in advance!