DXOMARK-Research/PIQ2023

About the training stratage

chencn2020 opened this issue · 1 comments

Hi, I find that some jod-score is negative.

I'd like to know when you training a IQA model, like HyperIQA, do you do some process with jod scores, or do you just let the model predict negative jod.

Looking forward to your replay :)

Hello,

Yes since the scores are on a JOD scale (Just Objectionable Difference), they don't need to be strictly positive. Since the scores are open the prediction can be negative yes, we don't use any sigmoid here.

I hope I answered your question.

Nicolas