Hello, guys! I'm a novice in image quality assement. I found that the MOS in training set are very large, should i normalize the label or scaling it into [0, 1] to train my CNN model? If i tansform the target, how can i transform the predictions back to real MOS value to create the submission file? Need your help, thanks.
Posted by: wantsu @ Feb. 19, 2021, 3:54 a.m.Hi,
In theory, you don't need to scale your results. Scaling does not change the results of SRCC. And before calculating the PLCC, the scoring program will fit your results.
Thanks,
Posted by: JinjinGu @ March 2, 2021, 12:24 p.m.