Significance of rgb_mean

[terrainer]

Hi,

Thank you for your code! I have been very interested in one aspect in particular. Many papers' code implementations (including SwinIR, HAT, DAT, and more recently, SPAN) make use of the rgb_mean code that is commented out here.
I have noticed that doing the same with the SPAN code implementation code greatly increases training stability, but breaks compatibility due to its interaction with the network features.

I was wondering if you could explain why you've chosen not to include it in your implementation, and what pros and cons you know of when using it?

Kind regards,
terrainer

[XY-boy]

Hi,

Thank you for your code! I have been very interested in one aspect in particular. Many papers' code implementations (including SwinIR, HAT, DAT, and more recently, SPAN) make use of the rgb_mean code that is commented out here. I have noticed that doing the same with the SPAN code implementation code greatly increases training stability, but breaks compatibility due to its interaction with the network features.

I was wondering if you could explain why you've chosen not to include it in your implementation, and what pros and cons you know of when using it?

Kind regards, terrainer

Hi. As far as I know, this operation is used to boost the performance on specific training dataset, such as DIV2K. As I train TTST on remote sensing datasets, therefore I comment this operation.