Using Imagenet mean/std for geobench classification

[raviolihui]

Hello, great work on DINO!

I am replicating you classification results of geobench. I see that you first transform the datasets to uint8 with max/min band stats of each dataset (eurosat, bigearthnet.. etz) and then it seems you normalize with Imagenet mean std.

I am wondering, why use Imagenet specifically and not the stats of the dataset you used for pretraining? I attempted using other statistics (for example bigearthnet when evaluating classification of bigearthnet) and it gives me worse results. Do you have an explanation of why using Imagenet stats (which seems counter intuitive) works better than any other stats?

Thank you for your time!

Best,
Carmen