Suggested normalization for pretrained models

[adamjstewart]

Thanks for publishing these pretrained models! We added them to TorchGeo, but aren't sure what normalization we should be using with them. Looking through the datasets, I'm seeing very different normalizations used for each dataset:

• BigEarthNet: compute mean/std, normalize using clip((x - mean - 2 * std) / (4 * std) * 255)
• EuroSAT: no normalization
• OSCD: compute min/max, normalize using clip((x - min) / (max - min) * 255)
• SeCo: compute min/max, normalize using clip((x - min) / (max - min) * 255), followed by ImageNet normalization

To most effectively use these pretrained models, we would like to be able to use the same normalization used to create the weights. So I'm guessing we should use the SeCo normalization technique?

[symhsym]

Hi, may I ask where did you find the BigEarthNet normalizations formula?

[adamjstewart]

https://github.com/ServiceNow/seasonal-contrast/blob/main/datasets/bigearthnet_dataset.py#L111

[symhsym]

Thanks