You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Oct 2, 2024. It is now read-only.
TCI (this is three channels, RGB; divide by 255)
B05, B06, B07, B08, B11, B12 (divide by 8160, clip to 0-1)
I think b05-b12 have the extra clipping step because of more sever outliers.
The normalization step is like max norm per band, but instead of max for bands B05-B12, they use a different arbitrary value and then clip to the expected 0-1 range after division. While max norm would produce an equivalent range, the input distribution would be skewed.
So I think we need to represent this kind of normalization. Maybe we can make a new norm_type called "clip_norm" and specify "divide_value"= 8160 and "clip_range"=(0,1).
also, this type of normalization is only applied to some bands. so maybe we need to specify norm_type per band in a list rather than as a scalar for the entire model input.
The text was updated successfully, but these errors were encountered:
Right now, and as I see in the PR, this has been solved as norm_with_clip and then indicate the values by which to make the clip for each band in norm_with_clip_values. Looks good to me, as it allows to use different values for different bands and reading the description in the README it is well understood. Important is the part about each value is used to divide each band before clipping values between 0 and 1., otherwise could be confusing.
SATLAS, a project that produces pretrained models that generate featuremaps, finetuned task specific models, and a dataset, provides normalization steps that #2 doesn't capture: https://github.com/allenai/satlas/blob/main/Normalization.md#sentinel-2-images Different steps are applied for different bands.
I think b05-b12 have the extra clipping step because of more sever outliers.
The normalization step is like max norm per band, but instead of max for bands B05-B12, they use a different arbitrary value and then clip to the expected 0-1 range after division. While max norm would produce an equivalent range, the input distribution would be skewed.
So I think we need to represent this kind of normalization. Maybe we can make a new norm_type called "clip_norm" and specify "divide_value"= 8160 and "clip_range"=(0,1).
also, this type of normalization is only applied to some bands. so maybe we need to specify norm_type per band in a list rather than as a scalar for the entire model input.
The text was updated successfully, but these errors were encountered: