search

allenai / satlas-super-resolution

Apache License 2.0
190 stars 24 forks source link

Can't reproduce the results on new Sentinel-2 imagery (but on the test set, it works perfectly) #34

Closed ettelephonne closed 1 month ago

ettelephonne commented 1 month ago

First, thank you very much for this great work ! I used the 8 images model available on your github. It was easy to reproduce the results on the test dataset as shown here: image

After this try, I downloaded from the GEE 8 Sentinel 2 16 bits images of Montreal (8 different dates). I converted these images to 8 bits png. I cut these images in patches (8 32 32 * 3). You can see exemples here: image To me, these patches I made look exactly the same as those from the test sets: image But the result is not as expected : image image But I used the same model as for the test set... Do you have any clue to help me understand what did I miss ? Thank you very much for your time and your help.

piperwolters commented 1 month ago

Hi, thanks for your interest in our work!

To confirm, did you download Sentinel-2 L1C imagery, and use the TCI file provided by ESA? The range of these images should be 0-255 and be 8-bit. And did you reproject the imagery to Web-Mercator (pseudocode provided here).