Cannot match reported SalsaNext performance with the released model

[chahyon-ku]

Hi authors!
I've been trying to reproduce the SalsaNext results you have on your paper using the provided pretrained model (https://drive.google.com/file/d/1DxuzlnFKnU8EpSuODRywJUrJlieUUheg/view?usp=sharing).


My confusion comes from the fact that I am getting identical IOUs for the latter 11 classes, but much lower IOUs for the first 3 classes (grass, tree, bush), reducing the total mIOU from 43.07 to 37.8.
I see that the table you have on your README also suffers from this: all other IOUs are identical to the ones on the paper except the first 3 classes, reducing the mIOU from 43.07 to 40.20.
Do you guys know why this might be?
Was the dataset/labels updated after you got your results?
Thanks for the help!