Need help in reproducing the results #3
Description
Hi,
Thank you for this interesting work. We wanted to reproduce the results, but we are unable to match the results reported in the paper. We used iSAID as training dataset. First, we tried with cat_seg/cat_seg_model_ours.py and cat_seg/modeling/transformer/model_ours.py and got the following results:
ISAID val: 83.60
DLRSD: 22.58
Potsdam: 15.4
Vaihingen: 13.4
Next, we noticed SAU module was not present in cat_seg/modeling/transformer/model_ours.py, so we integrated necessary changes from cat_seg/modeling/transformer/model_attention.py file. This yields the following results:
ISAID val: 81.44
DLRSD: 23.33
Potsdam: 15.40
Vaihingen: 14.98
These models are trained for 100K iterations. Is there any other change that we need to do for reproducing the results?
Thanks in advance.