The paper didn't say much about the feature transform module(FTM), so I'm not sure whether it should be a 19->19 channel 1*1 conv or 19->3 channel conv. In my experiment, 19->19 channel FTM along with 19 channel SISR feature calculating FA loss can achieve a mIoU of 0.6225, while 19->3 only achieves 0.5563. The normalization method in FA module also confuzes me, and I actually found that these normalizations make the result worse.(So the code here by default doesn't uses normalization. If you want to try using them please remove the comment of line 16,18,23,25 in utils/fa_loss.py)
The paper didn't say much about the feature transform module(FTM), so I'm not sure whether it should be a 19->19 channel 1*1 conv or 19->3 channel conv. In my experiment, 19->19 channel FTM along with 19 channel SISR feature calculating FA loss can achieve a mIoU of 0.6225, while 19->3 only achieves 0.5563. The normalization method in FA module also confuzes me, and I actually found that these normalizations make the result worse.(So the code here by default doesn't uses normalization. If you want to try using them please remove the comment of line 16,18,23,25 in utils/fa_loss.py)