Open CurryKd7 opened 2 months ago
Hi,
We have tried CAST-S and CAST-SD on cityscapes. Please check section A.3 in the paper.
My backbone uses deit_cast_stall_deep, and the pre trained model has a maximum of 2GB model checkpoint_0099.pth.tar. Why is it that after training on the Cityscapes dataset for 400 epochs, miou can only reach 58%?
May I ask which backbone you use to train cityscpes for semantic segmentation tasks