Fine-tuning on the clothing dataset did not yield good results

[PaPaQQQ]

We did not get good results in fine-tuning the semantic segmentation task on the clothing dataset. I want to ask if there is any problem with the training Settings. We used the 0.6b fine-tuning config and only changed the input size to 576*576. Compared with mask2former scheme of swin-L backbone, there is a big gap.

[PaPaQQQ]

I have no experience with fine-tuning vit models for downstream tasks and would like to ask if there are any tricks that could improve performance.