About finetuning on RSICD and RSITMD

[zzl0107]

Dear author, thank you for your excellent work.
We have attempted to finetune on RSICD and RSITMD using the pretrained weights from RS5M (i.e., RS5M_ViT-B-32.pt). We have strictly followed the hyperparameter settings you provided in A question about finetuning on RSICD #12.
However, our results on RSICD and RSITMD are still about 2–3 percentage points lower in mR compared to the results reported in your paper.
Would it be possible for you to release the training code for this part? We would really appreciate it!

[zilunzhang]

Dear author, thank you for your excellent work. We have attempted to finetune on RSICD and RSITMD using the pretrained weights from RS5M (i.e., RS5M_ViT-B-32.pt). We have strictly followed the hyperparameter settings you provided in A question about finetuning on RSICD #12. However, our results on RSICD and RSITMD are still about 2–3 percentage points lower in mR compared to the results reported in your paper. Would it be possible for you to release the training code for this part? We would really appreciate it!

Hi zzl,

We use Delong Chen's ITRA framework to fine-tune the models. Please refer his codebase for ft: https://itra.readthedocs.io/en/latest/Contents/introduction/overview.html

[zzl0107]

Thanks a lot! I’ll try it.