add fine-tuning examples for dense prediction tasks including regression and semantic segmentation tasks

[xiong-zhitong]

Integrate DOFA backbone with methods like ViT-CoMer for dense prediction tasks,
including regression and semantic segmentation tasks

[andrewaf1]

Hi! I found this paper and code really interesting and have begun using it via torchgeo.

I'm very curious how this method would work with ViT-CoMer. My understanding of that method is that it employs a ViT branch, a convolutional feature pyramid branch, and a module to fuse the resulting features. For DOFA, would a separate convolutional branch for each satellite source be required? I would also be very interested in seeing some examples of semantic segmentation, particularly how the features are transformed into a feature pyramid for UPerNet.

[xiong-zhitong]

Thanks for your interest. We have added example of using DOFA for the image classification task.
We are working on adding the semantic segmentation examples.
Please stay tuned!

[xiong-zhitong]

An example of using DOFA for segmentation tasks is added. We use DOFA as the backbone and UperNet as the segmentation head.