Custom Dataset with Rotation

[DruncBread]

Hi,

Thanks for your great work!

I tested the pretrained RoMa on my images (not real world images; with large rotation angles), and the alignment results were not good. I have a few questions:

1. Rotation Augmentation
Is it possible that I fine-tune RoMa with rotation augmentation to make it support rotation alignment?

2. Custom Dataset
It would be a lot harder for me to build a dataset like MegaDepth, since there are only images and exact pixel correspondence between image pairs in my own datasets. ( Images are generated from one image with affine or other transformation).
Is these data sufficient to fine-tune RoMa? If you could provide a simple training demo for my own dataset (image pairs+pixel correspondence), it would mean a world to me!

Any advice would be appreciated!

Best regards.