Johan Edstedt
Johan Edstedt
I've published wheels from my fork without issues. I dont currently have any plan to merge.
Fixed in #146
The notebook is public since a while back, you can see the approach there.https://www.kaggle.com/code/johanedstedt/tiny-roma-imc2022?scriptVersionId=130698232 
If you are able to generate output for 3 pairs you just change it to not dryrun and submit it no?
> > If you are able to generate output for 3 pairs you just change it to not dryrun and submit it no? > > Thanks, I am runing it....
Hi there! Unfortunately we extensively use the features themselves (not just the correlation), so if you swap the backbone you get big issues.
You could try mapping the new backbone to DINOv2 on some reasonable in distribution images, for example attach a linear head at the end of it. Then roma should be...
Hi. If you look inside our loss, we basically convert the depth into corresps. For your finetuning you would have to edit the loss to not use depth but rather...
It's broadly consistent but not exactly the same. It only stored thr paird (not overlap matrix) for example.
I would avoid matching when the roma confidence is very low. In the provided images I would guess that is the case.