Closed hihosilvers closed 10 months ago
Hi @hihosilvers, sorry for the late reply. This is still on our (quite long) ToDo list (see https://github.com/cvg/glue-factory/issues/22) but needs more tuning. We are welcoming any help here.
Thanks. I'll close this issue then
@Phil26AT Hi, do you have any suggestions on training the open SP+LG? Can I directly use the .yaml file in the repo?
Hi @noahzn
Yeah I'd recommend starting there. We obtained good results on homography pretraining with it, but it did not translate 1:1 to MegaDepth for us. You could try tuning the learning rate (which we never did with SP-open), or add some data augmentations for better generalization. If you manage to get good results, we'd welcome a contribution here. :)
Hi, wondering if it'd be possible to add the open version of SuperPoint (https://github.com/rpautrat/SuperPoint) to this LightGlue implementation like what you guys have in Glue Factory?
I'd like to train LightGlue myself on this version of SuperPoint but don't have the resources :(