Add MatchAnything Models

[aliejabbari]

Hi! Thanks for this great repo — it’s super clean and really helpful. I recently tested the MatchAnything models (Efficient LoFTR + ROMA), and they performed surprisingly well, especially on cross-modality image pairs. It would be great if this repo could also include these two matchers, since they fit nicely with the existing unified interface. The MatchAnything project provides ready pretrained weights for both Efficient LoFTR and ROMA and there’s also a helpful HuggingFace version here: https://huggingface.co/spaces/LittleFrog/MatchAnything/blob/main/imcui/third_party/MatchAnything/README.md. If someday you consider adding more models, these two could be a nice addition.

Thanks again for the awesome work !

[alexstoken]

Hi @aliejabbari,

Happy to hear you've found the repo helpful! Also nice to hear the MatchAnything models are useful for cross modal. We are tracking that in #13 and I hope to add it before the end of the year.

If you'd like, we're also accepting PRs to add models. If you already have a working implementation of MatchAnything and can fit it into our template, that would be a great contribution! We have a contribution guide as well.

[aliejabbari]

Hi @alexstoken,

Thanks a lot for the quick reply and for pointing me to the contribution guide. I’d be very happy to take a stab at adding MatchAnything.

In my experiments, I was able to load the MatchAnything weights for Efficient LoFTR and ROMA directly, and they worked quite smoothly on my cross-modality image pairs with minimal effort.

I’ll go through the template and contribution guide and see if I can put together a PR to add these models to the repo.

[aliejabbari]

I’ve opened PR #49 adding MatchAnything (ELoFTR + RoMa) via the HF Space submodule, with README instructions for fetching weights (gdown) and the ROMA editable install. Happy to tweak anything to fit.