Which branch to use?

[celidos]

Hi all!

I want to try my hand at implementing WinCLIP. With version 1 of the library coming out soon, should I branch off from the v1 branch?

Btw, is it okay to take the unofficial implementation from caoyunkang here https://github.com/caoyunkang/WinClip as a basis (distributed under the MIT license)?

I'm new and haven't contributed to such projects before.

P.S. Thank you very much for your library, it is extremely handy ❤️

[blaz-r]

Hello. It's really great to hear you like the library. ❤️

I think that you should use the v1 as a base, since it's going to be released very soon.

Regarding the implementation, I think it's fine if you take that one as a basis, but make sure you credit the original author in readme 😄. Also read the contributing guidelines.

@samet-akcay could you please confirm all this?

[samet-akcay]

Hi @celidos, happy to hear this!

As @blaz-r suggested, you could use v1. However, not sure about WinCLIP implementation. @djdameln is already working on the winclip implementation, which is almost complete. To avoid working on the same thing, I would suggest implementing another model.

Maybe we could communicate better about upcoming features and models to avoid such duplication

[blaz-r]

Oh, I didn't know that. I agree that for model implementations there should be some way of communicating what is in the works. Maybe a draft PR once someone starts implementing.

I think that there are currently no diffusion based models in Anomalib like DiffusionAD, DDAD or TransFusion. They are top 3 in detection AUROC on VisA and very recent, so that would be a very valuable addition.

There was also M3DM at CVPR this year for 3D anomaly detection. Even better performing and very recent model for 3D anomaly detection is 3DSR, but there is not public code yet.

[celidos]

Okay!

I was curious about (maybe list?) of hot features that would be great to add to anomalib, but which no one is working on now (just to avoid duplication)

There is a great awesome-list of recent advances
https://github.com/M-3LAB/awesome-industrial-anomaly-detection

Maybe you have some ideas?

UPD: just noticed your response, thank you! I'll try diffusion models you've mentioned above

[blaz-r]

Cool. Addition of diffusion model would be really great, thanks. Both DiffusionAD and DDAD have available code that is MIT licensed. @samet-akcay just to confirm, is any of these models already in development, and do you have any preference which one would be better addition?