Unsupervised Domain Adaptation

phizaz / diffae

Official implementation of Diffusion Autoencoders

https://diff-ae.github.io/

MIT License

821 stars 123 forks source link

Unsupervised Domain Adaptation #37

Open rik030 opened 1 year ago

rik030 commented 1 year ago

Can this method be used for unsupervised domain adaptation?

phizaz commented 1 year ago

In what sense of domain adaptation on you which to apply DiffAE, mind providing more specific details? One thing I can say though. DiffAE might have some hope on unsupervised image translation problem, e.g. transforming male to female or a cat to a dog and etc. To what extent DiffAE is applicable is a bit underexplored in that regard.

rik030 commented 1 year ago

Yes, I was asking about the unsupervised image-to-image translation. To be more specific, translating ceT1 modality to hrT2 modality. These modalities are unpaired. So, can DiffAE be used for an unpaired image-to-image translation?

phizaz commented 1 year ago

They seem to be medical images, e.g. from https://www.researchgate.net/publication/359141212_Automated_Koos_Classification_of_Vestibular_Schwannoma/figures?lo=1.

We have even less clue on medical images. Do both modalities share mostly the same structure? Like cropped the same way? Are the differences between domains are spatially located on the same correspoding regions?

My feeling tells me that you have a good chance on this task. Good luck!

If you have any success, mind you keep we keep in touch? I'm curious how well could in work on medical image domains. You can mail me directly at konpatp (at) gmail.com

rik030 commented 1 year ago

Yes, sir, the two modalities share an almost identical structure. I am going through your paper and the methodology used in your code to check whether domain translation on medical images can be possible with diffae. Sure, I will be in touch with you :)

han-yeol commented 1 year ago

well, I am interested in same object

I try to T1w mri image (2nd column) to T2w mri image (1st column) image translation all model trained mixed T1w and T2w mri image . and given xT = T1w, cond = T2w slice (similar slice, because of using only T1w) but prediction(3rd column) is not good... any advice?

han-yeol commented 1 year ago

This is my best, honestly, This is cherry-picking...

phizaz commented 1 year ago

Without the domain (medical) knowledge, I am not sure what's the problem with the prediction you have?