search

avinabsaha / ReIQA

Official implementation for CVPR2023 Paper "Re-IQA : Unsupervised Learning for Image Quality Assessment in the Wild"
https://arxiv.org/abs/2304.00451
MIT License
87 stars 7 forks source link

Is it truly unsupervised or more of pretraining for quality estimation? #7

Closed kartheekmedathati closed 10 months ago

kartheekmedathati commented 10 months ago

The paper seems to use linear regression to map the pre-trained embeddings to the visual quality. I am wondering if it is right to call it unsupervised learning for image quality estimation?

avinabsaha commented 10 months ago

Hi, Thanks for your interest in our work!

We intend to refer to the representation learning framework as unsupervised. That is why our title says Unsupervised Learning for Image Quality Assessment.

Also, you could refer to this work for this amazing survey on types of self/unsupervised methods.

Thanks!

kartheekmedathati commented 10 months ago

Thanks for pointing the survey, not sure if that is relevant. "Unsupervised representation learning" instead of "unsupervised learning" could have made it more clear. When I read the title I was excited that some label free mechanism was figured out for quality estimation...

avinabsaha commented 10 months ago

@kartheekmedathati Our method can also be used label-free, but only in Full Reference Quality Estimation. We have not reported the results, but we have some internal results that show that in the FR/2AFC setting, just using the L1 distance between embeddings from the Re-IQA-Quality-Aware model can be useful for Quality estimation. Although the results may not be state-of-the-art across datasets, but are still usable. Also, using more image scales (>2) has improved performance in unsupervised settings!

avinabsaha commented 10 months ago

Closing this issue for now. Feel free to re-open to discuss further.