search

Vandermode / TFPnP

Tuning-free Plug-and-Play Proximal Algorithm for Inverse Imaging Problems (ICML 2020 Award Paper & JMLR 2022)
https://arxiv.org/abs/2002.09611
87 stars 18 forks source link

Can TFPnP deal with unlabeled data? #2

Closed lixyxd closed 1 year ago

lixyxd commented 2 years ago

RL need reword which is PSNR. Thus, how can we deal with the data which do not have label?

Vandermode commented 2 years ago

Yes, you could define the reward function tailored to your need. For unsupervised cases, perhaps you could use non-reference image quality metrics e.g., NIQE, as a reward function (See [1] for details). Other relevant non-reference IQA such as [2] are also possible to be exploited

[1] Learning the Non-differentiable Optimization for Blind Super-Resolution, CVPR 2021 [2] RankIQA: Learning from Rankings for No-reference Image Quality Assessment, ICCV 2017

lixyxd commented 2 years ago

Yes, you could define the reward function tailored to your need. For unsupervised cases, perhaps you could use non-reference image quality metrics e.g., NIQE, as a reward function (See [1] for details). Other relevant non-reference IQA such as [2] are also possible to be exploited

[1] Learning the Non-differentiable Optimization for Blind Super-Resolution, CVPR 2021 [2] RankIQA: Learning from Rankings for No-reference Image Quality Assessment, ICCV 2017

Thank you, I will try!