search

The-Learning-And-Vision-Atelier-LAVA / DASR

[CVPR 2021] Unsupervised Degradation Representation Learning for Blind Super-Resolution
MIT License
388 stars 50 forks source link

The gaint difference on 4x isotropic model with the paper writing? #41

Open xiab3369 opened 2 years ago

xiab3369 commented 2 years ago

Thanks for the author providing interesting work! Can you provide your checkpoint on 4x isotropic model? When we train your DASR as your setting on the paper, the performance of DASR is even lower than predictor+SRMDNF 0.8dB on Set5 under 4x iso! (We have train 5 parallel DASR models, and select the best one.) Now, my experiment shows that unsupervised representation has no superiority! Besides, the retraining predictor+SRMDNF is 1.7 dB higher than your paper shows on isotropic sigma=2.6 on Set14.
We wish the author to provide checkpoint on 4x isotropic model to help us find the issues.

LongguangWang commented 2 years ago

Hi @xiab3369, we have uploaded pretrained models for x3 and x4 SR. In our experiments, we used pretrained SRMDNF and only trained a predictor network to estimate blur kernel for evaluation.

xiab3369 commented 2 years ago

Thanks for your response. We also use pretrain SRMDNF to replace the generator of IKC, and fit its weithts, and only train predictor as IKC setting (remove corrector). The results show that SRMDNF+Predictor achieves 31.88dB on 4x Set5 sigma=2.4 iso kernel. Besides, the parameter of SRMDNF is much less than DASR. Thus, we propose my concerns. But unsupervised method actually performing worse than the supervised method is common. All in all, thanks for your interesting work.