search

jixiaozhong / RealSR

Real-World Super-Resolution via Kernel Estimation and Noise Injection
Apache License 2.0
748 stars 102 forks source link

Variance cutoff used for Noise Injection #14

Open vinbhaskara opened 4 years ago

vinbhaskara commented 4 years ago

Can you please answer 5 questions because I could not find the details in your paper: 1) What is the variance cutoff you used for selecting the noise patches 2) What were the sizes of those noise patches while training (i.e. LR resolution while training). 3) Do you apply the variance cutoff on the variance of gray scale image or something else? 4) How many kernels and noise patches did you have in your Pool? 5) What is the Clean Up scale factor used?

Thanks!

vinbhaskara commented 4 years ago

@jixiaozhong Could you please respond? These settings were not made clear in the paper for reproducibility purposes.

yoon28 commented 4 years ago

Since the paper did not present the parameter settings in detail, I also want to know the secret. In addition to @bsvineethiitg 's five questions, how do you set the initial learning rate and learning rate schedule?

ALLinLLM commented 4 years ago

I guess they need waiting for the patent, just keep patient

yoon28 commented 4 years ago

@bsvineethiitg I implemented the paper unofficially. I saw that noises are very well removed but the sharpness of my results is not comparable to the original results. As you know, reproducing the paper is difficult because the author did not share many of the hyper-parameter settings. https://github.com/yoon28/realsr-noise-injection

ALLinLLM commented 4 years ago

@yoon28 That's true. I implemented the paper using KernelGAN to estimate kernel and create a kernel pool, then using the ESRGAN to train paired images. However, my results is not comparable to the original results too

qibao77 commented 3 years ago

@yoon28 That's true. I implemented the paper using KernelGAN to estimate kernel and create a kernel pool, then using the ESRGAN to train paired images. However, my results is not comparable to the original results too

+1

yoon28 commented 3 years ago

The author unveiled the training code. Check this out. https://github.com/Tencent/Real-SR After I read the author's implementation, I confirmed that my code was quite close to the author's except for some important hyperparameters. I appreciate authors to share their codes.