GuoShi28 / CBDNet

Code for "Toward Convolutional Blind Denoising of Real Photographs", CVPR 2019

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

some artifacts in generate ground truth data

conson0214 opened this issue · comments

I try to use ./SomeISP_operator_python/ISP_implement.py to generate synthetic noisy image & ground truth image(gt), but in many images, i saw artifacts like this

cbd_synthetic_issue_gt

and the raw srgb noisy-free image is like this

cbd_synthetic_issue_clear

I wonder if these artifacts are reasonable in synthetic ISP process, and why these happens?

In Python code, the clean image is converted to RAW and then convert back to the sRGB image. In my understanding, these artifact is caused by the demosaic process. In original Matlab code (strictly following the paper), the original sRGB image is directly treated as groundtruth image. However, when I read more paper, the original setting is kindly like denoising+demosaic, so in the python code, I change a little bit. You can try to use python code to train or directly use original RGB as groundtruth (as in my paper), if you find other questions, please feel free to contact me.

Hi thank you for your great work first!
Actually it seems that you are taking the rgb values from sRGB image and putting them in the corresponding Raw RGB location, and I don't think it's a real inverse process of demosaicing in ISP, thus caused the shown artifacts

Hi thank you for your great work first!
Actually it seems that you are taking the rgb values from sRGB image and putting them in the corresponding Raw RGB location, and I don't think it's a real inverse process of demosaicing in ISP, thus caused the shown artifacts

Hi, thanks for your concern about this issue.
I wonder if its not a real inverse process of demosaicing in ISP, whats the real inverse process like?
If you may supply some improving direction?

Hi thank you for your great work first!
Actually it seems that you are taking the rgb values from sRGB image and putting them in the corresponding Raw RGB location, and I don't think it's a real inverse process of demosaicing in ISP, thus caused the shown artifacts

Hi, thanks for your concern about this issue.
I wonder if its not a real inverse process of demosaicing in ISP, whats the real inverse process like?
If you may supply some improving direction?

https://link.springer.com/article/10.1007/s11263-017-1056-0, check this out
A fully inverse of demosaicing is not a simple thing, but if you only want to make a simulation, you can just fellow demosaicing process since it's just a non-linear process, or more precise, interpolation process.
In my opinion, if the two process match, there won't be any artifacts