search

sjmoran / CURL

Code for the ICPR 2020 paper: "CURL: Neural Curve Layers for Image Enhancement"
205 stars 34 forks source link

The codes is not designed for RGB images #2

Closed hermosayhl closed 4 years ago

hermosayhl commented 4 years ago

I have made lots of efforts to run the codes with MIT-Adobe FiveK datasets. However, the codes are designed for RAW images rather than RGB pairs. It's consuming my time. Would you please release one version of codes for RGB images?

sjmoran commented 4 years ago

Hi hermosayhl, firstly thank you for your interest in our work! The RGB variant of the low-level network is similar to the RAW, but you are correct that there are some key differences (see supplementary material of the paper). I will upload an RGB network for you.

hermosayhl commented 4 years ago

Hi hermosayhl, firstly thank you for your interest in our work! The RGB variant of the low-level network is similar to the RAW, but you are correct that there are some key differences (see supplementary material of the paper). I will upload an RGB network for you.

Thanks for your reply!

I have managed to run the codes with RGB image pairs although some hyber parameters are changed. As long as one can dedicated himself to the codes.

I'm not sure whether my changed version is correct. However, the initial trained results are pleasing so far. After I finished training, I may upload the changed codes and pretrained models for all friends.

hermosayhl commented 4 years ago

Hi hermosayhl, firstly thank you for your interest in our work! The RGB variant of the low-level network is similar to the RAW, but you are correct that there are some key differences (see supplementary material of the paper). I will upload an RGB network for you.

Listed below are some of the images for testing:

image

image

image

sjmoran commented 4 years ago

Good to hear that you got the RBG variant working. And thank you for contributing the code and trained models to the community. To clarify, in your output images, from left-to-right: the left most image is the input. Which of the other two images are the prediction and ground-truth?

hermosayhl commented 4 years ago

Good to hear that you got the RBG variant working. And thank you for contributing the code and trained models to the community. To clarify, in your output images, from left-to-right: the left most image is the input. Which of the other two images are the prediction and ground-truth?

Firstly

the images produced are: [input image, label image, enhanced image].

Secondly

The training process is coming to the end. I will change the code and trained models as soon as possible.