cameronfabbri / underwater-color-correction

Using GANs to correct color distortion in underwater images.

Underwater GAN (UGAN)

[Project Page] [arxiv] [Video]

Color correction for underwater vision. This project uses CycleGAN to perform style transfer on "clean" underwater photos to appear to be underwater.

We used this video from Youtube as our distorted domain, and hand picked images from Imagenet for our "clean" domain. CycleGAN was used to distort clean images to provide us a dataset.

Results

CycleGAN Comparison

Other videos in consideration:

https://www.youtube.com/watch?v=OSdrb1XNXZI

https://www.youtube.com/watch?v=ALN6y2PLCq0

underwater-color-correction's Issues

Info.pkl for pre-trained model

I don't fully understand the purpose of the info.pkl file in the convert_video file. Is there any chance you can supply a sample one for the pre-trained model so I can look at it? Thanks for any help you can give.

issues about CycleGAN

i can't find that how long is it spent on CycleGAN training to generate distorted images in your paper.
Does GTX 1080 have the ability to train CycleGAN?cause i found CycleGAN has nine residule_blocks!
I am not sure about what CycleGAN input is? distortion images and clear images？Where do these distortion images come from?
I just started learning GAN, thank you very much for replying！

can not find The flickr/ folder images

Couldn't find the 56 flickr images mentioned in the paper to compare with cyclegan
If you still have a copy, could you send me a copy? Thank you very much
[email protected]

model problem

Dear owner,I just want to run the code and model,but I can not download the checkpoints and datasets from you project page.Maybe the reason is that my IP is bannd,could you send me your datasets and checkpoints to my mail box,[email protected] you very much.

Can you tell me which line in the code represents patchGAN?

No images in tests/test_images in ICRA branch

No images are found in tests/test_images. Where can I get ?

dataset

Hi,

Could you please share the link to the dataset.

Thanks.

Should I set the is_training flag for the BN layer?

I encountered a different output of the same image during the test.
Looked at the nets/pix2pix.py file, it seems that the batch_norm function does not set the is_training flag. And when defining the optimizer, it does not add a dependency on UPDATE_OPS.
tf.contrib.layers.batch_norm
Is it because the network training can't learn a good mean and variance, so that the is_training state of the BN layer is forced to be True? But this results in different outputs when the batch size is different.

Looking forward to your reply! Thank you.

Artifacts in reconstructed images

After downloading your checkpoints (pretrained network) and generating images on different test images, the generated images include artifacts. Please check the attached images for more details. Could you please check your code or your checkpoints again? Thanks :)

how to use it for testing

hello @cameronfabbri , @junaedsattar ,
I tried to use model directly on a test image not sure if its possible.
I want to use it for some images but there are no instructions so not getting how to use it.
Thank you for the awesome work and Help :)

either dataset or pretrained model

hi @cameronfabbri ,
Is there the dataset that can we try it or the pretrained model?

pre-training model for reference

Hi,
Is there a pre-training model?May I take it as a reference?
Could I do a comparative experiment?
Thanks,

About the convert_video.py

The netG function is not found if I use the pix2pix network, how I can change it? If I use the resnet network, the netG function is netG(x), not netG(x, loss_method)!

What are the data you have used for the training?

What is the data you have used for the training?
Could you create a requirement.txt for us to try to run your code?