z-bingo / kernel-prediction-networks-pytorch Goto Github PK

Could you release the code of Single frame? Thank you

1. I find the dataset which you specfic in dataset_specs/full_dataset.conf is adobe5k.
   I download adobe5k, the dataset's file type is .dng which is raw pic type. So....What preprocessing do you run, to convert .dng to .bmp or .jpg ?

2. Do you finish the color mode?
   I change your code in train_eval_syn.py, KPN(color=True),and TrainDataSet(color = True), but it's occurred a error:
   RuntimeError: Expected 4-dimensional input for 4-dimensional weight 64 27 3 3 0, but got 5-dimensional input of size [32, 9, 3, 128, 128] instead

Thank you ~

Hi z-bingo,
I'm sorry to bother you with a question. I get a wrong points less than 20 dB and result with the gain of 8 while the result is correct with the gain of 1 by MFIR's compute_scores.py which has a large gap with the points in the paper. I'm looking forward to your reply.
Thanks,

Hey z-Bingo!
I was trying to download and extract the model_best.pth.tar file, though the software I have (7z, Windows 10) fails to extract the file.
might the file be corrupt? Could you provide the weight file in any other compressed format (or just leave uncompressed)?

Thanks
Nati

The current version of the code defaults to uniformly cropping input images during training and testing. In practice, when testing, we often don't want to do too many operations on the image, including cropping.
Without cropping the image, an unforeseen problem arose. nn.Conv2d cuts off edges by default for odd-sized images to make them even in size, which facilitates the flow of feature maps in the network. However, for U-Net, this results in inconsistent image sizes before and after skip concatenation, which will eventually result in an error.
A possible solution is to trim the edges of odd-sized images in the data provider during testing, so as to preprocess their sizes into even-numbered ones.
This solution idea has almost no effect on the functionality of the original method. Just dropping one or two edges of the image does not affect the evaluation of the model's denoising performance. At the same time, since the size of most images is even, only performing additional trimming operations for a few odd-sized images hardly adds additional time consumption.

When the input is from the real life picture, how to estimate the correspondding noise parameters?