There are some weird things？

clausmichele / ViDeNN

ViDeNN - Deep Blind Video Denoising

MIT License

224 stars 34 forks source link

There are some weird things？ #1

Closed liqiang311 closed 5 years ago

liqiang311 commented 5 years ago

Thanks for your trained model, and I run it on my own video.

As a result, there are some weird things on the image, as shown in the figure.

noisy noisy denoised noisy noisy denoise denoise

What's that? Can you help me?

clausmichele commented 5 years ago

Hi! As far as I see those images/frames contain artifacts due to compression, which of course can be seen as noise, but it was not a type of degradation that was in my training set. Maybe that's the explanation. I will try to denoise those two images on my pc and see if there are differences.

liqiang311 commented 5 years ago

Yes, you are correct, the sequence I am using is coded and there is quantization noise.

I have used the original sequence to test, and the bad phenomenon still exists, but it is much better.

denoised noisy

clausmichele commented 5 years ago

So, I've tried your noisy input and I don't get any strange output, tested with both checkpoint I provide. Which tensorflow version are you using? NOISY: DENOISED: 0001

liqiang311 commented 5 years ago

Thank you for your testing.

Python version: 3.6.4
Tensorflow version: tensorflow-gpu (1.4.0)

I only change the "xrange" to "range" in code. And I am running a series of images.

clausmichele commented 5 years ago

The only think I can try is to use your data on my machine, attach your files here if you want.

lalithnag commented 5 years ago

Hi, thanks a lot for the trained model!

I too have the same issue. Is it may be because the model is overfitting, as it has not seen these artefacts in the training data?

Python 3.6.7
tensorflow==1.12.0

(I am sending you the images via mail, as I cannot put it up online). Please also note that I can see these artefacts in all the frames!

Thanks!

clausmichele commented 5 years ago

If you use the gaussian trained checkpoint you won't have any of those weird results. Just run it with: python main.py --ckpt_dir=./ckpt_vidcnn-g

lalithnag commented 5 years ago

Thank you, I'm having a problem with running python main.py --ckpt_dir=./ckpt_vidcnn-g so I've opened this new issue.

MexWayne commented 1 year ago

The only think I can try is to use your data on my machine, attach your files here if you want.

Hi I‘m so interested in your work and I meet the same problem，it's the input png which is compressed will cause these artifacts？