Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

flickering across video transfer #14

Closed
CJHFUTURE opened this issue Feb 5, 2018 · 3 comments
Closed

flickering across video transfer #14

CJHFUTURE opened this issue Feb 5, 2018 · 3 comments

Comments

@CJHFUTURE
Copy link

Hi there is a patchy areas that jump around between images in video transfer.
This isnt so much temporal issues with the style transfer of features as much as little patches of luminance change that flicker around....Is this something to do with theWhiten-Color Transform as it doesnt happen when using the Adain method with your implementation?

@eridgd
Copy link
Owner

eridgd commented Feb 6, 2018

Deleted my earlier comment because I misread the question - in regards to flickering with video transfer, that's a common problem with style transfer approaches to varying degrees. See my response here #6 for other approaches that are more stable for video.

@CJHFUTURE
Copy link
Author

CJHFUTURE commented Mar 8, 2018

I wonder if it's possible to add a -seed option, so can get repeatable results. Try running twice, with same -seed, eg -seed 123, and you should get the same output image each time? I see this has bee implemented in the original neuralStyle Johnson repo. What do you think?This may increase temporal coherence in video stylisation.

@eridgd
Copy link
Owner

eridgd commented Mar 8, 2018

Unfortunately, that won't solve the issue because the flickering is due to variability between image frames, not randomness in the net itself. For example, if you run stylize.py to stylize a single image file, then run it again with the same input, the outputs will be identical because the parameters are fixed and all ops are deterministic.

The jcjohnson repo is the original style transfer algorithm, which involves an optimization over a random noise image and so has a component that can be controlled with a seed. Optimization in the WCT algorithm occurs during the decoder training phase only.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants