Colorization on 1929 movie

[ttoinou]

Hi

Thanks for this script ! I tried your colorization (v2 with class rebalancing) on some frames from Man With a Movie Camera (Vertov, 1929) here are the results :
https://drive.google.com/open?id=0B5-1OeNPsecwdTZ4dS0wUUhIaUU

• The results are all oranges :( what do you think I can do to have a better quality ?
• Should I train my own network with frames from documentaries ?
• Do you think I should start training from scratch or start from your network v2 ?
• Could we introduce some kind of temporal consistency across a bunch of frames for movie colorization ?

Thanks in advance !
Antoine.

[richzhang]

Our network is trained on modern photos, so there's a domain gap when applied to old movies. I think training your own network with old color movies is a great idea and can help bridge that gap. I would recommend starting from our network and fine-tuning on old movies.

Yes, temporal consistency would also make the results look nicer.

[ttoinou]

Thank you for your input, I'm first going to extract color frames from movies and documentaries. If I need ~500k iterations how many iter/images ? (not 1 I suppose)

[ttoinou]

I've extracted 12k frames from movies, now I need to create an LMDB file, do you think that tutorial would work http://deepdish.io/2015/04/28/creating-lmdb-in-python/ ?
Sorry for the noob question

[theColourSpace]

Hi ttoinou,

I'm also trying to use this engine for colorization on some historical images. Would have any interest in collaborating on this?

[ttoinou]

Hi, yes of course !
There's also another software for colorization : https://github.com/gustavla/autocolorize
But I didn't succeed in installing it (yet)

[ttoinou]

My email : tit_toinou@hotmail.com

[ttoinou]

If someone wants to process an entire folder he can use my script : https://github.com/ttoinou/colorization/blob/master/color.py