Today I'm putting online a demo for my comics cutter, Kumiko 📘
The idea is to enable reading a comic book panel-by-panel rather than page-by-page, for better readability especially on mobile.
The demo with examples: https://kumiko-demo.njean.me/
The project: https://framagit.org/nicooo/kumiko
@nicooo Neat. It fails a little when comic borders overlap horizontally or vertically somewhat, e.g. https://files.explosm.net/comics/Dave/comicnewton1.png gets chopped into two panels instead of four. Same issue with https://extrafabulouscomics.com/wp-content/uploads/2019/05/Forever-After-e1558458111634.png.
Also, I've no idea what this one's doing: https://files.explosm.net/comics/Kris/safe.png
And it doesn't cope with background noise (e.g. https://extrafabulouscomics.com/wp-content/uploads/2017/05/8-e1494652694789.jpg); perhaps a preprocessing step could fix that? That's the only thing on this list I can actually see how to improve.
(These are all evil examples.)
Thanks a lot for testing it out and providing evil examples! 🤗
Your second one is the worst (safe.png), with panels overlapping *and* not well defined. 😀
Kumiko is indeed far from perfect; for now it works with pretty plain contour detection. I haven't coded many pre- or post-processing steps, which I sure hope to.
Another idea would be to provide processing options from within the demo: background color, opencv parameters, etc. to better tune Kumiko.
Have you seen that you can switch from panel view to page view? ('p' key or bottom-left icons)
This will show you the panels that Kumiko extracted, giving more insight on what was computed internally.
@nicooo Yes, I have. That's how I filtered for the evil examples.
If the comics are in panels with a black border and a horizontal and vertical line >= 1 pixel wide of white between them, it has a very high success rate. Unfortunately, most comics are not like that.
Le réseau social de l'avenir : pas de publicité, pas de surveillance institutionnelle, conception éthique et décentralisation ! Gardez le contrôle de vos données avec Mastodon !