I made this largely for my own amusement/education, but others may find it useful:
https://github.com/hanysz/img2sgf
It's specifically for converting diagrams to SGF so that I can scan pages from printed books on my shelf and load them up into Lizzie+KataGo more quickly than I can input all the moves by hand. It's not great with photos (can't handle perspective distortion) or coloured images, although I've found that it can handle screenshots from some YouTube videos. I wanted something that's not put off by stone numbers or labels on the diagrams, which Imago seems to have trouble with.
Sorry that installing it is such a hassle. I don't have the chance to try it out on a range of computers. So if you manage to make it work, and find it useful, please post your tips here!
Diagram image to SGF: experimental software
-
Ferran
- Lives in gote
- Posts: 664
- Joined: Mon Apr 22, 2019 1:04 am
- Rank: OGS ddk
- GD Posts: 0
- KGS: Ferran
- IGS: Ferran
- OGS: Ferran
- Has thanked: 177 times
- Been thanked: 121 times
Re: Diagram image to SGF: experimental software
Great! And thanks a lot. I'll try it. Have you considered (or tried) against old Chinese numbering? There are lots of games that would become available. But I imagine it might be "a tad" harder.xela wrote:It's specifically for converting diagrams to SGF so that I can scan pages from printed books on my shelf and load them up[...]
Take care.
一碁一会
-
xela
- Lives in gote
- Posts: 652
- Joined: Sun Feb 09, 2014 4:46 am
- Rank: Australian 3 dan
- GD Posts: 200
- Location: Adelaide, South Australia
- Has thanked: 219 times
- Been thanked: 281 times
Re: Diagram image to SGF: experimental software
At this stage I'm ignoring move numbers. The output is just a static board position. You get the final position of the diagram, not the sequence of moves.
The reason is that I can tell the difference between black and white stones using "old school" image recognition technology. But to recognise numbers, I'd have to train a neural net, which means I need to collect and label a bunch of training data. Technically speaking, it's shouldn't be that hard, but it would eat up more time than I want to put into it. Chinese numbers shouldn't be any harder than Arabic numbers.
The reason is that I can tell the difference between black and white stones using "old school" image recognition technology. But to recognise numbers, I'd have to train a neural net, which means I need to collect and label a bunch of training data. Technically speaking, it's shouldn't be that hard, but it would eat up more time than I want to put into it. Chinese numbers shouldn't be any harder than Arabic numbers.
-
Ferran
- Lives in gote
- Posts: 664
- Joined: Mon Apr 22, 2019 1:04 am
- Rank: OGS ddk
- GD Posts: 0
- KGS: Ferran
- IGS: Ferran
- OGS: Ferran
- Has thanked: 177 times
- Been thanked: 121 times
Re: Diagram image to SGF: experimental software
Well, the ones I've found are in older books, ie. not as easy to read or clearly printed.xela wrote:Chinese numbers shouldn't be any harder than Arabic numbers.
Take care. Thanks for the effort.
一碁一会