and for the very curious, my bachelor thesis is on shapecatcher, http://shapecatcher.com/B_Milde%20-%20On%20The%20Security%20...

There is a whole chapter on shape contexts in it, which I use with shapecatcher, too.

I think there is away to not host images - I could just record strokes and play that back with js.

Right. Hiragana and Katakana ist trivial (just a couple of characters), but no support for Kanjis currently

1) it does count edge pixels - recognition is based on the shape of what you draw, so for the algorithm, a black picture is like a white one (but I admire your endurance to paint it all black) 2) server is under heavy load atm, it might drop some requests - just retry

You have to draw the snowflakes

There is a prior probability in the results - but it's hard to balance it so that latin characters don't dominate the results if you search something else. That is why I opted for the way it is now - this is a tool to find unicode characters, you have latin characters on your keyboard. It's not really meant for OCR

You could do prior probability conditional on not being available on a US keyboard, or at least make that an option.

edit: just read your comment elsewhere on the page indicating that you do something like what I described.

I was working on this last weekend - sorry I didn't know someone would post this is again to ycombinator ;)

Exactly, this is why priors for latin alpha characters are not set that high - you have them on your keyboard already.

I would like to compile it on my own with emscripten. Did you use emconfigure or did you write you own compilation script? Couldn't find any pointers to that in the Git, but maybe I didn't look around long enough to find it. How did you deal with kpathsea?