I created a method similar to this one for curve extrapolation based on sample curves. I took the second derivative of each point on the training data, added it's contributions to a vector field.
To extrapolate, you can take 'position' and 'velocity' of live data and integrate over the vector field to produce an extrapolation. I enjoyed the project and it worked fairly well. I think there's plenty of room for extension around this method.
Weren't going to work so I switched to the style in the font:
╱│
∠--┼
│
And it really liked to give 7, 6, or 2 if I went in the left,upRight,down pattern.
I think I found out the trick though: if you make a sharp 4 like the font you get random results but if you just make smooth loop of it (almost looks like a tilted roller coaster) it gets it every time.
Even went I deleted the preset glyph patterns and spent 5 minutes re-training it for 1-9 my training pattern for '4' still didn't work. I tried drawing it 2 different ways. It's neat that you can train it so quickly with the hold-shift feature though.
It appears to be a dot product of a vector field and the curve vector (i.e. a line integral).
I think this approach can't handle certain cases where a path is retraced in opposite directions, like with the handwritten letter 't' or 'i'. Perhaps curvature integrals could be added as well to improve it.
I don't think so.
I draw my 9s by drawing the circle, and then the line down.
Looks like you go the other way. This is all about direction and gesture as opposed to OCR.
https://assets.opentoken.com/sha256/N1ShZ7kcVmH_-auZEGaGOS_b...
Ah nevermind, it's matching the curve against template fields so it makes sense that it would show the best match even when the match is poor.