They didn't think to weight the prior probabilities by usage frequency* - drawing a reasonable ? gives me ȓ, ᕉ, ╔, ᣑ, Ѓ, ק, ᒌ, ŕ, ᒤ, ᒦ, ņ, ᒯ, ѓ, and finally ?.
I'm also guessing that they're directly comparing the handwritten character to some version of the unicode character rather than with human attempts to draw the character. Human drawings are often quite different (more slanted, stylised etc.) than typeface characters. This is much more forgiveable though because assembling a good dataset for human drawn characters is hard (especially for any reasonable chunk of the unicode set).
(*this is fairly easy to do: just find some large source of typical unicode, like Wikipedia in all languages, and index them).
They didn't think to weight the prior probabilities by usage frequency
I don't know if that is the right metric for this sort of tool. I'd guess the use case is for trying to find infrequently encountered characters. It should probably try to detect your current locale, and then say eliminate all ASCII characters when you are suspected of speaking English, etc., since you are already aware of how to type a question mark.
Yeah, it seems like they're taking the drawing and matching the edges and points up with edges and points in a font file. Hopefully they'll start weighing by frequency and add more glyphs soon!
Comments
They didn't think to weight the prior probabilities by usage frequency* - drawing a reasonable ? gives me ȓ, ᕉ, ╔, ᣑ, Ѓ, ק, ᒌ, ŕ, ᒤ, ᒦ, ņ, ᒯ, ѓ, and finally ?.
I'm also guessing that they're directly comparing the handwritten character to some version of the unicode character rather than with human attempts to draw the character. Human drawings are often quite different (more slanted, stylised etc.) than typeface characters. This is much more forgiveable though because assembling a good dataset for human drawn characters is hard (especially for any reasonable chunk of the unicode set).
(*this is fairly easy to do: just find some large source of typical unicode, like Wikipedia in all languages, and index them).
I don't know if that is the right metric for this sort of tool. I'd guess the use case is for trying to find infrequently encountered characters. It should probably try to detect your current locale, and then say eliminate all ASCII characters when you are suspected of speaking English, etc., since you are already aware of how to type a question mark.
I just drew a question mark, and it showed up as the first result with a score of 0.899462
Yeah, I just tried it with the inverted question mark (¿) and interrobang (‽) -- both attempts had the correct glyph for the first result.
Maybe we're training it?
Probably not training "matching", given that there is no place where you indicate what the correct thing was.
It could be gathering data about how people draw shapes in general, but it's not immediately obvious to me how much that can help.
Actually there is a "Good" and "Bad" voting feature for each result.
Retracted then. I hadn't noticed that.
I've been completely unable to get it to recognize an inverted interrobang (⸘) with no luck, though. Still, very very cool.
EDIT: Tried one more time and got it as the 9th suggestion.
Yeah, it seems like they're taking the drawing and matching the edges and points up with edges and points in a font file. Hopefully they'll start weighing by frequency and add more glyphs soon!
sounds like you split up the stem and the bowl.