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).