Speaking of stroke order, it's hard to imagine that the image based approach should be more efficient, seeing "the need to achieve stroke-order independence".

Speaking of stroke order, it's hard to imagine that the image based approach should be more efficient, seeing "the need to achieve stroke-order independence". It seems so much more difficult if you can't use the stroke order information, or the stroke direction. Even the speed over the stroke should be able to help you along, especially as you say for cursive characters. Without stroke order it seems almost impossible to decode those in a systematic way. I suppose this is a general feature of deep learning, it's still not clear how to make good use of all the structural information we possess about a problem, instead a very brute force approach is chosen.