This paper is a classic. It begins by explaining GRANDMA, a gesture-based sketching system. I'm not going to talk much about it, because that's not what's important about this work. In order to power GRANDMA, Rubine developed a gesture recognition system by combining 13 intuitive, but creative features with the classic linear discriminator. The real contribution of the work is the set of 13 features.
Rubine's features are really pretty cool. They turn a bunch of data into a handfull of really intuitive descriptors for a stroke. Humans really get "how curvey is", "how square it is", "how long it is", "how fast it was drawn". Apparently, computers really get it, too, because recognition is quite good.
My summary is pretty filled with opinion, so do read that even if you think you know what the paper's all about. To be quite honest, I often cringe a little when I think of Rubine. It's been around for so long, and I've heard it so many times, sometimes it makes me want to shout "JUST MOVE ON ALREADY". But there's a good reason that Rubine comes up so often. It's really simple (if you sort of ignore the math as just part of the algorithm) and it's got a really fast classification speed. I wonder if a significant improvement could be made by applying a newer, non-linear machine learning algorithm to Rubine's features.
I liked how the author explained both GRANDMA and the features. It's a great way to get started in this field. We see the application and the algorithm running it.
ReplyDelete