COMMENTS:
Jianjie Zhang
SUMMARY:
This paper begins with a brief description of gesture recognition and how it relates to sketch recognition. It then gives an overview of some foundational methods used for gesture recognition.
Hammond discusses Dean Rubine’s gesture recognition method and gives an explanation of each of the 13 stroke features that Rubine uses to recognize gestures. Hammond also discusses Christopher Long’s work on gesture recognition. Long’s method is an extension of Rubine’s that uses 22 features, 11 of which came from Rubine. Long did not feel that time features contributed to recognition and therefore did not include them in his method. Hammond also talks about Wobbrock’s $1 recognizer which uses a template matcher as instead of a feature-based method.
DISCUSSION:
This chapter is a great one-stop reference for a collection of gesture recognition methods. It provides a brief and clear example of each of Rubine’s features. I think the organization of the features along with their descriptions are a little easier to read here than in the Features section of Rubine’s Specifying Gestures by Example paper, but maybe that’s just me.
I kind of like the update notes that Aaron added (although some of them were repeated), particularly the one about how higher sampling rates can cause problems when the rotational change and smoothness features of the stroke. However, I don’t understand how deleting consecutive points isn’t accomplished anyway by resampling. Also, I was confused by some of the figures in the paper, but the ones that fit with the explanations were very helpful.
I completely feel the same way about the descriptions of the features in this reading. In the original paper they are just exposed, here they are explained in a friendly way easy to understand. In general I think the whole reading is easier to digest than the original, except perhaps for the positions of the figures which also confused me a little bit but I guess that can be blamed on LaTeX.
ReplyDeleteAs for the sampling problem I agree that resampling should eliminate this problem, however I understand that resampling is done only in the $1 recognizer, and this note is on the Rubine's where all the gesture data is kept with the original sampling rate, so it makes sense to take this preprocessing step into account.