This chapter surveys gesture recognition techniques. The chapter begins with Rubine's feature-based algorithm, which was one of the first gesture recognition algorithms, especially to achieve such high accuracy. Next, Hammond presents the Long extensions to Rubine's feature set. These two algorithms (both the original and the extended) are commonly used today and usually find good results.
Finally, Hammond introduces Wobbrock's $1 recognition algorithm, which is based on template matching as opposed to feature-based classification. The $1 algorithm is especially notable because it's so simple a small child could implement it, yet the accuracy is comparable to Rubine and Long.
This chapter gives a very good brief survey of gesture recognition techniques. Rubine is basically the seminal paper in the field, and Long is a natural extension. Wobbrock provides a good balance with a really different approach, and demonstrates that an algorithm does not need to be complex to work well. The Electronic Cocktail Napkin might be an appropriate inclusion because it is also vastly different from Rubine/Long or Wobbrock.
In the case of Rubine, the survey chapter has a much clearer focus on the meaning and importance of each feature, and gives less importance to GRANDMA, which has not aged as well as the core algorithm. After reading such a good, thorough overview of the algorithm, I'm questioning why we're reading the original paper, too. In the case of $1, the original paper is just as clear, just as focused, and more detailed. I would rather have just read that.
Since I'm simple minded and easily distracted, Aaron's notes and the missing figures were distracting.
I also found those comments distracting ;-) I'm just wondering how long does it take to train the system with for example 15 samples in each class.
ReplyDeleteIt does not take long according to my experience. The time cost varies due to different classifiers.
ReplyDeleteOverview is very useful for beginners to get familiar with their field and find which papers are worth reading. But if you want to learn more, you must read the original papers. Different persons have different views, even on the same object. Maybe you read the original paper, and have different view with the overview.
I found the paper rather informative and I agree with Drew that it was right for the paper to focus on the algorithm and its components instead of GRANDMA.
ReplyDeleteSimilarly, I appreciated the in-depth explanation of the algorithm components for all 3 of the algorithms explained.