Paper Reading #22: A $3 Gesture Recognizer – Simple Gesture Recognition for Devices Equipped with 3D Acceleration Sensors

Comments

Reference information
Authors: Sven Kratz and Michael Rohs
When/Where: IUI’10, February 7-10, 2010, Hong Kong, China

Summary
In this paper, the authors introduce a cheap 3D gesture recognition system, called $3 Gesture Recognizer, that can be implemented quickly. Unlike most other recognizers, it does not require large training sets for complex pattern recognition techniques; instead, it is based on a previous 2D recognizer that opted for simpler trigonometric and geometric methods. It compares user gestures, or traces, to a small library of gesture classes for recognition. The system also uses resampling in order to accommodate the varying gesture speeds between users (with N = 150 samples found to be optimal). But unlike the previous work, $3 Gesture Recognizer implements a heuristic to reduce the occurrence of false positives with the following algorithm:

• ε is defined as the threshold score.
• If and only if the highest-scoring candidate in the score table has a score > 1:1, return this candidate’s gesture ID.
• If and only if, within the top three candidates in the score table, two candidates exist of the same gesture class and have a score > 0:95, respectively, return the gesture ID of these two candidates.
• Else, return “Gesture not recognized!”

The designers' preliminary gesture vocabulary. Source: paper

For evaluation, the designers conducted a user study with twelve participants. Users performed fifteeen gestures each, while their gestures were evaluated. Average (correct) recognition rate was 80%. Between users, recognition rate varied between 58% and 98%, with a standard deviation of 11:4. While the designers noted some of the system's limitations, they believed the system's overall performance was comparable to other existing systems. 

Discussion
This was a pretty interesting paper, not only because of the implementation, but because I am interested in gesture and sketch recognition, or pattern analysis in general. I liked that the designers acknowledged the system's limitations, particularly that it could not detect gestures continuously and that it used a much smaller gesture library, I also liked that the recognizer was so rudimentary yet yield similar results to those with more advanced recognition techniques. The evaluation study seemed thorough, though I am curious about the participants, like the gender and age distributions.