There Are Plenty Of Letters Left In The Alphabet

Summary:

Mantyjarvi et al present a 2004 ACM article on enabling fast and effortless customization for accelerometer-based gesture interaction. The thesis of their work is that given the ubiquity of mobile computing, a limited-scale gesture recognition approach should be developed to accommodate the existing technology. In their approach they focus on discrete gesture commands – those hand movements that can be captured and used at any point; direct measurement and continuous gesture commands are not considered in this paper. To interpret the accelerometer data in the recognition system, a discrete Hidden Markov model approach is taken. The system consists of the two standard phases: (1) training and (2) recognition. Initially though, noise is added to each sensor input, which goes through pre-processing and vector quantization. The goal of the system is to minimize user effort in the training phase (and overall) – that is why the noise is added. A wireless prototype of the system was constructed to control a portable DVD player. Control gestures included: Play, Stop, Next, Previous, Increase, Decrease, Fast Forward, and Fast Rewind. A user study of one person was conducted. A recognition accuracy of 97- to 98-percent was achieved. The authors end with a conclusional summary and discussion of future work, which include expanding the sensor spectrum to other devices (e.g. gyroscopes).

Discussion:

This conference paper by Mantyjarvi et al presents a noise-based HMM approach to hand gesture recognition. I suppose their approach is worthwhile, given that the results seemed quite good. However, I am always a little suspicious about the performance results when we start injecting white noise to help capture signal variability. One sees this with other ‘black box’ approaches, such as neuro-fuzzy systems. Other than the approach, I have to call the authors out on two additional points. First, they only had a one person user study. I think Dr. Hammond has rubbed off on me too much because as soon as I got to that point in the paper where the user study was mentioned, I just cringed. I mean, I personally have coded plenty of sketch stuff such that if it is just me doing the input, I can get outstanding accuracy every time. The truth then comes out when I show it to, let’s say Dr. Hammond again for example, and then she easily breaks my stuff. So, long story short, the one person user study seriously diminished the quality of the paper. The other issue I had was that they did not include a Pause gesture. WTF?!? I pause DVDs, iTunes, etc. all of the time – so does everyone else. So, why not include a Pause? That is, unless to the authors, Stop = Pause. Could be. In the end, I guess it was an ok paper: two out of five stars.