There Are Plenty Of Letters Left In The Alphabet

Summary:

Wobbrock et al present a $1 recognizer for gesture recognition. Their thesis is that gesture recognition could be a significant benefit to those working in the field of user interfaces, versus limiting the technology to those working in the field of pattern recognition. The way they accomplish this is with the $1 recognizer, which is simply a way of saying their interface is “easy, cheap, and usable almost anywhere…” The authors begin their work by providing a background of related efforts and supporting materials. Next, the algorithm goals are outlined, with the code implementation being approximately 100 lines. The algorithm itself consists of four steps: (1) Resample the point path, (2) Rotate once based on the indicative angle, (3) Scale and translate, and (4) Find the optimal angle for the best score. There are some limitations to the $1 recognizer. Among these are: (1) $1 is rotation, scale and position invariant – it is not able to distinguish gestures whose identities are tied to specific orientations, aspect ratios, or locations, and (2) $1 is time invariant – that is, gestures cannot be differentiated on the simple basis of speed. To evaluate the $1 recognizer, a compilation of 4800 gestures (16 gesture types) from 10 different users was created using an HP iPAQ. Results indicated that $1 performed at 97-percent overall accuracy when including 1 recognition template, and at 99-percent overall accuracy when 3+ recognition templates were considered. This outperformed both Rubine and the Dynamic Time Warping recognition methods. A conclusional summary and pseudo-code section were provided.

Discussion:

This October 2007 UIST paper by Wobbrock et al presents a simple and easy to use (sketch) recognizer. It is essentially in the same spirit as Rubine in that the user is somewhat limited in their gesture recognition, due to the template matching (i.e. Rubine trains classifiers based on stroke-specific training data). I was quite impressed though with the accuracy – 99-percent, which outperformed Rubine and DTW. I think for basic applications, this approach would be quite good – after all, it’s only about 100-lines of code. I also thought the user study was very complete. Ten different users providing that many samples each was a fairly comprehensive study. Overall, and even with its limitations, I thought this paper was beneficial for understanding the art and science of gesture (sketch) recognition from a very simple and easy approach. I give it four out of five starts.