Small gestures go a long way: how many bits per gesture do recognizers actually need?
We investigate in this work the effect of bit depth on the performance of today's commonly used nearest-neighbor gesture recognizers. As current bit representations are typically an artifact of today's hardware and file formats, they are not reflective of the true cardinality of gesture data. We show that as few as 4-5 bits per gesture channel (x/y) are enough in order to attain peak recognition for Euclidean, Cosine, DTW, and Hausdorff distances. We also show how reduction in bit depth can lead to 85 times less memory for storing the training set without ruining recognition performance. The results will benefit practitioners of the next age of gesture sensing gadgets and devices that need to optimize speed, memory, and bit depth representation in their software and hardware designs.