Multi-context template-based gesture recognition for fast prototyping

The expansion of surface computers, low-cost tracking technologies, and several recognition modules has fostered the incorporation of gesture interaction in multiple applications. We compare the similarities and differences of sixteen template-based gesture recognizers that are easy-to-understand, -integrate, and -train, hence suitable for the fast prototyping of 2D touch interaction. Then, we evaluate their speed and accuracy on nine datasets and create a decision table to choose the right recognizer according to a set of descriptive and empirical criteria. Finally, we update this decision table by suggesting two new recognizers, μV and μF, based on the local shape distance to fill in two shortcomings found among state-of-art recognizers. The former, μV, is significantly more accurate than its peers and as accurate as $N to recognize rotation-invariant multi-strokes. However, unlike $N, it does not face a combinatorial explosion. The latter, μF, is significantly more accurate than its peers and as accurate as Jackknife to recognize 2D touch and 3D hand gestures. However, unlike Jackknife, it satisfies rotation-invariance and is suitable for the addition of new classes and templates in interactive time. We also release two tools, Gesto and Gester, to support the work of practitioners and researchers for collecting new gesture samples and evaluating the performance of one or several recognizers, respectively.