Canonical Superpixel-Graph

NTU hand digit dataset

HKU hand gesture dataset

HKU multi-angle hand gesture dataset

Microsoft Kinect & Leap Motion Dataset

Creative Senz3D Dataset

We performed two tests with different template selection schemes on the HKU hand gesture dataset [12] without utilizing the color and depth information. In the first test, the templates are randomly selected from the training data. In the second one, the templates are selected based on K-Medoids clustering [R7]. Each experiment is repeated for 20 times. The curves of mean accuracies with different numbers of templates are shown in Fig. R1. It can be seen that the performance is not very sensitive to the selection of the templates. For LOO CV, the accuracy drop is about 4% even when only 4 templates (one for each subject) are utilized. Although the recognition rates for L4O CV are quite low (about 80.9% and 88.5%) if only one template is available, the rates are improved significantly (about 92.6% and 93.3%) when the number of templates is increased to 4.

[R1] Y. Zhang, C. Wang, J. Zhao, L. Zhang, S.-C. Chan, “Template Selection based Superpixel Earth Mover’s Distance Algorithm for Hand Gesture Recognition,” ICSP, Chengdu, Nov. 2016.

Fig. Mean accuracies with different numbers of templates. Left: Templates are randomly selected. Right Templates are selected based on K-Medoids clustering [R7].