Hand pose estimation using support vector machines with evolutionary training

In this paper we report our study on improving hand pose estimation using support vector machines with evolutionary procedure for selecting the training set. There are many various approaches to extract and classify hand shape features, including histograms of oriented gradients, Hausdorff distances or shape contexts. Here, we explore how to exploit support vector machines to recognize a hand pose based on the shape context descriptors. Our solution consists in classifying a vector of differences between two shapes to determine whether they represent the same pose. Such a classification framework requires learning from large and imbalanced training sets. In order to make it appropriate for support vector machines, we select a representative training sample using evolutionary strategy. Experimental study reported in the paper confirms that the proposed approach is competitive and increases the performance of hand pose estimation.