Figure 1From: Novel Kernel-Based Recognizers of Human ActionsUnsupervised action recognition is performed by computing a feature vector for each frame of the sequence (left), as described in Section 6. A nonparametric probability distribution on the feature space is computed for each sequence (center); different sequences are compared by mapping the corresponding probability distributions to an RKHS by means of (right): in particular, sequences representing the same actions are clustered by using MMD as a distance metric.Back to article page