Person identification from action styles

We consider a problem of identifying people based on their styles in performing actions from an arbitrary predefined set of action types. We present a generative model describing the action instance creation process and derive a probabilistic identity inference scheme, which implicitly includes action type inference as one of its components. Our experiments validate the power of the approach. We report high recognition rates on four publicly available action recognition datasets and one dataset for person authentication, on which we obtain state-of-the-art results. We make use of existing action representations and show that combining them with an action-specific Mahalanobis metric, learned from examples, improves the results.