A method for learning a probabilistic model of human motion, in an unsupervised fashion, from unlabeled cluttered data is presented. The models obtained are mixtures of Gaussian models and conditional independence, described by a decomposable triangulated graph. A novel algorithm, inspired by the expectation-maximization (EM) algorithm [1], is used to learn the model structure, as well as model parameters, treating data labeling as hidden variables.
Song, Goncalves, and Perona demonstrate the effectiveness of their algorithm by using it to generate models of human motion automatically from unlabeled sequences. They report on the results of these experiments in the paper, and note that, although they have used point features in their experiments, the algorithm can also be applied to other types of features.
The paper is well written, and is self-contained with regard to the notions and algorithms described. A background in computer vision is required to properly understand the approach presented.