Latent model clustering and applications to visual recognition

We consider clustering situations in which the pairwise affinity between data points depends on a latent "context" variable. For example, when clustering features arising from multiple object classes the affinity value between two image features depends on the object class that generated those features. We show that clustering in the context of a latent variable can be represented as a special 3D hypergraph and introduce an algorithm for obtaining the clusters. We use the latent clustering model for an unsupervised multiple object class recognition where feature fragments are shared among multiple clusters and those in turn are shared among multiple object classes.