We describe an unsupervised, probabilistic method for learning visual feature hierarchies. Starting from local, low-level features computed at random locations, the method combines features hierarchically. At each level of the hierarchy, pairs of features are identified that tend to occur at stable positions relative to each other, by clustering the configurational distributions of observed feature cooccurrences using Expectation-Maximization. Stable pairs of features thus identified are combined into higher-level features. This learning scheme results in a graphical model that constitutes a probabilistic representation of a flexible visual feature hierarchy. For detection, evidence is propagated using Nonparametric Belief Propagation, a recent generalization of particle filtering. In experiments, the proposed approach demonstrates effective learning and robust detection of objects in the presence of clutter and occlusion.