Unsupervised learning of visual feature hierarchies

We propose an unsupervised, probabilistic method for learning visual feature hierarchies. Starting from local, low-level features computed at interest point locations, the method combines these primitives into high-level abstractions. Our appearance-based learning method uses local statistical analysis between features and Expectation-Maximization to identify and code spatial correlations. Spatial correlation is asserted when two features tend to occur at the same relative position of each other. This learning scheme results in a graphical model that constitutes a probabilistic representation of a flexible visual feature hierarchy. For feature detection, evidence is propagated using Belief Propagation. Each message is represented by a Gaussian mixture where each component represents a possible location of the feature. In experiments, the proposed approach demonstrates efficient learning and robust detection of object models in the presence of clutter and occlusion and under view point changes.