Learning an Object Model for Feature Matching in Clutter

We consider the problem of learning an object model for feature matching. The matching system is Bayesian in nature with separate likelihood and prior parts. The likelihood is based on Gabor filter responses, which are modelled as probability distributions in the filter response vector space. The prior model for the object shape is learnt in two stages: in the first stage we assume only the mean shape known, with independent variations for each feature point, and match ’easy’ images. We then estimate the characteristics of the shape variations for a realistic prior on the shapes. We demonstrate how incorporating the shape variation prior into the matching model enhances matching performance in the presence of clutter.