|
Weber, Markus Unsupervised Learning of Models for Visual Object Class Recognition We present a method to learn the structure of object class models used for recognition. We focus on a particular type of model where objects are represented as constellations of rigid features (parts). The variability within a class is represented by a joint probability density function (pdf) on the shape of the constellation and the output of feature detectors. The pdf may be estimated from training data once a model structure (type and number of features) has been specified. The method automatically identifies distinctive features in the training set and learns the statistical shape model. It is assumed that a set of generic feature detectors is available for the learning algorithm to choose from. The entire set of model parameters is learned using expectation maximization.
|