Part aggregation in a compositional model based on the evaluation of feature cooccurrence statistics

In this paper an appearance based, compositional approach to the recognition of deformable objects is presented. First, a hierarchical object model is proposed. On different levels of abstraction the model represents object categories, different views of an object, the parts of an object and basic feature vectors. Then, a training method based on multiple clustering steps is described. This paper addresses in particular the aggregation of features to parts and provides a statistical justification for feature clustering on the lowest level of the hierarchy. The performance of the proposed methods is demonstrated on a cartoon data base, where a high accuracy of 80% is achieved.