Generic Hierarchic Object Models and Classification Based on Probabilistic PCA

In this paper we tackle the problem of classifying objects, which are not known to the system but similar to some of the objects contained in the training set. This type of classification is referred to as generic object modeling and recognition and is necessary for applications were it is impossible to model all occurring objects. As no class for unknown objects exist, they are either rejected or assigned to the most similar class contained in the training set. Even in the case of soft assignments this can lead to wrong interpretation of the actual class membership. We present a new approach for generating appearance based hierarchical object models based on probabilistic PCA for generic object recognition. During the training step a hierarchical set of mixtures of probabilistic PCA models is generated. This represents a coarse-to-fine gradation with respect to the reconstruction ability of the training views at each hierarchy level. So coarse parts of the training views are covered on higher levels whereas the lower levels cover more details of the encoded training views. The mixture components are calculated at each hierarchy in an unsupervised manner using the expectationmaximization algorithm.