Towards an integrated framework for contour-based grouping and object recognition using Markov random fields

We present an integrated approach for contour-based grouping and object recognition. Domain knowledge and domain independent grouping laws are combined in a multi-layered Markov random field (MRF) framework. It provides a basis for propagating top-down knowledge between different processing cues or input modalities. Additionally, the domain dependent MRF-layer can be used in order to evaluate the grouping process with regard to relevant contours for object recognition. Initial results show the approach to be adequate for complex scenes and partially occluded objects.