In this paper, knowledge-based recognition of objects in a bureau scene is studied and compared using two different systems on a common data set: In the first system active scene exploration is based on semantic networks and an A✻-control algorithm which uses color cues and 2-d image segmentation into regions. The other system is based on production nets and uses line extraction and views of 3-d polyhedral models. For the latter a new probabilistic foundation is given. In the experiments, wide-angle overviews are used to generate hypotheses. The active component then takes close-up views which are verified exploiting the knowledge bases, i.e. either the semantic network or the production net.