We will discuss how hypotheses of geomemc structure can be generated in an active computer vision system. These hypotheses are built on sparse but reliable observations and can communicate with higher level hypotheses as well as the basic classified data hypotheses. We present experiments with such hypotheses, where we have chosen to form hypotheses about the existence of planar surfaces in a scene, given classified edges and junctions. The results of the experiments illustrate clearly the benefit of the approach. A goal directed active vision system could easily use this idea to make both powerful verifications of generated predictions as well as hew hypotheses and predictions at different level of abstraction.