Learning in an active hybrid vision system

This paper focuses on learning of object models for an active robot vision system. One of its main attributes is the generation of hybrid models of 3D objects, integrating implicit representations by neural networks and explicit descriptions by semantic networks. On both levels of the vision system, subsymbolic neural learning as well as symbolic semantic learning can be done completely unsupervised after defining a few constraints only. This allows us to adapt our vision system to new objects and domains without intensive training phases and without „handcrafting“ object models by an expert. Indeed, a new object has only to be presented once under good vision conditions to the robot vision system to be learnt for robust recognition.