In order to locate interesting areas of an image we describe a system for focus of attention; this is based on feedback strategies combining low-level features, and a high-level object model to recognise the object and to direct the search for missing information. We aim to improve on established single-pass hypothesis generation and verification approaches by applying our complex feedback strategies to recognise generic classes of objects. By using a complex feedback strategy we produce optimal sets of low level features and reduce the number of hypotheses generated. The system can extract simple and complex objects in a scale and rotation independent manner where the objects may be partially occluded. The method is illustrated for simple cubic objects and the results are expected to be applied for a mobile robot application.