Matching abstract CAD models with images is a well studied problem. It includes the problems of identifying modelled objects for which 3D CAD data is available in images, and of locating them with respect to a given reference frame. Some authors have concluded that this problem has no real general solution as the representation levels are too different (see for instance the discussion in the workshop of CAD model-based vision [Bow91]). We are developing an alternative approach which overcomes this problem by representing each CAD model by several images to which 3D CAD features are added. This representation allows to solve the standard vision problems to be solved much more easily such as 'where is this CAD feature in the image?' or 'what is the object pose?'. In addition it supports fast and robust recognition using recently developed hashing techniques. Several annotated images must then be stored along with CAD data. Experiments with different kinds of images illustrate the validity of the approach.