We address the problem of automatic model building for further recognition of objects. Our initial data are a set of images of an object. In a first stage, these images are put into correspondence using quasi-invariants, epipolar geometry and an approximation of the apparent motion by an homography. The different aspects of the objects may thus be computed and each aspect gives raise to a partial model of the object. In a second stage, these models and indexed in a data base which is used for recognition. This work is based on the idea that aspect graphs may (should?) be learned from examples rather than computed from CAD models, and that a planar representation associated with geometric quasi-invariants is a relevant tool for object recognition.