Modeling visual objects based on accidental alignment

A scene modeling system that preserves tangency and connectivity information is demonstrated. The model is built using an algebra of qualitative spatial relations that describes relative positions between different objects of a scene or among different components of the same object. The representation is qualitative in the sense that it distinguishes three qualitative spatial states: no contact, tangency, and overlap-contact. Other important attributes of visual objects are shape and size. It is shown how the model captures shape and size attributes by building simple operators that permit alignment-based reasoning in higher dimensions involving both rotated and translated objects. The model, which is much more compact than traditional quantitative geometric models, preserves much of the needed information for recognition and manipulation tasks.<<ETX>>