A MINIMUM DESCRIPTION LENGTH MODEL FOR RECOGNIZING OBJECTS WITH VARIABLE APPEARANCES (THE VAPOR MODEL)

Most object recognition systems can only model objects composed of rig id pieces whose appearance depends only on lighting and viewpoint. Man y real world objects, however, have variable appearances because they are flexible and/or have a variable number of parts. These objects can not be easily modeled using current techniques. We propose the use of a knowledge representation called the VAPOR (Variable APpearance Objec t Representation) model to represent objects with these kinds of varia ble appearances. The VAPOR model is an idealization of the object; all instances of the model in an image are variations from the ideal appe arance. The variations are evaluated by the description length of the data given the model, i.e., the number of information-theoretic bits n eeded to represent the model and the deviations of the data from the i deal appearance. The shortest length model is chosen as the best descr iption. We demonstrate how the VAPOR model performs in a simple domain of circles and polygons and in the complex domain of finding cloverle af interchanges in aerial images of roads.