Representations From Generative Systems

Abstract

Generative systems are capable of producing a consistent and comprehensive collection of patterns that meet certain constraints by means of formal components, which make explicit the primitive parts of the patterns of this collection, the allowed relations between primitive parts and underlying geometric reference frames. This explicitness suggests that generative systems can be reversed in purpose, that is, applied to the task of describing the patterns they produce. A reversed generative system describes a pattern by the sequence of steps used to generate an image (shape code). This shape code combines explicitness of potentially all image features with flexibility in retrieval and classification and with reduction of dimensionality. Description of an image by its shape code relates to cognitive theories of perception and recognition, notably recognition-by-components and the structural information theory. As an example of the application of such formalisms to the representation of architectural images the paper describes a reversal of the Palladian grammar’s room layout rule system into a compact representation of spatial arrangement. This representation can be applied to the indexing and retrieval of Palladian floor plans.