Toward the unsupervised interpretation of outdoor imagery

Abstract A new approach to the interpretation of outdoor scenes is described. It is based on a context-sensitive classifier which uses relative constraints to describe global relationships between object classes. Contextual models represent the structure of a scene's underlying feature space in terms of stable, physically based parameters. A discrete relaxation algorithm is used to find unambiguous labelings that satisfy a set of ordering relations between object classes. Unlike rule-based systems, these constraints provide a complete and consistent description of the scene. Scenes that are similar in structure are organized into contexts, each of which is represented by a consistent set of constraints. Instead of attempting to achieve a high degree of specificity and localization within limited domains, the methodology is geared toward recognizing general kinds of objects with little or no human intervention over a wider range of scenes. Several examples which demonstrate the recognition of simple objects in black-and-white and multispectral imagery acquired by aircraft, satellite, and at ground level are presented. Through a series of experiments, the ability of the system to degrade gracefully in performance when faced with new and unknown situations is demonstrated.