Natural Object Recognition

Much early machine-vision research in the modern signals-to-symbols paradigm was concerned with the interpretation of scenes from the “blocks world.” Line drawings of simple geometric objects were analyzed to infer the shapes of individual objects. More recent research has focused on the recognition of man-made objects, such as industrial parts in a factory setting, roads in an aerial photograph, and furniture in an office environment. In these systems, several complicating factors that were not present in the blocks world had to be addressed: namely noisy images, imperfect geometric models, and complex lighting. The complexity of description necessary for recognition was greater than that required for the blocks world. A logical next step in this progression is the interpretation of ground-level images of natural outdoor scenes. In the manufactured world, three-dimensional (3D) edges and surfaces are an adequate intermediate representation, but for the natural world, such shape descriptions are insufficient and perhaps inappropriate. By designing a vision system for interpreting ground-level scenes of the outdoor world, we hope to provide a new basis for a theory of computational image understanding in complex domains.

If robots are ever to attain versatile and autonomous behavior, it will be necessary to endow them with perceptual abilities that go far beyond the geometric reconstruction that modern robots perform. There is a tremendous difference between the expectations placed by robot designers on a perception system and the capabilities that the field of machine vision has so far provided.

Rather than study natural object recognition in the abstract, we have chosen to focus our research on the visual requirements of a particular task. Evaluation of the merits of any approach or theory can be carried out only within the scope of an intended purpose. Natural object recognition is too broad and ill-defined to serve as a useful goal for machine vision unless a task is chosen that allows the accomplishments of various approaches to be measured. Further, in defining the task we constrain the breadth of capabilities that must be developed before a practical contribution is attained and we establish a concrete foundation that can be referred to when design decisions are to be made.

This chapter provides details of our context-driven approach to machine vision. It describes the Condor architecture, gives details of the algorithms embedded within it, and provides an example of its application to natural object recognition.

The approach to machine vision that we have described is an attempt to overcome some of the fundamental limitations that have hindered progress in image understanding research. The ideas designed into that architecture embody a theory of computational vision for complex domains. To evaluate that theory, it is necessary to define a goal and to perform experiments that test how well the theory achieves that goal.

The key scientific question addressed here is the design of a computer vision system that can approach human-level performance in the interpretation of ground-level scenes of the natural world. Heretofore, no system has been constructed that demonstrates significant recognition competence in this domain and, worse, the field has not produced a theory about how such a system could be constructed. We offer a new paradigm for the design of computer vision systems that holds promise for achieving human-level competence, and report the experimental results of a system implementing that theory which demonstrates near-human recognition abilities in a natural domain of limited geographic extent.