Learning and Recognition In Natural Environments

Abstract

We present a system for learning descriptions of objects, and for subsequently recognizing learned objects, that functions in outdoor, natural environments. We describe in detail two modes of functioning within this system: (1) unguided, bottom-up learning of object descriptions directly from image data; (2) top-down recognition of objects whose approximate position and structure are known by use of image-level matching. We then argue that the system’s performance in these two modes indicates that robust outdoor performance can be achieved within the structure of our vision system.