Recognition of Visual Object Classes

Object recognition is both about recognizing speciic objects, e.g., \That is my dog Spot." and about recognizing classes of objects, e.g., \That is a dog." Our focus is on the latter problem, even though we do not ooer a precise deenition for what constitutes a class. In some cases, for example with human faces, the objects in a class are visually similar and form a visual object class. In other cases, say chairs, objects in the class may not look at all alike|the only similarities are in function. Recognition of functional object classes requires higher-level cognitive reasoning, we restrict here our attention to visual object classes. The main diiculty in object recognition is the problem of invariance. The pixel representation provided by the camera is dependent upon the lighting conditions, object pose, camera position, etc. Further, there is inherent variability between diierent instances from the same object class. Our approach to this problem is to model an object class as a set of local parts arranged in a deformable spatial connguration. For example, human faces consist of parts such as the eyes, nose, and mouth whose positions in the image plane vary with pose and expression as well as from individual to individual. The allowed object deformations are represented through shape statistics, which are learned from examples. Instances of an object in an image are detected by nding the appropriate features in the correct spatial connguration. The algorithm is robust with respect to partial occlusion, detector false alarms, and missed features. A 94% success rate is demonstrated for the problem of locating quasi-frontal views of faces in cluttered scenes.