Abstract

One of the major functions of intermediate-level vision is object discrimination—grouping stimuli into “things.” We propose a model in which the outputs of early visual processing are first grouped into proto-objects, which are then linked into objects. Proto-objects are defined as bounded, simply connected, and spatially continuous surfaces with associated surface attributes such as depth, color, and texture. We have constructed a physiologically based neural network implementation of this model that discriminates objects based on the extraction of depth from occlusion. We provide detailed mechanisms for binding edges and lines into contours, segementing contours of different objects, binding surface attributes to surfaces, and binding contours and surfaces together. Together, these processes identify proto-objects in the scene. The representation of a proto-object is distributed across multiple visual maps. Proto-objects are discriminated by strictly feedforward processing, but they are linked together to form objects by reentrant feedback from higher level areas. The linking processes depend upon both surface cues (depth) and contour cues (good continuation). Extensive simulations demonstrate that responses of the model are similar to human psychophysical responses to both real and illusory stimuli.

© 1992 Optical Society of America