Figure-Ground Organization in the Visual Cortex

A basic task of vision is the interpretation of 2D images in terms of objects in a 3D world.  We can understand the phenomenon of figure-ground organization as the product of neural mechanisms designed to deal with this fundamental problem.  Since V1 transforms the retinal pixel arrays into oriented-feature maps, making contrast borders explicit, the figure-ground problem translates to the task of relating contrast borders to object contours and assigning 'border ownership'. We found that neural signals in V2 code not only for position and local orientation of contrast borders, but also for the side of the figure to which a border belongs.  These signals reflect processing of global shape; displays that are locally identical, but differ in border ownership, produce different responses.  We do not yet understand the underlying mechanisms, but they appear to implement rules similar to the gestalt laws; for example, a region enclosed by a contour tends to become figure, junctions are split according to 'good continuation' etc.  In many cells, side-of-figure information is combined with selectivity for stereoscopic edges or selectivity for borders defined by dynamic occlusion.  These cells are selective for edge polarity (as given by stereo or motion cues), and the preferred 'near' side of the local edge is generally the same as the preferred figure side.  This shows that mechanisms using completely different computational strategies are combined in a specific manner:  The contrast borders are interpreted as occluding contours as if the figure were an object in 3D space.  Various observations point to ‘early’ mechanisms: Border ownership signals emerge with short latency; individual cells show a variety of cue combinations, as if different mechanisms are being combined in stages.

In conclusion, we are beginning to understand the function of area V2 as the stage of figure-ground segregation.  The gestalt phenomenon of border ownership can be traced to this representation.  V2 might provide the interface between local feature representations and the central processes involved in selective attention, search and recognition.