Must Figure-Ground Organization Precede Object Recognition? An Assumption in Peril

Abstract

The assumption that figure-ground segmentation ml/st precede object or shape recognition has been celltral to theories ofvisual perception. We showed that assumption to be incorrect in an experiment in whiciz observers reported the first perceivedfigure-ground organization ofbriefly exposed stimuli depicting two regions sharing a figure-ground border. We manipulated the symmetry ofthe two regions and their orientationdependellt denotivity (roughly, their meaningfulness), and measured how each of these variables influenced figure-ground reports when the stimuli were exposedfor 14, 28, 57, 86, or 100 ms, andfollowed immediately by a mask. Influences onfigureground organization from both sym"metry and orientationdependent object recognition processes were found; both were observed first in the 28-ms condition. Object recognition inputs did not dominate symmetry inputs. We suggest that object recognition processes may operate simultaneously on both sides of edges detected before figure-ground relationships are determined.Figure-ground organization entails segmenting the visual field into regions that are figures and regions that are backgrounds. Figures are patently endowed with shape and appear to occlude backgrounds; the latter appear to be shapeless near the borders they share with figures. Since the early 20th century, most theories of visual perception and object recognition have been grounded on the assumption that regions must be registered as figures before they can be matched to representations of objects in memory (Biederman

Graded effects in hierarchical figure-ground organization: Reply to Peterson (1999).

“…Under this interpretation of our model, figure-ground processing does not occur prior to objectrecognition processes because figure-ground organization is attempted but fails. Therefore, according to Peterson (1999), our model violates the "figure-ground-first" assumption that would be required to successfully argue against Peterson's (1994) prefigural account of figme-ground segregation in favor of a hierarchical account. A consequence of the ambiguity assumption according to Peterson (1999) is that our interactive model would be unable to explain some of the behavioral data from Peterson and colleagues (e.g., Peterson & Gibson, 1993, 1994.…”

“…Under Peterson's approach, object representations initially receive input from luminance contours or edges, not from figure-ground processes (as would be required under other models; e.g., Biederman, 1987;Kosslyn, 1987;Marl 1982). The input from these edge-based processes allows objects to be recognized before any figure-ground processing has occurred; thus, these object-recoguition processes are termed "prefigural" recognition processes (Peterson, 1994). The outputs of the recognition process can then influence figure-ground organization, allowing more familiar, or high-denotative, regions to have a higher probability of being labeled as figure.…”

“…The outputs of the recognition process can then influence figure-ground organization, allowing more familiar, or high-denotative, regions to have a higher probability of being labeled as figure. We contend that Peterson's (1994Peterson's ( , 1999 account is a feedforward model of visual processing in which (a) edges are extracted from the visual image, (b) prefigural object recognition occurs based on edge information, and (c) figtue-ground organization is influenced by the outputs of the prefigural recognition processes. This position is corroborated by the presence of Pathway D in Figure 5 from Peterson's (1999) commentary (equivalent to the luminance contour pathway depicted in Figure 8a in Vecera & O'Reilly, 1998), which illustrates the prefigural recognition process.…”

“…The critical question, and the question at the heart of the present exchange, is how object-level information influences figure-ground organization. Peterson's (1994;Peterson & Gibson, 1993, 1994) approach as to how object-level information influences flgure~ground segregation was to reject the logical requirement that figure--ground precedes object identification. Under Peterson's approach, object representations initially receive input from luminance contours or edges, not from figure-ground processes (as would be required under other models; e.g., Biederman, 1987;Kosslyn, 1987;Marl 1982).…”

Simultaneous Segmentation and Figure/Ground Organization Using Angular Embedding

“…Hoiem et al [15] fix an initial oversegmentation and iterate region-merging and depth estimation steps. It is not yet known where figure/ground discrimination occurs in biological visual systems [22], with, as noted by Ren et al [24], some evidence for early availability of a contour ownership signal [34].…”