Global effects in Figure/Ground segregation by
a model with only local interactions
with N. Rubin
Abstract
Figure/Ground segregation is a fundamental problem in visual
processing. Edges often arise because of occlusion, along the
bounding contour of the occluding object. Having detected an edge,
the visual system therefore has to decide whether it is there due to
occlusion, and if so, which side of it belongs to the front surface
(``the Figure''). This is known as the border ownership
problem. Determining the border ownership of an edge cannot be done
locally: it requires integrating information from an image region
which contains the entire Figure or large portions of it. Thus, a
neuron whose receptive field is small compared to the Figure cannot
compute border ownership of an edge of that Figure in isolation.
Recently it was reported that the responses of V2 cells to an edge can
be strongly modulated by the polarity of border ownership of the edge
(Zhou et al. 2000). These effects must therefore be the product of
computations done by a network of neurons. One possibility is that
such networks rely on direct long-range connections (of the scale of
the Figure) and/or feedback from areas with cells of large receptive
fields. The model presented here offers an alternative. It is shown
that network of small receptive field cells which interact only
locally can nevertheless compute Figure/Ground segregation. The
long-range effects emerge as a result of iterative propagation of
information through a cascade of short-range connections. The model
produces results similar to those observed perceptually: it prefers
enclosed, convex and/or smaller regions as the Figure. Our results
suggest that considerable portions of Figure/Ground segregation can be
accomplished by early visual cortex, without a need for feedback from
higher areas.
click here for the *.ps file of
the article.
click here for the *.pdf file of
the article.
Acknowledgments: We thank Davi Geiger, Jean-Michel Hupe, J. Anthony
Movshon, Ken Nakayama, Robert Shapley, Eitan Sharon, Shimon Ullman and
Yair Weiss for helpful discussions and comments on the manuscript. NR
is supported by NSF grant IBN-9720305 and an Alfred P. Sloan
Fellowship . MP is supported by NSF grants DMS97-21430 and
DMS99-71392 and an Alfred P. Sloan Fellowship.