Redwood Neuroscience
Title "Perceptual Organization of
Occluding Contours”
Eric
Saund
Abstract:
An important class of visual
illusions show that configurations
of simple figural elements induce the perception of "illusory contours"---perceived
figure/ground boundaries where there are no illumination changes in the image
data. The classic example is the Kanizsa Triangle, which inspired the use of the Pac-Man in
visual psychophysics. Most investigators
now believe that Kanizsa-type illusory contours arise
through the visual system's attempts to sort out the relative depth relations
of scene surfaces. This talk will present
a computational model for this process.
The fundamental computational theory articulates constraints and biases
on the inference of physical surface overlap from local evidence consisting of L-type
and T-type junctions arising from local contrast edges. For any given scene, an energy or cost
function is constructed over interpretation labels for nodes of a sparse graph,
or belief net. Soft constraint
propagation techniques encourage the formation of consistent
interpretations. The model leads to
correct interpretations (in the sense of agreeing with human percepts) of popular
simple ``Colorforms'' figures known to induce
illusory contours, as well as more difficult figures where interpretations acknowledging
accidental alignment are preferred. I
will also sketch steps for extending this model to transparent and moving
surfaces. Toward these goals, I will
present an interactive (Java Applet) demonstration of interpreting surface
overlap and transparency at image X-junctions.