Item

Abstract

When a camouflaged animal sits in front of the appropriate background, the animal is
effectively invisible. As soon as the animal moves, however, it is easily visible despite
the fact that at any given instant, there is no shape information. This process, referred to
as Spatiotemporal Boundary Formation (SBF), can be initiated by a wide range of texture
transformations, including changes in the visibility, shape, or color of individual texture
elements. Shipley and colleagues have gathered a wealth of psychophysical data on SBF,
and have presented a local motion vector model for the recovery of the orientation of
local edge segments (LESs) from as few as three element changes (Shipley and Kellman,
1997). Here, we improve and extend this model to cover the extraction of global form
and motion.
The model recovers the orientation of the LESs from a dataset consisting of the relative
spatiotemporal location of the element changes. The recovered orientations of as few as
two LESs is then be used to extract the global motion, which is then used to determine the
relative spatiotemporal location and minimal length of the LESs. To complete the global
form, the LESs are connected in a manner similar to that used in illusory contours. Unlike
Shipley and Kellman’s earlier model, which required that pairs of element changes be
represented as local motion vectors, the present model merely encodes the relative spatiotemporal
locations of the changes in any arbitrary coordinate system.
Computational simulations of the model show that it captures the major psychophysical
aspects of SBF, including a dependency on the spatiotemporal density of element
changes and a sensitivity to spurious changes. Interestingly, the relative encoding scheme
yields several emergent properties that are strikingly similar to the perception of aperture
viewed figures (Anorthoscopic Perception).
The model captures many of the important qaulities of SBF, and offers a framework
within which additional aspects of SBF may be modelled. Moreover, the relative encoding
approach seems to inherently encapsulate other phenomenon, offering the possibility
of unifying several phenomena within a single mathematical model.