From: las@ai.mit.edu (Lynn Andrea Stein)
Date: Thu, 19 Oct 1995 18:03:02 -0400
Message-Id: <199510192203.SAA03118@puffies.ai.mit.edu>
To: Marney Smyth <marney@ai.mit.edu>
Cc: Brian Scassellati <scaz@ai.mit.edu>
Subject: forwarded message from Brian Scassellati

Marney -- Here is one project description.  I am working on another,
but the relevant student seems to have vanished.  If you don't hear
from me tomorrow, it's not coming.

                                             Lynn

Brian Scassellati:  High-Level Perceptual Contours From A Variety Of
Low-Level Physical Features.


For years, researchers in machine vision have focused on extracting
object boundary information from luminance derivatives, color
contrast, depth information, and textural patterns.  However, no
single one of these methods is sufficient for detecting all of the
contour information that humans perceive.  Humans have little
difficulty discriminating depth-based contours from the fused images
in a stereogram, extracting illusory contours from a Kanizsa square,
or even in completing contours through the blind spot.

This project reformulates the problem of ``edge detection'' into the
recognition of high-level perceptual features which we call
``contours''.  It involves the development of a novel visual
processing framework that deals with contours as high-level image
features.  We will demonstrate the operation of this new model on
real-world images using a prototype active vision head designed to
support this visual architecture.


Additional avenues of research can be split into three categories:
additional components to our implementation of contour integration,
additional high-level behavioral goals to utilize this system, and
evaluation of this methodology on other aspects of vision.

There are many additions that should be made to the subset of the
integrated contour model implemented.  The addition of depth will
provide the first real utilization of the unique qualities of the
active vision platform.  Depth also provides a strong means of contour
information that has not yet been well exploited.  Another interesting
addition would be a module to implement the filling-in effects
demonstrated by the blind spot.  The filling-in effects may be helpful
in providing information for occluded objects, or even in integrating
information between the wide-angle and foveated cameras.  Motion would
be an excellent addition for rapid discrimination, and also aid in the
development of processing that was not pipeline-based.  Additional
experiments should also focus on the use of multiple C40 boards for
parallel computation of the low-level modules.  The use of feedback in
the integrated contour model would allow for investigations of
real-time processing and of the interactions between modalities.
Additional high-level procedures, like a model of shape recognition,
would allow the study of how expectations can influence the low-level
modules.  Finally, the entire integrated contour software should be
ported to run on Cog to make use of the richer variety of sensory
inputs and behavioral responses.

Additional behavioral goals should also be investigated.  Contour
segmentation is only a partial test of the integrated contour model.
A high-level behavioral goal, like object tracking or recognition, or
learning to follow human faces, would give a better sense of
completeness to this project.  By specifying other unique behavioral
goals, the relative importance and influence of each modality can be
more clearly studied.  Learning algorithms for the weightings between
modalities would need to be researched and developed.  Future
behaviors would also allow the use of the active vision platform as
more than a static camera base.

Finally, it would be interesting to apply the methodology used in
constructing a model of perceptual contours to other areas of machine
vision and artificial intelligence.  This type of modeling could be
useful for other machine vision tasks, such as navigation or visual
planning for object manipulation.  The methodology of differentiating
carefully between physical and perceptual stimuli could also be
applied to other senses, such as audition or tactile sensation.  The
technique of using many simple modules for a common problem may
also be useful to the study of other high-level mental tasks such as
memory retrieval or path planning, but the direct extension of this
technique is not clear at this time.