\documentstyle[10pt,psadobe-kl]{article}

% original settings
%\setlength{\evensidemargin}{0.4cm}
%\setlength{\oddsidemargin}{0.4cm}
%\setlength{\textwidth}{6.25in}
%\setlength{\textheight}{9in}
%\setlength{\topmargin}{-1.7cm}
%\setlength{\headheight}{8.333pt}
%\setlength{\headsep}{0.5in}
     
\setlength{\evensidemargin}{0.2cm}
\setlength{\oddsidemargin}{0.2cm}
\setlength{\textwidth}{6.5in}
\setlength{\textheight}{9.2in}
\setlength{\topmargin}{-1.7cm}
\setlength{\headheight}{8.333pt}
\setlength{\headsep}{0.5in}
     
\setlength{\parskip}{\medskipamount}
     
\setlength{\itemsep}{0in}
\setlength{\unitlength}{1mm}
\parindent 0pt


%\input psadobe
%\ArborTextfalse

\newcommand{\psblock}[3] {
 \begin{minipage}[t] {#1}
 \begin{center}
  \psadobepercent{#2}{#3}   
 \end{center}
 \end{minipage}
}

\def\thebibliographya#1{\section*{Other References}\list
 {[\arabic{enumi}]}{\settowidth\labelwidth{[#1]}\leftmargin\labelwidth
 \advance\leftmargin\labelsep
 \usecounter{enumi}
}
 \def\newblock{\hskip .11em plus .33em minus .07em}
 \sloppy\clubpenalty4000\widowpenalty4000
 \sfcode`\.=1000\relax}
\let\endthebibliographya=\endlist

\newcounter{max}
\def\thebibliographyb#1{
 \setcounter{max}{\theenumi}
 \list
 {[\arabic{enumi}]}{\settowidth\labelwidth{[#1]}\leftmargin\labelwidth
 \advance\leftmargin\labelsep
 \usecounter{enumi}
}
 \setcounter{enumi}{\themax}
 \def\newblock{\hskip .11em plus .33em minus .07em}
 \sloppy\clubpenalty4000\widowpenalty4000
 \sfcode`\.=1000\relax}
\let\endthebibliographyb=\endlist

\pagestyle{empty}
\begin{document}
\begin{center}
{ \bf High-Level Perceptual Contours From A Variety Of
Low-Level Physical Features }\\[8mm]
Brian Scassellati\\
\end{center}
\vspace{0.4cm}

{\bf The Problem:}

Extraction of robust contour information from real-world images using
many different visual modalities such as disparity, luminance, color,
and texture.

{\bf Motivation:}

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.

{\bf Previous Work:}

The problem of contour detection has been focused on by a variety of
machine vision researchers.  The best known luminance-based contour
detectors arose from the work of John Canny \cite{canny}, while
spectral-based contour detection was the focus of Mollon
\cite{mollon}.  The psychophysical aspects of texture were first
explored by Julesz \cite{julesz}, and followed by the computational
work of Bergen and Landy \cite{bergen-landy}, Malik and Perona
\cite{malik-perona}, and Voorhees and Poggio \cite{voorhees-poggio}.
The use of disparity as a contour detector was first explored by
psychophysicists with the use of random-dot stereograms and later Marr,
Poggio, and Grimson investigated the computational aspects
\cite{marr-poggio-stereo,grimson-stereo}.

The psychophysics of contour identification has been explored for
centuries \cite{petry}.  However, the most relevant work to this
project concerns the work of Gaetano Kanizsa and the Gestalt
psychologists.  Kanizsa demonstrated a large class of illusory
contours that were psychologically highly salient, yet could not be
identified using luminance, color, or disparity information
\cite{kanizsa}.  The Kanizsa square and triangle are perhaps the most
widely known examples of these phenomena.

{\bf Approach:}

Our model of contour processing allows for the integration of
independent low-level edge detection modules (such as luminance and
texture) with other higher-level sources of information (such as
Gestalt effects and expectations). The central idea behind the
integrated contour model is the proposition that complex, high-level
perceptions can be created by using simple compositions of low-level
routines coupled with input from other high-level processes.

The model itself is relatively simple: allow for each of the low-level
edge detectors to operate independently, accept information from
higher level processes, and combine all of the information at a single
integration point. The integrated contour model also allows for the
results from the integration to be passed back to each of the
sub-modules. In this way, building a contour is an iterative process
that can rely on multiple stages of processing; that is, the
integration box does not perform one-shot processing on an input
image, but rather allows for successive refinements in the contour
image produced.

We are currently constructing an active vision platform to allow for
continuous real-time testing of the model presented above.  The system
will perform simple eye-tracking for identified contours.

% {\bf Data:}

%\begin{figure}
%\fbox{
%\begin{minipage}[b]{6.25in}
%\begin{center}
%\psadobepercent{100}{office_out.ps} \hfill
%\psadobepercent{100}{lab_out.ps}
%\caption{Some results from a face detection application implemented
%according to our general feature and pattern detection methodology.
%Each detected face is enclosed by a dotted box.}
%\label{fig:pictures}
%\end{center}
%\end{minipage}
%}
%\end{figure}



%% THE DATA:  
%% 
%% Include a page of representative pictures showing the heart of your
%% work.


%% {\bf Difficulty:}

{\bf Impact:}

The robust object segmentation available through this method will
allow for better continuous-time processing of visual input for
robotic systems.  The difficulties of dealing with real-world, noisy
images are mediated by the presence of multiple, overlapping detectors
that combine positively to create a single unified mapping.  Robot
navigation, automated contour processing for medical and military
applications, and automated visual searches for large image databases
should all profit from this approach.

% \bibliographystyle{HvdHK94}
{\bf References:}
\begin{thebibliographyb}{10}

\bibitem[1]{canny}
John~F. Canny.
\newblock Finding edges and lines in images.
\newblock Memo 720, Massachusetts Institute of Technology Artificial
  Intelligence Laboratory, Cambridge, Massachusetts, June 1983.

\bibitem[2]{mollon}
J.~D. Mollon and L.~T. Sharpe, editors.
\newblock {\em Color Vision: Physiology and Psychophysics}.
\newblock Academic Press, 1983.

\bibitem[3]{julesz}
Bela Julesz.
\newblock Experiments in the visual perception of texture.
\newblock {\em Scientific American}, 232:34--43, April 1975.

\bibitem[4]{bergen-landy}
M.~S. Landy and J.~R. Bergen.
\newblock Texture segregation and orientation gradient.
\newblock {\em Vision Research}, 31, 1991.

\bibitem[5]{malik-perona}
Jitendra Malik and Pietro Perona.
\newblock Preattentive texture discrimination with early vision mechanisms.
\newblock {\em Journal of the Optical Society of America}, 7, 1990.

\bibitem[6]{voorhees-poggio}
H.~Voorhees and T.~Poggio.
\newblock Computing texture boundaries from images.
\newblock {\em Nature}, 333, 1988.

\bibitem[7]{marr-poggio-stereo}
T.~Poggio and D.~Marr.
\newblock Cooperative computation of stereo disparity.
\newblock {\em Science}, 194:283--287, 1976.

\bibitem[8]{grimson-stereo}
W.E.L. Grimson.
\newblock {\em From Images to Surfaces}.
\newblock MIT Press, 1981.

\bibitem[9]{petry}
Susan Petry and Glenn~E. Meyer, editors.
\newblock {\em The Perception of Illusory Contours}.
\newblock Springer-Verlag, 1987.

\bibitem[10]{kanizsa}
Gaetano Kanizsa.
\newblock {\em Organization in Vision: Essays on Gestalt Perception}.
\newblock Praeger Publishers, 1979.

\end{thebibliographyb}

% \footnotesize
% \bibliography{/home/vi/sung/papers/references/vision}
% \bibliographystyle{plain}

\normalsize




\end{document}