{\rtf1\mac\ansicpg10000\uc1 \deff4\deflang1033\deflangfe1033{\upr{\fonttbl{\f0\fnil\fcharset256\fprq2{\*\panose 02020603050405020304}Times New Roman{\*\falt Times};}{\f4\fnil\fcharset256\fprq2{\*\panose 02000500000000000000}Times;}
{\f47\froman\fcharset77\fprq0{\*\panose 00000000000000000000}Times-Bold{\*\falt Times};}}{\*\ud{\fonttbl{\f0\fnil\fcharset256\fprq2{\*\panose 02020603050405020304}Times New Roman{\*\falt Times};}
{\f4\fnil\fcharset256\fprq2{\*\panose 02000500000000000000}Times;}{\f47\froman\fcharset77\fprq0{\*\panose 00000000000000000000}Times-Bold{\*\falt Times};}}}}{\colortbl;\red0\green0\blue0;\red0\green0\blue255;\red0\green255\blue255;\red0\green255\blue0;
\red255\green0\blue255;\red255\green0\blue0;\red255\green255\blue0;\red255\green255\blue255;\red0\green0\blue128;\red0\green128\blue128;\red0\green128\blue0;\red128\green0\blue128;\red128\green0\blue0;\red128\green128\blue0;\red128\green128\blue128;
\red192\green192\blue192;}{\stylesheet{\widctlpar\adjustright \f4\cgrid \snext0 Normal;}{\s1\keepn\widctlpar\outlinelevel0\adjustright \b\f4\cgrid \sbasedon0 \snext0 heading 1;}{\*\cs10 \additive Default Paragraph Font;}{\s15\li720\widctlpar\adjustright 
\f4\cgrid \sbasedon0 \snext15 Body Text 2;}{\*\cs16 \additive \ul\cf2 \sbasedon10 Hyperlink;}{\s17\widctlpar\adjustright \i\cf1\loch\af4\hich\af4\dbch\f4\cgrid \sbasedon0 \snext17 Body Text;}}{\info{\title Notes on Texture Segmentation:}
{\author Marc Talusan}{\operator Marc Talusan}{\creatim\yr2003\mo10\dy2\hr17\min19}{\revtim\yr2003\mo10\dy20\hr15\min27}{\version6}{\edmins24}{\nofpages2}{\nofwords468}{\nofchars2671}{\*\company MIT}{\nofcharsws3280}{\vern115}}
\widowctrl\ftnbj\aenddoc\hyphcaps0\formshade\viewkind1\viewscale125\pgbrdrhead\pgbrdrfoot \fet0\sectd \linex0\endnhere\sectdefaultcl {\*\pnseclvl1\pnucrm\pnstart1\pnindent720\pnhang{\pntxta .}}{\*\pnseclvl2\pnucltr\pnstart1\pnindent720\pnhang{\pntxta .}}
{\*\pnseclvl3\pndec\pnstart1\pnindent720\pnhang{\pntxta .}}{\*\pnseclvl4\pnlcltr\pnstart1\pnindent720\pnhang{\pntxta )}}{\*\pnseclvl5\pndec\pnstart1\pnindent720\pnhang{\pntxtb (}{\pntxta )}}{\*\pnseclvl6\pnlcltr\pnstart1\pnindent720\pnhang{\pntxtb (}
{\pntxta )}}{\*\pnseclvl7\pnlcrm\pnstart1\pnindent720\pnhang{\pntxtb (}{\pntxta )}}{\*\pnseclvl8\pnlcltr\pnstart1\pnindent720\pnhang{\pntxtb (}{\pntxta )}}{\*\pnseclvl9\pnlcrm\pnstart1\pnindent720\pnhang{\pntxtb (}{\pntxta )}}\pard\plain \widctlpar
\tx560\tx1120\tx1680\tx2240\tx2800\tx3360\tx3920\tx4480\tx5040\tx5600\tx6160\tx6720\adjustright \f4\cgrid {\tab \tab \tab \tab \tab \tab \tab \tab \tab \tab \tab Notes by Philip Meier
\par \tab \tab \tab \tab \tab \tab \tab \tab \tab \tab \tab pmmeier@mit.edu
\par \tab \tab \tab \tab \tab \tab \tab \tab \tab \tab \tab 10-7-03
\par }\pard \widctlpar\adjustright {
\par }\pard \fi720\li1440\widctlpar\adjustright {\scaps\fs28 Texture Segmentation II
\par }\pard \widctlpar\adjustright {
\par }{
\par Filter based theories tend to be composed of
\par 
\par Linear filtering + some non-linearities + edge detection
\par 
\par Assumption:  \ldblquote Whatever a single texture is, the appropriate statistics are uniform all over it.\rdblquote 
\par 
\par Two relevant terms: Stationary and Ergodic.
\par }\pard \widctlpar\brdrb\brdrdot\brdrw60\brsp20 \adjustright {
\par }\pard \widctlpar\adjustright {
\par 
\par Difference of Guassians, half wave rectification, lateral inhibition
\par 
\par Why does one do a half wave rectification?}{ See Aaron\rquote s GUI.}{
\par 
\par How does the type of representation of a texture matter?  For example, }{d}{oes the representation have an explicit notion of local orientation?  Can it capture the linear extension of a texture of stalks of grass?
\par 
\par The question again is \ldblquote What types of statistics are used when humans discriminate or segment textures?\rdblquote  To determine discrimination, show two textures and ask, \ldblquote Are these the same?\rdblquote  T o determine segmentatio}{
n, show }{two textures side by side and ask, \ldblquote Where is the boundary between these two textures.\rdblquote  Interestingly, you ca
n discriminate textures more rapidly than you can segment them.  It takes longer to localize the boundary between to textures.  Possibly this observed time difference is the result of separate mechanisms/computations that are used for localizing boundarie
s after or parallel to having detected their presence\line 
\par \tab \tab \tab \tab \tab Segments pre-attentively?
\par Mean?  \tab \tab \tab \tab \tab Yes.\tab \tab 
\par Mean and variance?\tab \tab \tab \tab Yes.
\par Whole distribution of variance?\tab \tab }{ }{-
\par 
\par 
\par The Landy-Bergen Model of segmenting textures with varying orientations basically approximates the mathematical formula for statistical differences in orientation.  This suggests that there is not that much of a difference between \ldblquote 
filter based models\rdblquote  and \ldblquote statistical models.\rdblquote 
\par \line (I think the source is }{\f0\cf1 Landy, M. S., and Bergen, J. R. (1991). Texture segregation and orientation gradient. }{\i\f0\cf1 Vision Research 31}{\f0\cf1 , 679-691.)}{
\par 
\par We raised the point that the psychophysical task matters for the representation that you are interested in studying.
\par 
\par David Martin has written a paper arguing that \ldblquote the }{\f47 proper, explicit treatment of texture is required to detect boundaries in natural images.\rdblquote  
  Basically you can do a lot to find edges using color and luminance, but adding in textures is very important for natural scenes.
\par }{\field{\*\fldinst {\f47  HYPERLINK http://www.cs.berkeley.edu/~dmartin/papers/}{\cf1 pami-boundary.pdf}{\f47  }{\f47 {\*\datafield 
00d0c9ea79f9bace118c8200aa004ba90b02000000170000003d00000068007400740070003a002f002f007700770077002e00630073002e006200650072006b0065006c00650079002e006500640075002f007e0064006d0061007200740069006e002f007000610070006500720073002f00700061006d0069002d006200
6f0075006e0064006100720079002e007000640066000000e0c9ea79f9bace118c8200aa004ba90b7a00000068007400740070003a002f002f007700770077002e00630073002e006200650072006b0065006c00650079002e006500640075002f007e0064006d0061007200740069006e002f007000610070006500720073
002f00700061006d0069002d0062006f0075006e0064006100720079002e007000640066000000}}}{\fldrslt {\cs16\ul\cf2 http://www.cs.berkeley.edu/~dmartin/papers/pami-boundary.pdf}}}{\cf1 
\par 
\par Roland brought up the fact that there are certain \ldblquote texture\rdblquote  properties that are useful for finding object boundaries even in a Lambertian world of smooth objects with a single light source.  There
 reason for this is that there tend to be strong parallel and characteristic orientations in the region of an object near an edge. 
\par 
\par }\pard\plain \s17\widctlpar\adjustright \i\cf1\loch\af4\hich\af4\dbch\f4\cgrid {\ldblquote \hich\af4\dbch\af4\loch\f4 Its not just tiger stripped p\hich\af4\dbch\af4\loch\f4 otatoes that benefit from segmentation by texture\hich\af4\dbch\af4\loch\f4 ; wh
\hich\af4\dbch\af4\loch\f4 ite ones benefit t\hich\af4\dbch\af4\loch\f4 oo!\rdblquote 
\par }\pard\plain \widctlpar\adjustright \f4\cgrid {\cf1 
\par Adelson summarized the converging methods for investigating vision.
\par 
\par X = some visual task, such as segmenting an image of the world into objects.
\par 
\par Psychophysics:\tab How good are people at X?
\par Physiology\tab \tab What are the mechanisms that implement X?
\par Pure Statistics\tab \tab How much information is available to an ideal observer doing X?
\par Ecological Optics\tab How are statistics of a visual environment exploitable for X? 
\par 
\par }{\f47 
\par 
\par }{
\par }{
\par }}