/[gzz]/manuscripts/Paper/paper.tex
ViewVC logotype

Diff of /manuscripts/Paper/paper.tex

Parent Directory Parent Directory | Revision Log Revision Log | View Patch Patch

revision 1.34 by tjl, Wed Mar 19 10:25:58 2003 UTC revision 1.35 by jvk, Thu Mar 20 11:52:29 2003 UTC
# Line 1045  XXX: http://www.opticsexpress.org/abstra Line 1045  XXX: http://www.opticsexpress.org/abstra
1045  %  subjective preference for text displays and empirical measures of their  %  subjective preference for text displays and empirical measures of their
1046  %  readability."  %  readability."
1047    
1048    It can be argued that
1049    the backgrounds clutter the display
1050    visually, making the user interface more confusing,
1051    and reduce text readability.
1052    Indeed, one of the most difficult aspects of the work was making the random
1053    color selection produce acceptable results.
1054    However, by tuning the color selection and the gamma
1055    correction of the display, we were able to (in our opinion) avoid
1056    the above problems.
1057    It is important that the colors chosen are light and that the palettes
1058    have a relatively small range of colors.
1059    Also, text readability on the generated backgrounds depends
1060    strongly on the text scale; making it easy for the user to zoom
1061    fluidly in and out helps.
1062    
1063    We have not tried to attain infinite zoomability\cite{furnas00infinity}
1064    with the current implementation,
1065    but only the more modest goal of zooming within a range
1066    that would be reasonable for a single PDF document,
1067    i.e., approximately 100-fold
1068    difference between mininum and maximum zoom.
1069    % This is in line with our
1070    % intended use, since it would be unreasonable
1071    % to expect a texture to be recognizable
1072    % if a sub-pixel area were zoomed to the full screen.
1073    It could be possible\cite{furnas00infinity}
1074    to make the unique background look similar
1075    at different scales, but
1076    this would remove the use of the texture as a cue of scale.
1077    Our
1078    nonlinear use of the register combiners
1079    does have some ill effects when zooming the texture out
1080    to a very small scale: mipmapping will not give the correct
1081    average color value.
1082    It may be possible to alleviate this by modeling the texture mathematically
1083    and calculating the correct average and placing corrective terms to the
1084    equations.
1085    However, in the intended zooming range
1086    the current system is quite satisfactory.
1087    
1088    % It is important that the background can be zoomed
1089    % to different resolutions.
1090    
1091    % leads to aliasing: .... modeling textures mathematically , ...
1092    
1093  TJL  TJL
1094    
1095  The most commonly asked question about this work concerns  The most commonly asked question about this work concerns
# Line 1070  On a darker background, this approach wo Line 1115  On a darker background, this approach wo
1115    
1116  JVK  JVK
1117    
 There has been  
 lot of texture perception work on texture discrimination.  
 However, in our application texture discrimination is not as  
 much of an issue as memorizability and recognizability of  
 previously seen textures.  
 Furthermore, our textures are on a higher level,  
 more like complete pictures than the usually studied microstructure.  
   
   
 XXX  
   
 %\subsection{Texture perception}  
   
1118  Psychophysical studies on texture perception have mostly concentrated  Psychophysical studies on texture perception have mostly concentrated
1119  on \emph{texture discrimination}, the ability of human observers to  on \emph{texture discrimination}\cite{julesz62visualpattern},
1120  discriminate pairs of textures.    the ability of human observers to discriminate pairs of textures.  
1121  The term is often used interchangably with \emph{texture segregation},  %The term is often used interchangably with \emph{texture segregation},
1122  the more specific task of finding the border between differently textured  %the more specific task of finding the border between differently textured
1123  areas (different phases of local characteristics at the  %areas (different phases of local characteristics at the
1124  border can segregate otherwise indiscriminable textures).  %border can segregate otherwise indiscriminable textures).
1125    %
1126  First experiments on computer-generated, unnatural textures in the 60s  %First experiments on computer-generated, unnatural textures in the 60s
1127  \cite{julesz62visualpattern} led to proposals of discrimination models  %\cite{julesz62visualpattern} led to proposals of discrimination models
1128  based on the $N$th-order statistics of textures  %based on the $N$th-order statistics of textures
1129    %(the joint distributions of the values at the corners of a randomly
1130    %placed (translated) $N$-gon for all different $N$-gons).
1131    %%and connectivity structures of certain micropatterns.
1132    %
1133    First discrimination models were based
1134    on the $N$th-order statistics of textures
1135  (the joint distributions of the values at the corners of a randomly  (the joint distributions of the values at the corners of a randomly
1136  placed (translated) $N$-gon for all different $N$-gons).  placed (translated) $N$-gon for all different $N$-gons).
1137  %and connectivity structures of certain micropatterns.  However, the order of similarity in the statistics did not
1138    consistently explain discrimination performance, and certain
1139    distinctive local features were conjectured.
1140    
1141  Attempt to explain texture discrimination by the densities of textons  Julesz\cite{julesz81textons} proposed that discrimination could be explained
1142  \cite{julesz81textons}, fundamental texture elements, such as  by the densities of textons, fundamental texture elements, such as
1143  elongated blobs, line terminators, line crossings, etc.    elongated blobs, line terminators, line crossings, etc.  
1144  However, the textons are hard to define formally.  However, the textons are hard to define formally.
1145    
1146  Much simpler filtering-based models can explain texture discrimination  Much simpler filtering-based models can explain texture discrimination
1147  just as well \cite{bergen88earlyvision}.  just as well \cite{bergen88earlyvision}.
1148  Essentially a bank of linear filters is applied to the texture followed  In this approach, a bank of linear filters is applied to the texture followed
1149  by a nonlinearity and then another set of filters.  by a nonlinearity and then another set of filters to extract densities
1150  In \cite{heeger95pyramid}, new textures with appearance similar  of features (see, e.g., \cite{heeger95pyramid} for an application).
1151  to a given texture are created by matching certain histograms  %In \cite{heeger95pyramid}, new textures with appearance similar
1152  of filter responses.  %to a given texture are created by matching certain histograms
1153    %of filter responses.
1154    
1155  Mapping texture appearance to an Euclidian texture space  In our application texture discrimination is not as
1156  (see \cite{gurnsey01texturespace} and the references therein):  much of an issue as memorizability and recognizability of
1157  in the reported experiments, three dimensions have been sufficient  previously seen textures.
1158  to explain most of the variation in the similarity judgements for  Furthermore, in most texture perception work
1159  artificial textures.  texture is considered as the output of a stochastic
 However, the texture stimuli have been somewhat simple  
 (no color, lack of frequency-band interaction, etc.).  
 For some natural texture sets (see, e.g., \cite{rao96texturenaming}),  
 three dimensions have also been  
 sufficient, but often semantic connections cause the  
 similarity to be context-dependant, making it hard to assess the  
 dimensionality.  
 % XXX: this is something we should experiment with our textures  
   
 XXX: reviews  
   
 XXX: physiological knowledge of visual perception  
 (see, e.g.,~Bruce et al\cite{bruce96visualperception});  
   
 XXX: in most work, texture is considered as the output of a stochastic  
1160  process that produces certain repeating features.  process that produces certain repeating features.
 Different samples from the process are considered as the same texture.  
1161  The textures created by our algorithm, although repeating, are more like  The textures created by our algorithm, although repeating, are more like
1162  complete images rather than microstructure.  complete images than statistical microstructure.
1163  Therefore, higher level processes of vision are also involved  Therefore, higher level processes of vision are also involved
1164  in the perception and recognition.  in the perception and recognition.
1165    
 theories of structural object perception  
 (see, e.g., Biederman\cite{biederman87})  
   
   
 XXX  
   
1166  Experiments on black-and-white %(faces,)  Experiments on black-and-white %(faces,)
1167  ink blots, and snow crystals  ink blots, and snow crystals
1168  \cite{goldstein71visualrecognition} show that  \cite{goldstein71visualrecognition} show that
# Line 1182  also easily be associated with the docum Line 1201  also easily be associated with the docum
1201    
1202  XXX: refs?  XXX: refs?
1203    
1204  XXX: texture set dimensionality studies?  Mapping texture appearance to an Euclidian texture space
1205    (see \cite{gurnsey01texturespace} and the references therein):
1206    in the reported experiments, three dimensions have been sufficient
1207    to explain most of the variation in the similarity judgements for
1208    artificial textures.
1209    However, the texture stimuli have been somewhat simple
1210    (no color, lack of frequency-band interaction, etc.).
1211    For some natural texture sets (see, e.g., \cite{rao96texturenaming}),
1212    three dimensions have also been
1213    sufficient, but often semantic connections cause the
1214    similarity to be context-dependant, making it hard to assess the
1215    dimensionality.
1216    % XXX: this is something we should experiment with our textures
1217    
1218    XXX: reviews
1219    
1220    XXX: physiological knowledge of visual perception
1221    (see, e.g.,~Bruce et al\cite{bruce96visualperception});
1222    
1223    theories of structural object perception
1224    (see, e.g., Biederman\cite{biederman87})
1225    
1226    
1227    XXX
1228    
1229  %\section{Software availability}  %\section{Software availability}
1230    
# Line 1228  amount of unique backgrounds cheaply, ma Line 1270  amount of unique backgrounds cheaply, ma
1270  to give, e.g., all academic articles stored on a user's hard  to give, e.g., all academic articles stored on a user's hard
1271  drive their own background.  drive their own background.
1272    
1273  \subsection{Problems}  %\subsection{Further work}
   
 It can be argued that  
 the backgrounds clutter the display  
 visually, making the user interface more confusing,  
 and reduce text readability.  
 Indeed, one of the most difficult aspects of the work was making the random  
 color selection produce acceptable results.  
 However, by tuning the color selection and the gamma  
 correction of the display, we were able to (in our opinion) avoid  
 the above problems.  
 It is important that the colors chosen are light and that the palettes  
 have a relatively small range of colors.  
 Also, text readability on the generated backgrounds depends  
 strongly on the text scale; making it easy for the user to zoom  
 fluidly in and out helps.  
   
 We have not tried to attain infinite zoomability\cite{furnas00infinity}  
 with the current implementation,  
 but only the more modest goal of zooming within a range  
 that would be reasonable for a single PDF document,  
 i.e., approximately 100-fold  
 difference between mininum and maximum zoom.  
 % This is in line with our  
 % intended use, since it would be unreasonable  
 % to expect a texture to be recognizable  
 % if a sub-pixel area were zoomed to the full screen.  
 It could be possible\cite{furnas00infinity}  
 to make the unique background look similar  
 at different scales, but  
 this would remove the use of the texture as a cue of scale.  
 Our  
 nonlinear use of the register combiners  
 does have some ill effects when zooming the texture out  
 to a very small scale: mipmapping will not give the correct  
 average color value.  
 It may be possible to alleviate this by modeling the texture mathematically  
 and calculating the correct average and placing corrective terms to the  
 equations.  
 However, in the intended zooming range  
 the current system is quite satisfactory.  
   
 % It is important that the background can be zoomed  
 % to different resolutions.  
   
 % leads to aliasing: .... modeling textures mathematically , ...  
   
 \subsection{Further work}  
1274    
1275  So far, we have concentrated mostly on low-end hardware, and  So far, we have concentrated mostly on low-end hardware, and
1276  have not even tapped the full potential of the NV25 architecture.  have not even tapped the full potential of the NV25 architecture.

Legend:
Removed from v.1.34  
changed lines
  Added in v.1.35

savannah-hackers-public@gnu.org
ViewVC Help
Powered by ViewVC 1.1.26