![Choosing Effective Colours for Data Visualization
Christopher G. Healey1
Department of Computer Science, University of British Columbia
ABSTRACT
In this paper we describe a technique for choosing multiple colours
for use during data visualization. Our goal is a systematic method
for maximizing the total number of colours available for use, while
still allowing an observer to rapidly and accurately search a display
for any one of the given colours. Previous research suggests that
we need to consider three separate effects during colour selection:
colour distance, linear separation, and colour category. We describe
a simple method for measuring and controlling all of these effects.
Our method was tested by performing a set of target identification
studies; we analysed the ability of thirty-eight observers to find a
colour target in displays that contained differently coloured back-
ground elements. Results showedour method canbe usedto selecta
group of colours that will provide good differentiation between data
elements during data visualization.
CR Descriptors: H.5.2 [Information Interfaces and Presen-
tation]: User Interfaces - ergonomics, screen design (graphics,
colour); I.3.6 [Computer Graphics]: Methodology and Tech-
niques - ergonomics,interaction techniques
1 INTRODUCTION
Scientific visualization in computer graphics is a rapidly expanding
area of research. This is due in large part to the dramatic increase
in both the size and the number of datasets that need to be visu-
alized [5, 17]. To date, many application-specific tools have been
built to help analyse individual datasets. Much less work has fo-
cused on developing guidelines for the design of visualization tech-
niques [16]. Our work is intended to address one aspectof this more
general question.
A typical method of visualizing a dataset involves mapping data
attributes to visual features (e.g., shape, size, spatial location, and
orientation). Colour is an important and frequently-used feature.
Examples include colour temperature gradients on maps and charts,
colour-coded vector fields in flow visualization, or colour icons dis-
played by real-time simulation systems.
If we use colour to represent our data, an important question to
ask is: “How can we chooseeffective colours that provide good dif-
ferentiation between data elements during the visualization task?”
We addressthis problem by trying to answerthree related questions:
How can we allow rapid and accurateidentification of individ-
ual data elements through the use of colour?
What factors determine whether a “target” element’s colour
will make it easy to find, relative to differently coloured “non-
target” elements?
How many colours can we display at once, while still allowing
for rapid and accurate target identification?
1Departmentof ComputerScience,University of British Columbia, 2366
Main Mall, Vancouver, British Columbia, V6T 1Z2, Canada
Previous work has addressed the issue of choosing colours for
certain types of data visualization. For example, Ware and Beatty
describe a simple colour visualization technique for displaying cor-
relation in a five-dimensional dataset [21]. Research on the design
of military systems sometimes quotes anecdotal evidence that sug-
gests “:::the generalguideline for computer-generated images is no
more than five to seven colours at a time:::” [10], although they of-
fer no explanation for why this might be the case. Robertson, Ware,
Rheingans and Tebbs, and Levkowitz and Herman discuss vari-
ous methods for building effective colour gamuts and colourmaps
[14, 20, 13, 12]. Recent work at the IBM Thomas J. Watson Re-
search Center has focused on a rule-based visualization tool that
considers how a user perceives visual features like hue, luminance,
height, and so on [15, 2].
None of these techniques were designed to investigate the rapid
and accurate identification of individual data elements based on
colour. Also, since the colour gamut and colourmap work uses con-
tinuous colour scales to encodeinformation, they do not addressthe
question of how many colours we can effectively display at once,
while still providing gooddifferentiation betweenindividualdata el-
ements.
An intuitive first step to gaining more control over colour would
be to use a perceptualcolourmodellike CIE LUV, CIE Lab,or Mun-
sell [22]. Unfortunately, fixing the colour distance in a perceptual
colourmodel to a constantvalue doesnot guarantee that each colour
will be equally easy to detect. Other factors can affect how groups
of coloured elements interact with one another.
Our technique does, in fact, use the CIE LUV colour model to
provide control over colour distance and isoluminance. We also ex-
ploit two specific results related to colour target detection: linear
separation [6, 1] and colour category [11]. These effects are con-
trolled to allow for the rapid and accurate identification of colour
targets. Target identification is a necessary first step towards per-
forming other types of exploratory data analysis. If we can rapidly
and accurately differentiate elements based on their colour, we can
apply our results to other important visualization techniqueslike de-
tection of data boundaries, the tracking of data regions in real-time,
and enumeration tasks like counting and estimation [19, 18, 8].
2 BACKGROUND
Before we describe our technique in detail, we provide a brief
overview of the CIE LUV colour model, as well as a description of
the linear separation and colour category effects.
2.1 CIE LUV
The CIE LUV colour model was proposed by the Commission In-
ternationale de L’Èclairge (CIE) in 1976 [22]. Colours are speci-
fied using the three dimensions L (which encodes luminance), u,
and v
(which together encode chromaticity). CIE LUV provides
two useful properties for controlling perceived colour difference.
First, colours with the same L
are isoluminant. Second, Euclidean
distance and perceived colour difference (specified in E units)
can be interchanged,since the colourdifference between two colour](https://image.slidesharecdn.com/viz-220418054648/85/Choosing-Effective-Colours-for-Data-Visualization-1-320.jpg)
![stimuli x and y is roughly:
E
xy =
p(L
xy)2 +(u
xy)2 +(v
xy)2 (1)
2.2 Linear Separation
The linear separation effect was originally described by D’Zmura
[1991] . He was investigating how the human visual system finds
a target colour in a sea of background non-target colours. D’Zmura
ran experiments that asked observers to determine the presence or
absence of an orange target. Two groups of differently coloured
non-target elements were also present in each display (e.g., in one
experiment half the non-targets in each display were coloured green
and half were coloured red). Results showed that when the target
could be separated by a straight line from its non-targets in colour
space(Figure 1, target T and non-targets A and C), the time required
to determine the target’s presence or absence was constant, and in-
dependent of the total number of elements being displayed. This
suggests detection occurs in parallel in the low-level visual system.
When the target was collinear with its non-targets (Figure 1, target
T and non-targets A and B), the time required to identify the tar-
get was linearly proportional to the number of elements being dis-
played. Observers had to search serially through each display to de-
termine whether the target was present or absent.
T
B
A
C
Figure 1: Example of a target with both collinear and separablenon-
targets, the colours are shown in a u;v-slice from the CIE LUV
colour model, notice that target T is equidistant from all three non-
targets A, B, and C; in Case 1, target colour T is collinear with the
non-target colours A and B; in Case 2, target T is linearly separable
from its non-targets A and C
Work by Bauer [1996] provided a number of results that
strengthen D’Zmura’s hypothesis. First, he showed that perceptual
colour models cannot be used to overcome linear separation. In his
experiments the target–non-target distance was fixed to a constant
value in CIE LUV (Figure 1). In spite of this, the target required
serial searching when it was collinear with its non-targets, but was
significantlyeasierto detectwhenit wasseparable. Bauerreplicated
his initial findings in three additional colour regions: green, blue,
and green-yellow. This suggestslinear separation applies to colours
from different parts of the visible colour domain.
2.3 Colour Category
Kawai [1995] reported results that suggest that the time required to
identify a colour target depends in part on the named colour regions
occupiedby the target and its non-targets. If target identification de-
pended only on the colour distance between a target and its non-
targets, the time taken to detect presence or absence of the target
should decrease uniformly as target–non-target distance increases.
Kawai ran a set of experiments that showed this was not true.
Kawai tested displays that contained a uniquely colour target and
a constant number of uniformly coloured non-targets. He divided
blue
boundary
purple
boundary
green
boundary
T
A
B g-axis
j-axis
Figure 2: A target T and two non-targets A and B shown in a (j;g)
slice from the OSA colour model; the boundariesof the blue, green,
and purple colour regions are shown as thick lines; T and B occupy
the same named colour region “blue”, but A does not
an isoluminant colour slice used during his experiments into named
colour regions. His results showed that search times decreased dra-
matically whenever the non-target was moved outside the target’s
colour region. For example, finding a target coloured T in a set
of non-targets coloured B was significantly more difficult than than
finding T in a set of non-targets coloured A (Figure 2). The target–
non-target distances TAand TB are equal; Kawai suggeststhe dif-
ference in performance is due to the fact that both T and B are lo-
cated in the blue colour region, but A is not. Similar search time
asymmetries for separate experiments testing green, purple, and red
targets were also explained by a difference in colour regions.
3 COLOUR SELECTION TECHNIQUE
Results discussed in the background section suggest that choosing
effective colours for data visualization dependson at least three sep-
arate criteria:
colour distance: the Euclidean distance between different
colours as measured in a perceptually balanced colour model
linear separation: the ability to linearly separate targets from
non-targets in the colour model being used
colour category: the named colour regions occupied by both
the target and non-target elements
We began our investigation by controlling colour distance and
linear separation, but not colour category. We proceeded under the
assumption that the user might choose to search for any one of the
available data elements at any given time. This is typical during ex-
ploratory data analysis; users will often change the focus of their in-
vestigation based on the data they see as the visualization unfolds.
Our requirement meant that the colourselection techniquehad to al-
low for rapid and accurateidentificationof anyof the elementsbeing
displayed.
We ran four studies to investigate the tradeoff between the num-
ber of colours displayed and the time required to determine the pres-
enceor absenceof a target element. Eachstudy displayeda different
number of unique colours, which we identify using the names red,
yellow-red, yellow, green-yellow, green, blue-green, blue, purple-
blue, purple, and red-purple (or R, YR, Y, GY, G, BG, B, PB, P, and
RP):
three-colour study: each display contained three different
colours (i.e., one colour for the target and two for the non-
targets) which we named R, GY, and PB](https://image.slidesharecdn.com/viz-220418054648/85/Choosing-Effective-Colours-for-Data-Visualization-2-320.jpg)
![five-colour study: each display contained five different
colours: R, Y, GY, B, and P
seven-colour study: each display contained seven different
colours: R, Y, GY, G, BG, P, and RP
nine-colour study: each display contained nine different
colours: R, YR, Y, GY, G, BG, PB, P, and RP
Every colourin a given study was tested as a target. For example,
the three-colour study was run three times, first with an R element
acting as a target (and GY and PB elements acting as non-targets),
next with a GY target (and R and PB non-targets), and finally with a
PB target (and R and GY non-targets). Faster search times for cer-
tain targets would have to be explained in terms of colour distance,
linear separation, or colour category. Each of the four studies were
themselves divided in the following manner:
elements in each display were drawn as coloured squares, and
were randomlylocatedon an underlying99grid thatcovered
the entire viewing area of the monitor
half of the displays were randomly chosen to contain an ele-
ment that used the target colour; the other half did not
one-third of the displays contained a total of 17 elements (one
target and 16 non-targets if the target was present, or 17 non-
targets if the target was absent); one-third of the displays con-
tained a total of 33 elements; one-third of the displays con-
tained a total of 49 elements
Colours for each of the four studies were chosen such that the
colour distance between pairs of colours and the linear separation
for each colour were fixed to constant values. Our results showed
that detection was rapid and accurate for all colours from both the
three-colour and five-colour studies. Results from the seven-colour
and nine-colour studies were mixed; some colours gave better per-
formance than others. This difference was explained when we ex-
amined the colour regions occupied by each of the colours.
3.1 Method
We began the colour selection process by obtaining the chromatic-
ities of our monitor’s triads. We also measured the luminance of
the monitor’s maximum intensity red, green, and blue with a spot
photometer. These values were needed to convert colours from
CIE LUV into the monitor’s RGB gamut.
We wanted to ensure that the colours we chose had the same per-
ceived intensity. Previous research has suggested that random vari-
ation in intensity can interfere with an observer’s ability to perform
visualization tasks based on colour [4]. In order to guarantee iso-
luminance, all the colours were chosen from a single u;v-slice
through the CIE LUV colour space at L = 67:1. We wanted to
maximize the number of available colours, while still maintaining
control over colour distance and linear separability. To do this, we
computed the boundary of the monitor’s gamut in the L = 67:1
slice. We then found the largest circle inscribed within the gamut
(Figure 3a).
Given the maximum inscribed circle, we chose colours that were
equally spaced around its circumference. For example, during the
five-colour study we chose colours at positions 14, 86, 158,
230, and 302 counterclockwise rotation from the x-axis (Fig-
ure 3b). This method ensured that neighbouring colours had a con-
stant colour distance. It also ensured that any colour acting as a tar-
get had a constant linear separation from every other (non-target)
colour. A similar technique was used to select colours for the three-
colour, seven-colour, and nine-colour studies. This gave us the fol-
lowing colour distances d and linear separations l:
L* = 67.1
(67.1, 21.1, 11.2)
∆Ε∗
r = 75.0
(a)
Green-
Yellow
Blue
Purple
Red
Yellow
d
l
d
(b)
Figure 3: Choosing colours from the monitor’s gamut: (a) the
boundary of the gamut at L
= 67:1, along with the maximum
inscribed circle centered at (L
;u
;v
) = (67:1;21:1;11:6), ra-
dius 75.0 E
; (b) five colours chosen around the circle’s circum-
ference; each element has a constant colour distance d with its two
neighbours, and a constant linear separation l from the remaining
(non-target) elements
Study d l
three-colour 129.9 E
112.5E
five-colour 88.2 E 51.9 E
seven-colour 65.1 E 28.4 E
nine-colour 51.3 E 17.6 E
Thirty-eight users with normal or corrected acuity participated as
observers during our studies. After ensuring they were not colour
blind, each observer was asked to complete one or more target
blocks. A target block consisted of 360 displays testing a single
colour target from one of the four studies. A total of 66 target
blocks were completed. We used a Macintosh computer with an
8-bit colour display to run our studies. Responses (either “target
present” or “target absent”)and the time to respond for each display
an observer completed were recorded for later analysis.
3.2 Results
Observers had very little difficulty identifying targets during the
three-colour and five-colour studies. Graphs of mean response time
across display size were relatively flat for every colour (Figures 4a
and 4b; note that the y-axis ranges from 400msec to 800msec
in these graphs). Mean response times ranged from 459msec
to 549 msec during the three-colour study, and from 508 msec to
661msec during the five-colour study. Mean response error dur-
ing both studies was approximately 2.5%. We concluded that users
could accurately identify the target in all cases,and that the time re-
quired to do so was relatively independentof display size. This sug-
gests that, even when using five different colours, the visual system
can search for any one of the colours in parallel.
Target identification became significantly more difficult for cer-
tain colours during the seven-colour and nine-colour studies (Fig-
ures 4c and 4d; note that the y-axis ranges from 100msec to
3500msec in these graphs). Mean response error during the seven-
colour study was still low, at approximately 3.3%. The P, Y, R,
BG, and RP targets all exhibited relatively flat responsetime graphs.
Mean response time for these elements ranged from 611msec to
870msec. The G and GY targets, however, gave response times
typical of serial search. An increase in the number of elements be-
ing displayed broughton a corresponding increasein responsetime.
The increase for target-absent displays (approximately 19 msec per
additional element for GY, and 17 msec per element for G) was
roughly twice that for target-present displays (7 msec and 8 msec](https://image.slidesharecdn.com/viz-220418054648/85/Choosing-Effective-Colours-for-Data-Visualization-3-320.jpg)
![per element for GY and G, respectively). Observers had to search
through, on average, half the elements before they found the tar-
get in target-present displays. In target-absent displays, however,
they had to search through all the elements to confirm that no target
existed. This explains why per item search time increased roughly
twice as fast for target-absent displays.
A similar setof results was obtained during the nine-colourstudy.
Overall mean responseerror increasedto 8.1%; it was lowest for the
P, Y, and PB targets (approximately 3.4%), and highest for the G,
GY, and RP targets (approximately 14%). The P, Y, and PB targets
displayed relatively flat response time graphs during target-present
displays. The remaining targets showed some form of serial search.
The effect was weakest for the R and YR targets, and strongest for
the G and GY targets.
3.3 Colour Category Integration
Results from our four studies showed that controlling colour dis-
tance and linear separation alone is not enough to guarantee con-
sistently good (or consistently bad) target identification for every
colour. Results from Kawai suggest that colour category can also
have a strong effect on the amount of time required to identify a
colour target. We decided to see whether colour category results
could explain the asymmetric response times we observed during
the seven-colour and nine-colour studies.
In order to do this, we needed a method of placing individual
colours within a named colour region. We used a simple technique
described by Healey and Enns [9]. This technique places individual
colours within user-named colour regions as follows:
Step 1:
Individual colours are mapped into the Munsell colour model.
Colours are specified in the Munsell model using the three di-
mensions hue, chroma (which controls saturation), and value
(which controls luminance) [3]; the hue dimension is divided
into ten regions that are identified using the ten names R, YR,
Y, GY, G, BG, B, PB, P, and RP. A colour is assignedthe name
of its hue dimension in Munsell space.
Step 2:
The above techniqueis usedto subdivideall or part of a colour
space into named colour regions. We wanted to divide our
maximum inscribed circle into named colour regions. Healey
and Enns divided a similar circle at L
= 71:6 (this corre-
sponds closely to our circle from the L = 67:1 slice), as
shown in Figure 5.
Step 3:
Observers are asked to name representative colours from each
colour region. This converts Munsell names into user-chosen
names; it also provides a way of measuring how strongly a
particular name represents a given colour region. Healey and
Enns used the colour at the middle of each colour region in
Figure 5 to represent the region. Table 1 shows the names
chosen for each of the ten representative colours. For exam-
ple, 5.3% of the observers called the YR region’s representa-
tive colour “red”, 86.8% called it “orange”, and 7.9% called it
“brown”.
Each of the colours from our four studies was assigned a name
using the method described in Step 1 above; this also placed each
colour within one of the ten named colour regions. The user-chosen
colour names shown in Table 1 allow us to measure the similarity of
different colour regions. This category similarity is dependent on:
the range of user-chosen colour names assigned to a given re-
gion
-40.0
-20.0
0.0
20.0
40.0
60.0
80.0
-60.0 -40.0 -20.0 0.0 20.0 40.0 60.0 80.0
v*
u*
B
GY
G
BG
R
YR
Y
PB P
RP
Figure 5: Example of segmenting the circumference of a maximum
inscribed circle from a uv-slice through CIE LUV at L = 71:6;
ticks alongthe the circle mark the boundariesbetweenthe ten named
regions, points are the representative colours for each region
how “strongly” a region covers any user-chosen name in its
range
how the ranges of the two regions overlap
As an example, consider the P and R regions, which have ranges:
purple magenta pink red
P 86.9% 2.6% 5.2%
R 26.3% 71.0%
Colours from P and R overlap only at the “pink” name. Their
overlap is not that strong, since neither P nor R are strongly clas-
sified as pink. The amount of overlap is computed by multiplying
the percentages for the common name. This gives a P–R overlap of
5:2% 26:3% = 0:014. A closer correspondence of user-chosen
names for a pair of regions results in a stronger category similar-
ity, for example, G and GY have an overlap of 97:3%100:0% =
0:973. Categories that overlap over multiple names are combined
using addition, for example, YR and Y overlap in both the “orange”
and the “brown” name, giving a YR-Y overlap of (86:8%2:6%)+
(7:9%44:7%) = 0:058. These overlap values can be used to ex-
plain why certain colours were difficult to use for target identifica-
tion.
During the seven-colour study the P, Y, R, BG, and RP targets
gave good performance. The G and GY targets gave poor perfor-
mance. An examination of Table 2 suggests the following explana-
tion:
targets R, Y, P, and RP have a weak category similarity to the
other colours used during the study; Y has no category simi-
larity to the other colours, R and RP have a category similarity
(i.e., an R-RP overlap) of 0.145, and P and RP have a category
similarity of 0.173
target BG has a moderate category similarity to GY and G
(0.256 and 0.263, respectively)
targets G and GY have a moderate category similarity to BG,
and a very strong category similarity (0.973) to one another
For each target in the seven-colour study, an increase in its cat-
egory similarity to the other colours corresponded closely to an in-
crease in its mean response time. We can measure this correspon-
dence by computing Spearman’s correlation coefficient on the rank](https://image.slidesharecdn.com/viz-220418054648/85/Choosing-Effective-Colours-for-Data-Visualization-5-320.jpg)
![CW d CCW d l
R 42.0 E
42.0 E
12.2 E
YR 42.0 E 49.8 E 14.6 E
Y 49.8 E
112.6 E
39.0 E
G 112.6 E 67.5 E 52.9 E
B 67.5 E
63.2 E
29.1 E
P 63.2 E 48.6 E 20.4 E
RP 48.6 E
42.0 E
14.2 E
Table 3: The exactdistance in CIE LUV between a colour, its clock-
wise neighbour(CW d), andits counter-clockwiseneighbour(CCW
d), as well as the linear separation from the other colours when it
acted as a target
not be used to ensure consistently good identification based on
colour. Colour distance and linear separation need to be considered,
since they have an effect on search performance. We used a more
systematic colour selection technique to chooseanother set of seven
colours:
a single green was chosenfrom the combined G-GY colour re-
gion; observers see the entire region as green, which means it
can be used for only a single colour
the clockwise neighbour of our green was a yellow, chosen to
lie on the border between the GY and Y colour regions
the counterclockwiseneighbourofourgreen wasa blue-green,
chosen to lie in the center of the BG colour region (we did not
use the colour from the border between G and BG, because it
was too difficult to differentiate from our green)
the remaining four colours (chosen from the YR, R, RP, and
PB colourregions)were atequalsteps betweenour yellow and
our blue-green
This gave us a constant neighbour distance d and linear separa-
tion l (59:4E and 24:6E, respectively) between the Y, YR,
R, RP, PB, and BG colours (G had a larger d and l than the other
colours). Results from displays using these colours as targets are
shown in Figure 7. Mean response error was 5.2%. Response time
graphs for all seven colours are much flatter than in the original
seven-colourstudy, although G and Y still gavemixed results during
target-absentdisplays. We couldhavefurther differentiated theseel-
ements by choosing a yellow from the center of the Y colour region
(rather than at the GY-Y border). This might have resulted in poorer
performance for the other targets due to the reduction in colour dis-
tance and linear separation between our Y, YR, R, RP, PB, and BG
colours, however. It appears that seven isoluminant colours is the
maximum we can display at one time, while still allowing rapid and
accurate identification of any one of the colours.
4 VISUALIZING MEDICAL IMAGES
As we discussed in the introduction, our long-term goal is a set
of guidelines that help users maximize the total number of colours
available during visualization, while still allowing for the rapid and
accurate detection of any one of the given colours. Experimental
results showed how we can control colour distance, linear separa-
tion, and colour category to pick seven isoluminant colours that sat-
isfy our requirements. Extending these results to a practical appli-
cation is a non-trivial problem. However, even the specific results
reported here can help to design useful visualization applications.
To this end, we discusshow coloursfrom ourexperiments were used
to visualize slice data from a Computerized Tomography (CT) scan.
500
750
1000
1250
1500
1750
2000
2250
2500
2750
3000
3250
3500
17 33 49
Display Size (Elements)
Response
Time
(msec)
Colour Category Absent
RP
YR
R
PB
BG
Y
G
(a)
500
750
1000
1250
1500
1750
2000
2250
2500
2750
3000
3250
3500
17 33 49
Display Size (Elements)
Response
Time
(msec)
RP
PB
R
YR
Y
BG
G
Colour Category Present
(b)
Figure 7: Response time graphs for the second colour-category
study: (a) target absent displays; (b) target present displays
The scan was taken to locate an aneurism in a patient’s brain. This
resulted in 116 individual slices, each of which had a 512 512
resolution.
We began by displaying slices as greyscale images (Figure 8a);
values in each slice had been scaled to range from (0:::255) for
exactly this purpose. Higher values in a slice correspond to denser
material detected during the CT scan. Greyscale images are a stan-
dard method for displaying medical slice data [20, 2], since they
allow a user to easily locate individual regions of similar density.
Through consultation we identified seven intensity ranges that cor-
responded to seven individual regions of interest. Slices were then
viewed using seven different colours. Figure 8b displays the slice
using colours from our original seven-colourexperiment. Figure 8c
displays the slice using colours from our final colour category ex-
periment.
There are three important points to note. First, as with the
greyscale representation, it is easy to locate individual regions in
both colour slices. The densities (and therefore the intensity values)
of different regions of interest are consistent across slices. There-
fore, our colour displays work well for any individual slice chosen
by the user.
Second, as predicted by our experiment results, the colours used
in Figure 8b do not always provide good differentiation between re-
gions, compared to the colours used in Figure 8c. Consider the re-
gion reference by the arrow. In Figure 8b, it appears to made up of
a single type of element. In fact, it is made up of two types of ele-
ments; this is clear when examining the same region in Figure 8c.
Notice there is no corresponding tradeoff between Figure 8c and
Figure 8b, that is, there are no pairs of elements that are difficult to
differentiate in Figure 8c but easy to differentiate in Figure 8b.
Finally, colour images can provide visual cues that might not be
immediately apparentin a greyscale display. For example, consider](https://image.slidesharecdn.com/viz-220418054648/85/Choosing-Effective-Colours-for-Data-Visualization-7-320.jpg)
![the small spots of purple at the top of the region in the center of the
image. These spots represent areas of high relative density, and are
quickly identified as different from the surrounding material. The
same information is present in the greyscale image, but the visual
system is not immediately drawn to it, since the contrast between
the region and its neighbours is not as pronounced. One method of
visualizing slice data is to “fly” through the slice stack, rapidly dis-
playing individual slices one after another in a movie-like fashion
[7]. In this context, we expectusers could more accuratelydetectar-
eas with a specific density range using colour rather than greyscale.
5 DISCUSSION
This paper has presented a simple method for choosing effective
colours for use during data visualization. We accomplish this by
controlling colour distance, linear separation of colours, and colour
category similarity during colour selection. We found that an isolu-
minant slice through a monitor’s gamut yielded up to seven differ-
ent colours, any one of which could be rapidly and accurately de-
tected, even in the presence of all the others. Each colour needed
a minimum colour distance, linear separation, and colour category
differentiation to guarantee good performance. One of our future
goals is a set of investigations that explicitly measure the tradeoff in
observer performance when these effects are varied. This will also
allow us to identify the minimum required value for each effect.
The perceived intensity of each colour was held constant during
ourstudies to eliminate the potentialfor visualinterference. It might
be possible to use intensity in a controlled manner to increase the
number of colours we can display simultaneously. For example, we
might be able to choose multiple groups of colours, with each group
being restricted to an isoluminantslice through the monitor’s gamut.
A colour group would be assigned to a conceptually related set of
attribute values in the dataset to be visualized. Proper division by
intensity might allow a user to select individual colours either from
within a group (e.g., a dull red element or a bright green element) or
across groups (e.g., all red elements). This type of technique is not
possible if colours of varying intensity are chosen from the moni-
tor’s gamut in an ad hoc fashion.
Acknowledgements
I would like to thank Dr. Kellogg S. Booth and Dr. James T. Enns
for the technical advice they offered during my research. I would
also like to thank Ms. Jeanette Lum for coordinating and running
the most of the experiment sessions.
References
[1] BAUER, B., JOLICOEUR, P., AND COWAN, W. B. Visual
search for colour targets that are or are not linearly-separable
from distractors. Vision Research (1996), (in press).
[2] BERGMAN, L. D., ROGOWITZ, B. E., AND TREINISH, L. A.
A rule-basedtoolfor assistingcolormapselection. In Proceed-
ings Visualization ’95 (Atlanta, Georgia, 1995), pp. 118–125.
[3] BIRREN, F. Munsell: A Grammar of Color. Van Nostrand
Reinhold Company, New York, New York, 1969.
[4] CALLAGHAN, T. C. Dimensional interaction of hue and
brightness in preattentive field segregation. Perception Psy-
chophysics 36, 1 (1984), 25–34.
[5] CAMBELL, W. J., SHORT, JR., N. M., AND TREINISH, L. A.
Adding intelligenceto scientificdata management.Computers
in Physics 3, 3 (1989), 26–32.
[6] D’ZMURA, M. Color in visual search. Vision Research31, 6
(1991), 951–966.
[7] HEALEY, C. G., BOOTH, K. S., AND ENNS, J. T. Real-time
multivariate data visualization using preattentive processing.
ACM Transactions on Modelling and Computer Simulation 5,
3 (1995), 190–221.
[8] HEALEY, C. G., BOOTH, K. S., AND ENNS, J. T. High-speed
visual estimation using preattentive processing. ACM Trans-
actions on Computer-Human Interaction (1996), (in press,
scheduled for July 1996 issue).
[9] HEALEY, C. G., AND ENNS, J. T. A perceptual colour seg-
mentation algorithm. Tech. Rep. TR-96-09, Department of
Computer Science, University of British Columbia, 1996.
[10] HITT, E. F. Retrofit of flat panel color displays in the F-
16A/B. In ProceedingsIEEE/AIAA 11th Digital Avionics Sys-
tems Conference (Seattle, Washington, 1992), pp. 323–328.
[11] KAWAI, M., UCHIKAWA, K., AND UJIKE, H. Influence of
color category on visual search. In Annual Meeting of the
Association for Research in Vision and Ophthalmology (Fort
Lauderdale, Florida, 1995), p. #2991.
[12] LEVKOWITZ, H., AND HERMAN, G. T. Color scales for
image data. IEEE Computer Graphics Applications 12, 1
(1992), 72–80.
[13] RHEINGANS, P., AND TEBBS, B. A tool for dynamic explo-
rations of color mappings. Computer Graphics 24, 2 (1990),
145–146.
[14] ROBERTSON, P. K. Visualizing color gamuts: A user inter-
face for the effective use of perceptualcolor spacesin data dis-
plays. IEEE Computer Graphics Applications 8, 5 (1988),
50–64.
[15] ROGOWITZ, B. E., AND TREINISH, L. A. An architecture
for rule-based visualization. In Proceedings Visualization ’93
(San Jose, California, 1993), pp. 236–243.
[16] ROSENBLUM,L. J. Researchissuesin scientificvisualization.
IEEE ComputerGraphics Applications14, 2 (1994), 61–85.
[17] TREINISH, L. A., FOLEY, J. D., CAMPBELL, W. J., HABER,
R. B., AND GURWITZ, R. F. Effective software systems for
scientific data visualization. Computer Graphics 23, 5 (1989),
111–136.
[18] TRIESMAN, A. Search, similarity, and integration of fea-
ture between and within dimensions. Journal of Experimental
Psychology: Human Perception Performance17, 3 (1991),
652–676.
[19] VAREY, C. A., MELLERS, B. A., AND BIRNBAUM, M. H.
Judgments of proportions. Journal of Experimental Psychol-
ogy: Human Perception Performance 16, 3 (1990), 613–
625.
[20] WARE, C. Color sequences for univariate maps: Theory, ex-
periments, and principles. IEEE Computer Graphics Appli-
cations 8, 5 (1988), 41–49.
[21] WARE, C., AND BEATTY, J. C. Using colour dimensions to
display data dimensions. Human Factors 30, 2 (1988), 127–
142.
[22] WYSZECKI, G., AND STILES, W. S. Color Science: Con-
ceptsandMethods,Quantitative Data and Formulae,2nd Edi-
tion. John Wiley Sons, Inc., New York, New York, 1982.](https://image.slidesharecdn.com/viz-220418054648/85/Choosing-Effective-Colours-for-Data-Visualization-9-320.jpg)
Choosing Effective Colours for Data Visualization
- 1. Choosing Effective Colours for Data Visualization Christopher G. Healey1 Department of Computer Science, University of British Columbia ABSTRACT In this paper we describe a technique for choosing multiple colours for use during data visualization. Our goal is a systematic method for maximizing the total number of colours available for use, while still allowing an observer to rapidly and accurately search a display for any one of the given colours. Previous research suggests that we need to consider three separate effects during colour selection: colour distance, linear separation, and colour category. We describe a simple method for measuring and controlling all of these effects. Our method was tested by performing a set of target identification studies; we analysed the ability of thirty-eight observers to find a colour target in displays that contained differently coloured back- ground elements. Results showedour method canbe usedto selecta group of colours that will provide good differentiation between data elements during data visualization. CR Descriptors: H.5.2 [Information Interfaces and Presen- tation]: User Interfaces - ergonomics, screen design (graphics, colour); I.3.6 [Computer Graphics]: Methodology and Tech- niques - ergonomics,interaction techniques 1 INTRODUCTION Scientific visualization in computer graphics is a rapidly expanding area of research. This is due in large part to the dramatic increase in both the size and the number of datasets that need to be visu- alized [5, 17]. To date, many application-specific tools have been built to help analyse individual datasets. Much less work has fo- cused on developing guidelines for the design of visualization tech- niques [16]. Our work is intended to address one aspectof this more general question. A typical method of visualizing a dataset involves mapping data attributes to visual features (e.g., shape, size, spatial location, and orientation). Colour is an important and frequently-used feature. Examples include colour temperature gradients on maps and charts, colour-coded vector fields in flow visualization, or colour icons dis- played by real-time simulation systems. If we use colour to represent our data, an important question to ask is: “How can we chooseeffective colours that provide good dif- ferentiation between data elements during the visualization task?” We addressthis problem by trying to answerthree related questions: How can we allow rapid and accurateidentification of individ- ual data elements through the use of colour? What factors determine whether a “target” element’s colour will make it easy to find, relative to differently coloured “non- target” elements? How many colours can we display at once, while still allowing for rapid and accurate target identification? 1Departmentof ComputerScience,University of British Columbia, 2366 Main Mall, Vancouver, British Columbia, V6T 1Z2, Canada Previous work has addressed the issue of choosing colours for certain types of data visualization. For example, Ware and Beatty describe a simple colour visualization technique for displaying cor- relation in a five-dimensional dataset [21]. Research on the design of military systems sometimes quotes anecdotal evidence that sug- gests “:::the generalguideline for computer-generated images is no more than five to seven colours at a time:::” [10], although they of- fer no explanation for why this might be the case. Robertson, Ware, Rheingans and Tebbs, and Levkowitz and Herman discuss vari- ous methods for building effective colour gamuts and colourmaps [14, 20, 13, 12]. Recent work at the IBM Thomas J. Watson Re- search Center has focused on a rule-based visualization tool that considers how a user perceives visual features like hue, luminance, height, and so on [15, 2]. None of these techniques were designed to investigate the rapid and accurate identification of individual data elements based on colour. Also, since the colour gamut and colourmap work uses con- tinuous colour scales to encodeinformation, they do not addressthe question of how many colours we can effectively display at once, while still providing gooddifferentiation betweenindividualdata el- ements. An intuitive first step to gaining more control over colour would be to use a perceptualcolourmodellike CIE LUV, CIE Lab,or Mun- sell [22]. Unfortunately, fixing the colour distance in a perceptual colourmodel to a constantvalue doesnot guarantee that each colour will be equally easy to detect. Other factors can affect how groups of coloured elements interact with one another. Our technique does, in fact, use the CIE LUV colour model to provide control over colour distance and isoluminance. We also ex- ploit two specific results related to colour target detection: linear separation [6, 1] and colour category [11]. These effects are con- trolled to allow for the rapid and accurate identification of colour targets. Target identification is a necessary first step towards per- forming other types of exploratory data analysis. If we can rapidly and accurately differentiate elements based on their colour, we can apply our results to other important visualization techniqueslike de- tection of data boundaries, the tracking of data regions in real-time, and enumeration tasks like counting and estimation [19, 18, 8]. 2 BACKGROUND Before we describe our technique in detail, we provide a brief overview of the CIE LUV colour model, as well as a description of the linear separation and colour category effects. 2.1 CIE LUV The CIE LUV colour model was proposed by the Commission In- ternationale de L’Èclairge (CIE) in 1976 [22]. Colours are speci- fied using the three dimensions L (which encodes luminance), u, and v (which together encode chromaticity). CIE LUV provides two useful properties for controlling perceived colour difference. First, colours with the same L are isoluminant. Second, Euclidean distance and perceived colour difference (specified in E units) can be interchanged,since the colourdifference between two colour
- 2. stimuli x and y is roughly: E xy = p(L xy)2 +(u xy)2 +(v xy)2 (1) 2.2 Linear Separation The linear separation effect was originally described by D’Zmura [1991] . He was investigating how the human visual system finds a target colour in a sea of background non-target colours. D’Zmura ran experiments that asked observers to determine the presence or absence of an orange target. Two groups of differently coloured non-target elements were also present in each display (e.g., in one experiment half the non-targets in each display were coloured green and half were coloured red). Results showed that when the target could be separated by a straight line from its non-targets in colour space(Figure 1, target T and non-targets A and C), the time required to determine the target’s presence or absence was constant, and in- dependent of the total number of elements being displayed. This suggests detection occurs in parallel in the low-level visual system. When the target was collinear with its non-targets (Figure 1, target T and non-targets A and B), the time required to identify the tar- get was linearly proportional to the number of elements being dis- played. Observers had to search serially through each display to de- termine whether the target was present or absent. T B A C Figure 1: Example of a target with both collinear and separablenon- targets, the colours are shown in a u;v-slice from the CIE LUV colour model, notice that target T is equidistant from all three non- targets A, B, and C; in Case 1, target colour T is collinear with the non-target colours A and B; in Case 2, target T is linearly separable from its non-targets A and C Work by Bauer [1996] provided a number of results that strengthen D’Zmura’s hypothesis. First, he showed that perceptual colour models cannot be used to overcome linear separation. In his experiments the target–non-target distance was fixed to a constant value in CIE LUV (Figure 1). In spite of this, the target required serial searching when it was collinear with its non-targets, but was significantlyeasierto detectwhenit wasseparable. Bauerreplicated his initial findings in three additional colour regions: green, blue, and green-yellow. This suggestslinear separation applies to colours from different parts of the visible colour domain. 2.3 Colour Category Kawai [1995] reported results that suggest that the time required to identify a colour target depends in part on the named colour regions occupiedby the target and its non-targets. If target identification de- pended only on the colour distance between a target and its non- targets, the time taken to detect presence or absence of the target should decrease uniformly as target–non-target distance increases. Kawai ran a set of experiments that showed this was not true. Kawai tested displays that contained a uniquely colour target and a constant number of uniformly coloured non-targets. He divided blue boundary purple boundary green boundary T A B g-axis j-axis Figure 2: A target T and two non-targets A and B shown in a (j;g) slice from the OSA colour model; the boundariesof the blue, green, and purple colour regions are shown as thick lines; T and B occupy the same named colour region “blue”, but A does not an isoluminant colour slice used during his experiments into named colour regions. His results showed that search times decreased dra- matically whenever the non-target was moved outside the target’s colour region. For example, finding a target coloured T in a set of non-targets coloured B was significantly more difficult than than finding T in a set of non-targets coloured A (Figure 2). The target– non-target distances TAand TB are equal; Kawai suggeststhe dif- ference in performance is due to the fact that both T and B are lo- cated in the blue colour region, but A is not. Similar search time asymmetries for separate experiments testing green, purple, and red targets were also explained by a difference in colour regions. 3 COLOUR SELECTION TECHNIQUE Results discussed in the background section suggest that choosing effective colours for data visualization dependson at least three sep- arate criteria: colour distance: the Euclidean distance between different colours as measured in a perceptually balanced colour model linear separation: the ability to linearly separate targets from non-targets in the colour model being used colour category: the named colour regions occupied by both the target and non-target elements We began our investigation by controlling colour distance and linear separation, but not colour category. We proceeded under the assumption that the user might choose to search for any one of the available data elements at any given time. This is typical during ex- ploratory data analysis; users will often change the focus of their in- vestigation based on the data they see as the visualization unfolds. Our requirement meant that the colourselection techniquehad to al- low for rapid and accurateidentificationof anyof the elementsbeing displayed. We ran four studies to investigate the tradeoff between the num- ber of colours displayed and the time required to determine the pres- enceor absenceof a target element. Eachstudy displayeda different number of unique colours, which we identify using the names red, yellow-red, yellow, green-yellow, green, blue-green, blue, purple- blue, purple, and red-purple (or R, YR, Y, GY, G, BG, B, PB, P, and RP): three-colour study: each display contained three different colours (i.e., one colour for the target and two for the non- targets) which we named R, GY, and PB
- 3. five-colour study: each display contained five different colours: R, Y, GY, B, and P seven-colour study: each display contained seven different colours: R, Y, GY, G, BG, P, and RP nine-colour study: each display contained nine different colours: R, YR, Y, GY, G, BG, PB, P, and RP Every colourin a given study was tested as a target. For example, the three-colour study was run three times, first with an R element acting as a target (and GY and PB elements acting as non-targets), next with a GY target (and R and PB non-targets), and finally with a PB target (and R and GY non-targets). Faster search times for cer- tain targets would have to be explained in terms of colour distance, linear separation, or colour category. Each of the four studies were themselves divided in the following manner: elements in each display were drawn as coloured squares, and were randomlylocatedon an underlying99grid thatcovered the entire viewing area of the monitor half of the displays were randomly chosen to contain an ele- ment that used the target colour; the other half did not one-third of the displays contained a total of 17 elements (one target and 16 non-targets if the target was present, or 17 non- targets if the target was absent); one-third of the displays con- tained a total of 33 elements; one-third of the displays con- tained a total of 49 elements Colours for each of the four studies were chosen such that the colour distance between pairs of colours and the linear separation for each colour were fixed to constant values. Our results showed that detection was rapid and accurate for all colours from both the three-colour and five-colour studies. Results from the seven-colour and nine-colour studies were mixed; some colours gave better per- formance than others. This difference was explained when we ex- amined the colour regions occupied by each of the colours. 3.1 Method We began the colour selection process by obtaining the chromatic- ities of our monitor’s triads. We also measured the luminance of the monitor’s maximum intensity red, green, and blue with a spot photometer. These values were needed to convert colours from CIE LUV into the monitor’s RGB gamut. We wanted to ensure that the colours we chose had the same per- ceived intensity. Previous research has suggested that random vari- ation in intensity can interfere with an observer’s ability to perform visualization tasks based on colour [4]. In order to guarantee iso- luminance, all the colours were chosen from a single u;v-slice through the CIE LUV colour space at L = 67:1. We wanted to maximize the number of available colours, while still maintaining control over colour distance and linear separability. To do this, we computed the boundary of the monitor’s gamut in the L = 67:1 slice. We then found the largest circle inscribed within the gamut (Figure 3a). Given the maximum inscribed circle, we chose colours that were equally spaced around its circumference. For example, during the five-colour study we chose colours at positions 14, 86, 158, 230, and 302 counterclockwise rotation from the x-axis (Fig- ure 3b). This method ensured that neighbouring colours had a con- stant colour distance. It also ensured that any colour acting as a tar- get had a constant linear separation from every other (non-target) colour. A similar technique was used to select colours for the three- colour, seven-colour, and nine-colour studies. This gave us the fol- lowing colour distances d and linear separations l: L* = 67.1 (67.1, 21.1, 11.2) ∆Ε∗ r = 75.0 (a) Green- Yellow Blue Purple Red Yellow d l d (b) Figure 3: Choosing colours from the monitor’s gamut: (a) the boundary of the gamut at L = 67:1, along with the maximum inscribed circle centered at (L ;u ;v ) = (67:1;21:1;11:6), ra- dius 75.0 E ; (b) five colours chosen around the circle’s circum- ference; each element has a constant colour distance d with its two neighbours, and a constant linear separation l from the remaining (non-target) elements Study d l three-colour 129.9 E 112.5E five-colour 88.2 E 51.9 E seven-colour 65.1 E 28.4 E nine-colour 51.3 E 17.6 E Thirty-eight users with normal or corrected acuity participated as observers during our studies. After ensuring they were not colour blind, each observer was asked to complete one or more target blocks. A target block consisted of 360 displays testing a single colour target from one of the four studies. A total of 66 target blocks were completed. We used a Macintosh computer with an 8-bit colour display to run our studies. Responses (either “target present” or “target absent”)and the time to respond for each display an observer completed were recorded for later analysis. 3.2 Results Observers had very little difficulty identifying targets during the three-colour and five-colour studies. Graphs of mean response time across display size were relatively flat for every colour (Figures 4a and 4b; note that the y-axis ranges from 400msec to 800msec in these graphs). Mean response times ranged from 459msec to 549 msec during the three-colour study, and from 508 msec to 661msec during the five-colour study. Mean response error dur- ing both studies was approximately 2.5%. We concluded that users could accurately identify the target in all cases,and that the time re- quired to do so was relatively independentof display size. This sug- gests that, even when using five different colours, the visual system can search for any one of the colours in parallel. Target identification became significantly more difficult for cer- tain colours during the seven-colour and nine-colour studies (Fig- ures 4c and 4d; note that the y-axis ranges from 100msec to 3500msec in these graphs). Mean response error during the seven- colour study was still low, at approximately 3.3%. The P, Y, R, BG, and RP targets all exhibited relatively flat responsetime graphs. Mean response time for these elements ranged from 611msec to 870msec. The G and GY targets, however, gave response times typical of serial search. An increase in the number of elements be- ing displayed broughton a corresponding increasein responsetime. The increase for target-absent displays (approximately 19 msec per additional element for GY, and 17 msec per element for G) was roughly twice that for target-present displays (7 msec and 8 msec
- 4. 500 750 17 33 49 500 750 17 33 49 500 750 17 33 49 500 750 17 33 49 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 (a) (b) (c) (d) 5-Item Present Display Size (Elements) Response Time (msec) 5-Item Absent Display Size (Elements) Response Time (msec) Y P 3-Item Absent PB R Display Size (Elements) Response Time (msec) 3-Item Present PB R Display Size (Elements) Response Time (msec) P 7-Item Absent Display Size (Elements) Response Time (msec) Display Size (Elements) Response Time (msec) 7-Item Present 9-Item Absent Display Size (Elements) Response Time (msec) 9-Item Present Display Size (Elements) Response Time (msec) BG GY RP YR Y P GY GY R GY B R B GY Y P Y R BG RP GY G RP Y R BG G GY G YR PB Y RP R P G GY P PB BG R Figure 4: Response time graphs for the four studies, the graph on the left represents displays where the target was absent, while the graph on the right represents displays where the target was present: (a) response time as a function of display size (i.e., total number of elements shown in the display) for each target from the three-colour study; (b) response times for each target from the five-colour study; (c) response times for each target from the seven-colour study; (d) response times for each target from the nine-colour study
- 5. per element for GY and G, respectively). Observers had to search through, on average, half the elements before they found the tar- get in target-present displays. In target-absent displays, however, they had to search through all the elements to confirm that no target existed. This explains why per item search time increased roughly twice as fast for target-absent displays. A similar setof results was obtained during the nine-colourstudy. Overall mean responseerror increasedto 8.1%; it was lowest for the P, Y, and PB targets (approximately 3.4%), and highest for the G, GY, and RP targets (approximately 14%). The P, Y, and PB targets displayed relatively flat response time graphs during target-present displays. The remaining targets showed some form of serial search. The effect was weakest for the R and YR targets, and strongest for the G and GY targets. 3.3 Colour Category Integration Results from our four studies showed that controlling colour dis- tance and linear separation alone is not enough to guarantee con- sistently good (or consistently bad) target identification for every colour. Results from Kawai suggest that colour category can also have a strong effect on the amount of time required to identify a colour target. We decided to see whether colour category results could explain the asymmetric response times we observed during the seven-colour and nine-colour studies. In order to do this, we needed a method of placing individual colours within a named colour region. We used a simple technique described by Healey and Enns [9]. This technique places individual colours within user-named colour regions as follows: Step 1: Individual colours are mapped into the Munsell colour model. Colours are specified in the Munsell model using the three di- mensions hue, chroma (which controls saturation), and value (which controls luminance) [3]; the hue dimension is divided into ten regions that are identified using the ten names R, YR, Y, GY, G, BG, B, PB, P, and RP. A colour is assignedthe name of its hue dimension in Munsell space. Step 2: The above techniqueis usedto subdivideall or part of a colour space into named colour regions. We wanted to divide our maximum inscribed circle into named colour regions. Healey and Enns divided a similar circle at L = 71:6 (this corre- sponds closely to our circle from the L = 67:1 slice), as shown in Figure 5. Step 3: Observers are asked to name representative colours from each colour region. This converts Munsell names into user-chosen names; it also provides a way of measuring how strongly a particular name represents a given colour region. Healey and Enns used the colour at the middle of each colour region in Figure 5 to represent the region. Table 1 shows the names chosen for each of the ten representative colours. For exam- ple, 5.3% of the observers called the YR region’s representa- tive colour “red”, 86.8% called it “orange”, and 7.9% called it “brown”. Each of the colours from our four studies was assigned a name using the method described in Step 1 above; this also placed each colour within one of the ten named colour regions. The user-chosen colour names shown in Table 1 allow us to measure the similarity of different colour regions. This category similarity is dependent on: the range of user-chosen colour names assigned to a given re- gion -40.0 -20.0 0.0 20.0 40.0 60.0 80.0 -60.0 -40.0 -20.0 0.0 20.0 40.0 60.0 80.0 v* u* B GY G BG R YR Y PB P RP Figure 5: Example of segmenting the circumference of a maximum inscribed circle from a uv-slice through CIE LUV at L = 71:6; ticks alongthe the circle mark the boundariesbetweenthe ten named regions, points are the representative colours for each region how “strongly” a region covers any user-chosen name in its range how the ranges of the two regions overlap As an example, consider the P and R regions, which have ranges: purple magenta pink red P 86.9% 2.6% 5.2% R 26.3% 71.0% Colours from P and R overlap only at the “pink” name. Their overlap is not that strong, since neither P nor R are strongly clas- sified as pink. The amount of overlap is computed by multiplying the percentages for the common name. This gives a P–R overlap of 5:2% 26:3% = 0:014. A closer correspondence of user-chosen names for a pair of regions results in a stronger category similar- ity, for example, G and GY have an overlap of 97:3%100:0% = 0:973. Categories that overlap over multiple names are combined using addition, for example, YR and Y overlap in both the “orange” and the “brown” name, giving a YR-Y overlap of (86:8%2:6%)+ (7:9%44:7%) = 0:058. These overlap values can be used to ex- plain why certain colours were difficult to use for target identifica- tion. During the seven-colour study the P, Y, R, BG, and RP targets gave good performance. The G and GY targets gave poor perfor- mance. An examination of Table 2 suggests the following explana- tion: targets R, Y, P, and RP have a weak category similarity to the other colours used during the study; Y has no category simi- larity to the other colours, R and RP have a category similarity (i.e., an R-RP overlap) of 0.145, and P and RP have a category similarity of 0.173 target BG has a moderate category similarity to GY and G (0.256 and 0.263, respectively) targets G and GY have a moderate category similarity to BG, and a very strong category similarity (0.973) to one another For each target in the seven-colour study, an increase in its cat- egory similarity to the other colours corresponded closely to an in- crease in its mean response time. We can measure this correspon- dence by computing Spearman’s correlation coefficient on the rank
- 6. purple magenta pink red orange brown yellow green aqua blue other P 86.9% 2.6% 5.2% – – – – – – – 5.2% RP 15.7% 28.9% 55.3% – – – – – – – – R – – 26.3% 71.0% – – – – – – 2.6% YR – – – 5.3% 86.8% 7.9% – – – – – Y – – – – 2.6% 44.7% 47.4% – – – 5.2% GY – – – – – – – 97.3% – – 2.6% G – – – – – – – 100.0% – – – BG – – – – – – – 26.3% 57.8% 15.8% – B – – – – – – – – 7.9% 89.4% 2.6% PB 5.2% – – – – – – – – 92.1% 2.6% Table 1: Responsesfor the representative colours from the ten named regions, shown as the percentage of observers who chose a given name order of our colours in terms of total category similarity and mean response time. R Y GY G BG P RP R – 0 0 0 0 .014 .145 Y 0 – 0 0 0 0 0 GY 0 0 – .973 .256 0 0 G 0 0 .973 – .263 0 0 BG 0 0 .256 .263 – 0 0 P .014 0 0 0 0 – .173 RP .145 0 0 0 0 .173 – Total .159 0 1.229 1.236 .519 .187 .318 Table 2: Similarity table, showingindividualand total categorysim- ilarity values for each of the colours used during the seven-colour study Table 2 sums the category similarity measures for each of the seven colours we used. The rank order of total category similar- ity from lowest to highest is Y, R, P, RP, BG, GY, and G. Rank- ing our colours based on total mean response time (including both presentand absenttrials across all three display sizes) from smallest to largest gave an order of P, R, Y, BG, RP, GY, and G. The Spear- man correlation between these rankings is r = 0:821, confirming that higher mean response times for a given target correspond to a higher category similarity. A similarity–response time correlation can also be observed dur- ing the nine-colour study. Targets P, Y, YR, and PB had a weak cat- egory similarity to other colours used during the study. Targets R, RP, and BG had a moderate category similarity to each other and to G and GY. Targets G and GY had a moderate category similarity to BG, and a strong category similarity to one another. The Spearman correlation between the colours’ total similarity and mean response time ranking was r = 0:762. Results for YR and BG are somewhat anomalous; category similarity values suggest we should have ob- servedbetter performance for the YR target, andworse performance for the BG target. Results for the other colours correspond closely to their category similarity measures. 3.4 Colour Category Study Results from investigating colour category integration might imply that effective colours can be selected by controlling colour category alone. This was tested by selecting seven colours from colour re- gions that had low category similarity with one another. An exam- ination of observer responses in Table 1 shows that colours chosen from the R, YR, Y, G, B, P, and RP regions satisfy our restriction; the largest category overlap occurs between RP-R (0.145) and RP-P (0.173). The colours we used were exactly those chosen by Healey and Enns to represent the above seven regions. Seven observers with normal or corrected acuity participated as observers during the colour category study. A total of 360 displays for each target were completed by the observers. 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 R YR Display Size (Elements) Response Time (msec) Colour Category Absent B G P Y RP (a) 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 Display Size (Elements) Response Time (msec) B G P Y RP R YR Colour Category Present (b) Figure 6: Response time graphs for the first colour-category study: (a) target absent displays; (b) target present displays Mean response error was 3.4%, which shows responses were ac- curate. The pattern of response times, however, was similar to the original seven-colour study. Several colours (in particular R and YR) exhibited poor search performance. This can be explained by examining the distance between neighbouring pairs of colours, and the linear separation for each colour when it acted as a target (Ta- ble 3). Colours that gave the worst search performance (R, YR, and RP) had the smallest neighbour distances and linear separation; in fact, the R, YR, and RP targets had a linear separation that was smaller than the one used during the nine-colourstudy. Colours that gave the best search performance (Y, G, B, and P) had the largest neighbour distances and linear separation. From these results we concluded that colour category alone can-
- 7. CW d CCW d l R 42.0 E 42.0 E 12.2 E YR 42.0 E 49.8 E 14.6 E Y 49.8 E 112.6 E 39.0 E G 112.6 E 67.5 E 52.9 E B 67.5 E 63.2 E 29.1 E P 63.2 E 48.6 E 20.4 E RP 48.6 E 42.0 E 14.2 E Table 3: The exactdistance in CIE LUV between a colour, its clock- wise neighbour(CW d), andits counter-clockwiseneighbour(CCW d), as well as the linear separation from the other colours when it acted as a target not be used to ensure consistently good identification based on colour. Colour distance and linear separation need to be considered, since they have an effect on search performance. We used a more systematic colour selection technique to chooseanother set of seven colours: a single green was chosenfrom the combined G-GY colour re- gion; observers see the entire region as green, which means it can be used for only a single colour the clockwise neighbour of our green was a yellow, chosen to lie on the border between the GY and Y colour regions the counterclockwiseneighbourofourgreen wasa blue-green, chosen to lie in the center of the BG colour region (we did not use the colour from the border between G and BG, because it was too difficult to differentiate from our green) the remaining four colours (chosen from the YR, R, RP, and PB colourregions)were atequalsteps betweenour yellow and our blue-green This gave us a constant neighbour distance d and linear separa- tion l (59:4E and 24:6E, respectively) between the Y, YR, R, RP, PB, and BG colours (G had a larger d and l than the other colours). Results from displays using these colours as targets are shown in Figure 7. Mean response error was 5.2%. Response time graphs for all seven colours are much flatter than in the original seven-colourstudy, although G and Y still gavemixed results during target-absentdisplays. We couldhavefurther differentiated theseel- ements by choosing a yellow from the center of the Y colour region (rather than at the GY-Y border). This might have resulted in poorer performance for the other targets due to the reduction in colour dis- tance and linear separation between our Y, YR, R, RP, PB, and BG colours, however. It appears that seven isoluminant colours is the maximum we can display at one time, while still allowing rapid and accurate identification of any one of the colours. 4 VISUALIZING MEDICAL IMAGES As we discussed in the introduction, our long-term goal is a set of guidelines that help users maximize the total number of colours available during visualization, while still allowing for the rapid and accurate detection of any one of the given colours. Experimental results showed how we can control colour distance, linear separa- tion, and colour category to pick seven isoluminant colours that sat- isfy our requirements. Extending these results to a practical appli- cation is a non-trivial problem. However, even the specific results reported here can help to design useful visualization applications. To this end, we discusshow coloursfrom ourexperiments were used to visualize slice data from a Computerized Tomography (CT) scan. 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 Display Size (Elements) Response Time (msec) Colour Category Absent RP YR R PB BG Y G (a) 500 750 1000 1250 1500 1750 2000 2250 2500 2750 3000 3250 3500 17 33 49 Display Size (Elements) Response Time (msec) RP PB R YR Y BG G Colour Category Present (b) Figure 7: Response time graphs for the second colour-category study: (a) target absent displays; (b) target present displays The scan was taken to locate an aneurism in a patient’s brain. This resulted in 116 individual slices, each of which had a 512 512 resolution. We began by displaying slices as greyscale images (Figure 8a); values in each slice had been scaled to range from (0:::255) for exactly this purpose. Higher values in a slice correspond to denser material detected during the CT scan. Greyscale images are a stan- dard method for displaying medical slice data [20, 2], since they allow a user to easily locate individual regions of similar density. Through consultation we identified seven intensity ranges that cor- responded to seven individual regions of interest. Slices were then viewed using seven different colours. Figure 8b displays the slice using colours from our original seven-colourexperiment. Figure 8c displays the slice using colours from our final colour category ex- periment. There are three important points to note. First, as with the greyscale representation, it is easy to locate individual regions in both colour slices. The densities (and therefore the intensity values) of different regions of interest are consistent across slices. There- fore, our colour displays work well for any individual slice chosen by the user. Second, as predicted by our experiment results, the colours used in Figure 8b do not always provide good differentiation between re- gions, compared to the colours used in Figure 8c. Consider the re- gion reference by the arrow. In Figure 8b, it appears to made up of a single type of element. In fact, it is made up of two types of ele- ments; this is clear when examining the same region in Figure 8c. Notice there is no corresponding tradeoff between Figure 8c and Figure 8b, that is, there are no pairs of elements that are difficult to differentiate in Figure 8c but easy to differentiate in Figure 8b. Finally, colour images can provide visual cues that might not be immediately apparentin a greyscale display. For example, consider
- 8. (a) (b) (c) Figure 8: Examples of a single CT slice (note that printed colours do not match exactly the colours shown on the monitor): (a) the slice dis- playedusinga greyscaleintensity ramp; (b) colours from the original seven-colourexperiment,although the locationand bordersof eachregion are clearly visible, the makeup of certain regions is difficult to determine, compared to; (c) colours from the final seven-colour experiment, which show clearly that the region referenced by the arrow is made up of two different groups of elements
- 9. the small spots of purple at the top of the region in the center of the image. These spots represent areas of high relative density, and are quickly identified as different from the surrounding material. The same information is present in the greyscale image, but the visual system is not immediately drawn to it, since the contrast between the region and its neighbours is not as pronounced. One method of visualizing slice data is to “fly” through the slice stack, rapidly dis- playing individual slices one after another in a movie-like fashion [7]. In this context, we expectusers could more accuratelydetectar- eas with a specific density range using colour rather than greyscale. 5 DISCUSSION This paper has presented a simple method for choosing effective colours for use during data visualization. We accomplish this by controlling colour distance, linear separation of colours, and colour category similarity during colour selection. We found that an isolu- minant slice through a monitor’s gamut yielded up to seven differ- ent colours, any one of which could be rapidly and accurately de- tected, even in the presence of all the others. Each colour needed a minimum colour distance, linear separation, and colour category differentiation to guarantee good performance. One of our future goals is a set of investigations that explicitly measure the tradeoff in observer performance when these effects are varied. This will also allow us to identify the minimum required value for each effect. The perceived intensity of each colour was held constant during ourstudies to eliminate the potentialfor visualinterference. It might be possible to use intensity in a controlled manner to increase the number of colours we can display simultaneously. For example, we might be able to choose multiple groups of colours, with each group being restricted to an isoluminantslice through the monitor’s gamut. A colour group would be assigned to a conceptually related set of attribute values in the dataset to be visualized. Proper division by intensity might allow a user to select individual colours either from within a group (e.g., a dull red element or a bright green element) or across groups (e.g., all red elements). This type of technique is not possible if colours of varying intensity are chosen from the moni- tor’s gamut in an ad hoc fashion. Acknowledgements I would like to thank Dr. Kellogg S. Booth and Dr. James T. Enns for the technical advice they offered during my research. I would also like to thank Ms. Jeanette Lum for coordinating and running the most of the experiment sessions. References [1] BAUER, B., JOLICOEUR, P., AND COWAN, W. B. Visual search for colour targets that are or are not linearly-separable from distractors. Vision Research (1996), (in press). [2] BERGMAN, L. D., ROGOWITZ, B. E., AND TREINISH, L. A. A rule-basedtoolfor assistingcolormapselection. In Proceed- ings Visualization ’95 (Atlanta, Georgia, 1995), pp. 118–125. [3] BIRREN, F. Munsell: A Grammar of Color. Van Nostrand Reinhold Company, New York, New York, 1969. [4] CALLAGHAN, T. C. Dimensional interaction of hue and brightness in preattentive field segregation. Perception Psy- chophysics 36, 1 (1984), 25–34. [5] CAMBELL, W. J., SHORT, JR., N. M., AND TREINISH, L. A. Adding intelligenceto scientificdata management.Computers in Physics 3, 3 (1989), 26–32. [6] D’ZMURA, M. Color in visual search. Vision Research31, 6 (1991), 951–966. [7] HEALEY, C. G., BOOTH, K. S., AND ENNS, J. T. Real-time multivariate data visualization using preattentive processing. ACM Transactions on Modelling and Computer Simulation 5, 3 (1995), 190–221. [8] HEALEY, C. G., BOOTH, K. S., AND ENNS, J. T. High-speed visual estimation using preattentive processing. ACM Trans- actions on Computer-Human Interaction (1996), (in press, scheduled for July 1996 issue). [9] HEALEY, C. G., AND ENNS, J. T. A perceptual colour seg- mentation algorithm. Tech. Rep. TR-96-09, Department of Computer Science, University of British Columbia, 1996. [10] HITT, E. F. Retrofit of flat panel color displays in the F- 16A/B. In ProceedingsIEEE/AIAA 11th Digital Avionics Sys- tems Conference (Seattle, Washington, 1992), pp. 323–328. [11] KAWAI, M., UCHIKAWA, K., AND UJIKE, H. Influence of color category on visual search. In Annual Meeting of the Association for Research in Vision and Ophthalmology (Fort Lauderdale, Florida, 1995), p. #2991. [12] LEVKOWITZ, H., AND HERMAN, G. T. Color scales for image data. IEEE Computer Graphics Applications 12, 1 (1992), 72–80. [13] RHEINGANS, P., AND TEBBS, B. A tool for dynamic explo- rations of color mappings. Computer Graphics 24, 2 (1990), 145–146. [14] ROBERTSON, P. K. Visualizing color gamuts: A user inter- face for the effective use of perceptualcolor spacesin data dis- plays. IEEE Computer Graphics Applications 8, 5 (1988), 50–64. [15] ROGOWITZ, B. E., AND TREINISH, L. A. An architecture for rule-based visualization. In Proceedings Visualization ’93 (San Jose, California, 1993), pp. 236–243. [16] ROSENBLUM,L. J. Researchissuesin scientificvisualization. IEEE ComputerGraphics Applications14, 2 (1994), 61–85. [17] TREINISH, L. A., FOLEY, J. D., CAMPBELL, W. J., HABER, R. B., AND GURWITZ, R. F. Effective software systems for scientific data visualization. Computer Graphics 23, 5 (1989), 111–136. [18] TRIESMAN, A. Search, similarity, and integration of fea- ture between and within dimensions. Journal of Experimental Psychology: Human Perception Performance17, 3 (1991), 652–676. [19] VAREY, C. A., MELLERS, B. A., AND BIRNBAUM, M. H. Judgments of proportions. Journal of Experimental Psychol- ogy: Human Perception Performance 16, 3 (1990), 613– 625. [20] WARE, C. Color sequences for univariate maps: Theory, ex- periments, and principles. IEEE Computer Graphics Appli- cations 8, 5 (1988), 41–49. [21] WARE, C., AND BEATTY, J. C. Using colour dimensions to display data dimensions. Human Factors 30, 2 (1988), 127– 142. [22] WYSZECKI, G., AND STILES, W. S. Color Science: Con- ceptsandMethods,Quantitative Data and Formulae,2nd Edi- tion. John Wiley Sons, Inc., New York, New York, 1982.