Colour categories boundaries are better defined
in contextual conditions
Robert Benavente, C. Alejandro Párraga and Maria Vanrell
Computer Vision Center / Dept. Ciències de la Computació (UAB)
Building O, Campus UAB, 08193 Bellaterra (Barcelona), Spain
{robert, aparraga, maria}@cvc.uab.es
INTRODUCTION
Boundaries between basic colour categories (green, blue,
purple, pink, red, orange, yellow and brown) were measured.
• Colour samples were presented on isolation (presented on a
dark background and surrounded by a white “frame”)
Previous experiment (Parraga et al., 2009)
• Colour samples were determined by a QUEST protocol
Which is the colour • Subjects were forced to categorise colours (yes/no paradigm)
of the patch, green
or blue?
green
Problems: Some boundaries (e.g. green-blue) were very
diffuse and presented bimodal distributions, which were
attributed to the emergence of non-basic categories (e.g.
turquoise).
Boundaries that presented bimodal distributions:
L=36 red-brown, green-blue, purple-pink
L=58 orange-yellow, green-blue, purple-pink, pink-red
L=81 orange-yellow, yellow-green, green-blue, blue-purple,
purple-pink
blue
Colour name frontiers were measured using
a 2AFC paradigm and a Quest protocol
Exemplary set of previous results
Number of instances
The categorisation of colour (or colour naming)
by humans is an important problem both in
visual perception and computer vision.
Green
Turquoise?
Blue
New goal
In a new experiment, colour boundaries
were again measured but this time
colour samples were presented over a
coloured “mondrian” (randomised
colours) background to simulate
context and a new paradigm was used.
NEW EXPERIMENT
Apparatus:
Cambridge Research Systems VisaGe graphics
card, CRT calibrated monitor (Trinitron tube)
inside a dark room, and Logitech gamepad.
Subjects:
8 subjects, (speakers of English as a first
language). Colour vision assessed using the
Ishihara plates and the Farnsworth D-15 test.
Presentation:
Coloured patches on a coloured “mondrian”
background. Patches subtended 5 deg to the
observer. Viewing was unconstrained and
binocular (distance = 156 cm). At the bottom of
the screen there were two colour names.
Task:
Observers operated the gamepad buttons to
select a patch colour that was mid-way between
the colour represented by the names at the
bottom of the screen.
Change the
colour of the
patch to select
the colour midway between
green and
blue…
green
Segmentation of the Lab colour
space in different Luminance
levels: only boundaries in
three levels were measured
(intermediate levels were
interpolated)
blue
Stimuli:
Colour samples were selected using the
boundaries proposed by the parametric
model of Benavente et al (2008) in Lab
space as a guide.
Method:
Observers were allowed to manipulate
the hue of the central patch along arcs
of the same luminance and saturation.
Once a boundary was chosen by the
observer, a new “arch” was selected
across the same colour frontier and a
new random “mondrian” was set for
the next run. After 10 runs, a new
colour frontier was explored. A
complete experiment consisted of 190
runs, exploring 19 colour frontiers (at
three different Lab luminance levels:
36, 58 and 81).
(Sturges and Whitfield, 1995)
Lab colour space
(Sturges and Whitfield, 1995)
RESULTS
Unimodal boundaries
30
Modelling
Green-Blue
L= 36
Blue-Purple
10
5
6.2
8
6.5
3
6.7
9
6.2
8
6.5
3
6.7
9
5.5
3
6.0
3
5.5
3
6.0
3
5.2
8
5.7
8
5.2
8
5.0
3
5.7
8
4.5
2
4.7
8
4.5
2
5.0
3
4.2
7
4.2
7
4.7
8
3.7
7
4.0
2
3.7
7
4.0
2
3.0
2
3.5
2
3.5
2
2.7
6
3.2
7
Green-Blue
3.2
7
2.5
1
1.7
6
2.2
6
2.0
1
1.0
1
1.5
1
L= 58
1.2
6
0.2
5
0.7
5
0.0
0
0.5
0
0
25
Blue-Purple
Purple-Pink
Pink-Red
20
Red-Orange
15
Orange-Yellow
Yellow -Green
10
5
25
3.0
2
2.7
6
Green-Blue
2.5
1
1.7
6
2.2
6
2.0
1
1.0
1
1.5
1
1.2
6
0.7
5
0.0
0
30
0.2
5
0
0.5
0
Number of instances
Red-Brow n
Blue-Purple
L= 81
Purple-Pink
Pink-Orange
20
Orange-Yellow
Yellow -Green
15
10
The figure on the right shows the
results of adjusting the Benavente et al
parametric model according to its
original criteria (first column),
according to the previous (context-less)
boundary measures (second column)
and according to the results shown on
the left (third column). Each row
corresponds to one of the three
reference luminance levels.
5
6.2
8
6.5
3
6.7
9
6.0
3
5.2
8
5.5
3
5.7
8
4.5
2
4.7
8
5.0
3
3.7
7
4.0
2
4.2
7
3.5
2
3.2
7
2.5
1
2.7
6
3.0
2
1.7
6
2.0
1
2.2
6
1.0
1
1.2
6
1.5
1
0.7
5
0.5
0
0.2
5
0.0
0
0
Angle in Radians
Each coloured histogram corresponds to a
colour name boundary (as defined by
Benavente et al) including all observers (8) and
all saturation levels (10). Although the shape of
the histograms varies from boundary to
boundary, it is possible to appreciate from the
figures that most results are concentrated
around a certain mean value (the average
boundary angle)
The parameters of the model were
adjusted so that the width of the
boundary regions (in yellow) matches
the shape of the histograms on the left.
The resulting regions (third column)
show much sharper transitions than in
the case of the boundaries described by
Parraga et al (measured against a dark,
context-less background and using a
different paradigm).
CONCLUSIONS
• A new paradigm was used to measure colour name boundaries in Lab space. This
paradigm allows us to collect larger amounts of data than the previous one
(Parraga et al)
• Results show that frontiers measured using the new paradigm with colours
presented over random mondrian backgrounds are better defined than those
measured on dark backgrounds
• These results open a new question on whether they are due to the new paradigm
or to the insertion of context (random mondrian background).
Parraga et al
Context model
L plane = 36
Brow n-Green
L plane = 58
Pink-Red
15
30
Benavente et al
Purple-Pink
20
L plane = 81
25
REFERENCES AND ACKNOWLEDGEMENTS
Parraga C.A., Benavente, R., Vanrell, M. and Baldrich, R., Psychophysical measurements to model
inter-colour regions of colour-naming space, J. of Im. Sci. and Tech., 53(3), (2009).
Benavente, R., Vanrell, M., and Baldrich, R., Parametric fuzzy sets for automatic color naming. J. of
Opt. Soc. Am. A, 25(10), (2008).
Sturges, J. and Whitfield, T.W.A., Locating Basic Colors in the Munsell Space. Color Research and
Application. 20, 364-376. (1995).
Supported by projects TIN2007-64577 and CSD2007-00018 of the Spanish Ministry of
Science. RB was funded by the “Juan de la Cierva” fellowship (JCI-2007-627) and CAP
was funded by the “Ramón y Cajal” fellowship (RYC-2007-00484). Funded by Porject
Consolider Ingenio 2010.