Improving Color Constancy in the Presence of Multiple Illuminants
using Depth Information
Marc Ebner and Johannes Hansen
Ernst-Moritz-Arndt-Universit
¨
at Greifswald, Institut f
¨
ur Mathematik und Informatik,
Walther-Rathenau-Straße 47, 17487 Greifswald, Germany
Keywords:
Color Constancy, Space Average Color, Depth Map, Color, Kinect.
Abstract:
A human observer is able to judge the color of objects independent of the illuminant. In contrast, a digital sen-
sor (or the retinal receptors for that matter) only measure reflected light which varies with the illuminant. The
brain is somehow able to compute a color constant descriptor from the light falling onto the retina. We have
improved a well known color constancy algorithm based on local space average color. This color constancy
algorithm can be mapped to the different visual processing stages of the human brain. We have extended this
algorithm by incorporating depth information. The idea is that wherever there are depth discontinuities there
may also be a change of the illuminant in the image. Hence, depth discontinuities are used to separate differ-
ent illuminants. This allows us to better estimate the local illumination and allows us to compute an improved
color constant descriptor. We also compute local space average depth to decide locally whether to average data
from retinal sensors uniformly or non-uniformly. We show how our algorithm works on real world scenes.
Depth information is obtained from a standard Kinect sensor.
1 INTRODUCTION
Object color is an important cue in everyday life. We
use it to recognize or distinguish different objects.
However, color is a product of the brain (Zeki, 1993).
The brain somehow computes a color constant de-
scriptor from the data measured by the retinal recep-
tors (Ebner, 2007a). The ability to compute a color
constant descriptor is also very important for artifi-
cial vision systems. In particular, it is very important
in the area of autonomous mobile robotics whenever
robots have to work in several different environments.
In the human eye, the retinal receptors measure
light reflected by objects. Unfortunately, reflected
light varies with the spectral power distribution of
the illuminant. Suppose a white wall is illuminated
by an illuminant with a power distribution having a
maximum in the red part of the spectrum. Hence
the cones with a maximum absorption in the red part
of the spectrum will respond more strongly than the
cones with maximum absorption in the green and blue
parts of the spectrum. Similarly, if we take a digital
camera and take a digital photo of this scene, then
the image will have a strong reddish color cast to it.
The wall will appear red in the image. If the illu-
minant is known, we can compute a color corrected
image of the scene. The scene will then look as if it
had been illuminated by a uniform illuminant. Dig-
ital cameras assume that a single uniform illuminant
(sunlight, cloudy sky, flash, neon light, etc.) is illu-
minating the scene. Hence, a digital camera corrects
for a single illuminant. However, in practice this as-
sumption (that a single illuminant is illuminating the
scene) is not valid. We usually have multiple different
illuminants such as sunlight falling through a window
and artificial light turned on inside the building. Thus,
we need to estimate the illuminant locally in order to
correctly estimate object reflectance, i.e. the percent-
age of incident light which is reflected by an object.
This estimate of object reflectance can then be used
for object recognition as it is independent of the illu-
minant.
A number of different color constancy algorithms
have been proposed (Agarwal et al., 2006; Ebner,
2007a). Most algorithms assume that a single uniform
illuminant is illuminating the scene, e.g. the White
Patch Retinex algorithm or the gray world assumption
(Buchsbaum, 1980). Some algorithms assume that
the illuminant is somehow constrained (Finlayson and
Hordley, 2001). A color constancy algorithm based
on the gray-edge hypothesis has been proposed by
van de Weijer et al. (2007). Apart from the original
133
Ebner M. and Hansen J..
Improving Color Constancy in the Presence of Multiple Illuminants using Depth Information.
DOI: 10.5220/0004743601330140
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2014), pages 133-140
ISBN: 978-989-758-011-6
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Retinex algorithm (Land and McCann, 1971), only a
few work in the context of non-uniform illumination,
e.g. Barnard et al.s (1997) extension of the gamut
constraint algorithm. In practice, one usually has to
cope with a non-uniform illumination.
Most algorithms for color constancy cannot be
readily mapped to the human vision system. Ebner
(2007c; 2012) has proposed a model of human color
perception which can be mapped to the human vision
system. His method also works in the context of non-
uniform illumination. In its original form, this algo-
rithm only uses the output from the retinal receptors
to arrive at a color constant descriptor. It does not use
depth information. This algorithm has been extended
by Ebner and Hansen (2013) to incorporate depth in-
formation. Here, we also compute local space average
depth in order to decide locally whether to average
data from retinal senors uniformly or non-uniformly.
In addition, we better handle uncertainty in the posi-
tion of the detected edges.
Depth information is readily available inside the
human vision system. Gilchrist (1977) has put for-
ward the coplanar ratio hypothesis. According to this
hypothesis, lightness is determined primarily by ratios
within perceived planes. Our research is in line with
this hypothesis. How and if depth cues are actually
used by the human visual system to compute a color
constant descriptor is currently unknown. With this
contribution we explore how depth information may
be used to arrive at a color constant descriptor. For ar-
tificial vision systems, we can obtain depth informa-
tion from a variety of methods (Horn, 1986; Jain et al.,
1995). For our experiments, we have used the Kinect
sensor to obtain a RGB image and the so called depth
map which provides the distance to the corresponding
object point for each pixel of the image.
In Section 2 we briefly explain Ebner’s algorithm
and how it can be mapped to the individual stages of
the human vision system. In Section 3 we explain
how depth information can be integrated into this al-
gorithm. Section 4 describes how we have used the
Kinect sensor to obtain a dense depth map. Section
5 describes the experiments that we have performed.
Section 6 concludes this paper.
2 COLOR CONSTANCY BASED
ON LOCAL SPACE AVERAGE
COLOR
A color constant descriptor can be computed in var-
ious different ways. See Ebner (2007a) or Barnard
et al. (2002) for an overview and evaluation of sev-
eral different algorithms. A quite simple algorithm is
the gray world assumption which has been put for-
ward by Buchsbaum (1980). According to the gray
world assumption, the world is gray on average. This
assumption allows us to compute a color constant de-
scriptor. Using this assumption, we can obtain an es-
timate of the illuminant by simply averaging all pixel
values. Given this estimate, we can compute an out-
put image that is independent of the illuminant. For
the gray world assumption to work, it is necessary that
quite a large number of different surface reflectances
are contained in the scene being viewed.
Ebner (2009) has extended this algorithm to esti-
mate the illuminant locally for each image pixel. He
has also shown how this algorithm can be mapped to
the human vision system (Ebner, 2007c, 2012). The
algorithm runs on a grid of processing elements. It
is assumed that we have one processing element per
image pixel. For each pixel, we have three color
bands in the red, green and blue parts of the spec-
trum. The processing elements are laterally con-
nected to each other. Each processing element es-
timates the illuminant for the corresponding image
pixel by computing local space average color. Let
a(x,y) = [a
r
(x,y),a
g
(x,y),a
b
(x,y)] be local space av-
erage color estimated by processing element at posi-
tion (x, y). Let c(x,y) = [c
r
(x,y),c
g
(x,y),c
b
(x,y)] be
the measured color, i.e. the pixel color of the input
image, at position (x,y). It is assumed that c(x,y) cor-
responds linearly to the irradiance falling onto the im-
age sensor. Let N(x,y) be the neighborhood defined
for the processing element at position (x,y) and let p
c
be a small positive value. The following two update
equations are iterated until convergence:
a
0
i
(x,y) =
1
|N(x,y)|
(x
0
,y
0
)N(x,y)
a
i
(x
0
,y
0
) (1)
a
i
(x,y) =(1 p
c
)a
0
i
(x,y) + p
c
c
i
(x,y) (2)
with i {r,g,b}. The first equation takes local space
average color from neighboring processing elements
and averages it. The current element can also be in-
cluded in this averaging process. The second equation
adds a tiny amount of the measured color to the esti-
mated average.
The parameter p
c
determines the extent of the av-
eraging. If p
c
is rather small, then local space aver-
age color is computed over an extensive area. If p
c
is comparatively large, then local space average color
is computed over a very small area. The parameter
p
c
is usually set such that a sufficiently large number
of image pixels are included in the average, e.g. 30%
of all image pixels. Let ¯p
c
be the desired percentage
of the image over which local space average color is
computed and let s be the maximum of the width and
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
134
the height of the image in pixels, then p
c
is given by
p
c
= 1/(4 ¯p
2
c
s
2
).
Assuming narrow band sensors, then the mea-
sured irradiance is proportional to the reflectance
R
i
(x,y) and irradiance L
i
(x,y) at the images object
point for color band or wavelength i. In other words,
we have c
i
(x,y) = L
i
(x,y)R
i
(x,y) assuming a scaling
factor of 1. Using L
i
(x,y) a
i
(x,y) we can compute
a color constant descriptor by dividing the measured
color c
i
(x,y) by local space average color a
i
(x,y).
In Ebner’s (2012) model, the retinal receptors
measure the irradiance falling into the eye. The reti-
nal receptors have a logarithmic response curve. The
color space is rotated due to color opponent cells be-
fore reaching the visual cortex. Cells in V4 compute
local space average color. This local space average
color is subtracted from the data made available by
cells in V1. Because of the logarithmic response, lo-
cal space average color simply needs to be subtracted
from the data from V1 in order to arrive at a color
constant descriptor.
Within area V1, the visual stimulus is analyzed
with respect to all kinds of different aspects (Living-
stone and Hubel, 1984). Cells have been found whose
optimal stimulus is an oriented line. Other cells’ opti-
mal stimulus is light of a particular wavelength. Cells
usually respond more prominently to one eye or the
other. These cells are grouped in columns which are
called ocular dominance columns. It could be that vi-
sual information is also analyzed with respect to depth
discontinuities in order to improve color perception.
We will explore this possibility in the next section.
3 USING DEPTH INFORMATION
TO IMPROVE COLOR
CONSTANCY
In a natural scene, there are usually many different
illuminants. Sun light may be falling through a win-
dow, a desk lamp may be turned on and at the same
time neon light may illuminate the interior of the
room. If we take an image of the room, the top of the
desk may be illuminated by the light from the desk
lamp while the area below the desk may be illumi-
nated by ambient light which has been reflected mul-
tiple times by the objects contained in the room. If we
look at the desk, we see a depth discontinuity at the
edge of the desk which separates the top of the desk
from the floor.
Figure 1(a) shows another example. The first
room is illuminated by sunlight falling through a win-
dow while the corridor is illuminated by a blueish illu-
minant. For this image, the door frame provides a nice
separation between the two illuminants. The depth
discontinuity at the door frame has been highlighted
manually.
We will now show how we can integrate such
depth discontinuities into the algorithm to compute
local space average color. It does not make sense to
average local space average color across depth dis-
continuities because it is assumed that one illuminant
illuminates the area on one side of the edge while an-
other illuminant illuminates the area on the other side
of the edge. Figure 1(b) shows an estimate of the two
illuminants (computed by our algorithm). The two
illuminants are clearly separated by the door frame.
On the right hand side of the door we have a smooth
illumination gradient. Figure 1(c) shows the output
image which has been computed by dividing the mea-
sured color c(x,y) shown in Figure 1(a) by local space
average color a(x, y) shown in Figure 1(b).
Of course it is also possible that the same illumi-
nant illuminates both sides of an edge. However, this
will do no harm. In order to take depth discontinuities
into account, we only need to use a slightly modified
neighborhood N
d
(x,y) which replaces the uniform
neighborhood N(x, y) in Equation (1). Let d(x,y) be
the depth map. The depth map specifies the distance
from the camera to each object point. We only want
to average across processing elements whose corre-
sponding object points have approximately the same
depth. Hence, we can define N
d
(x,y) as follows
N
d
(x,y) ={(x
0
,y
0
) N(x,y)
with |d(x,y) d(x
0
,y
0
)| ε
d
(x,y)}
(3)
where ε
d
defines the edge threshold assuming that the
depth map has been scaled to the range [0, 1]. We av-
erage across discontinuities smaller than ε
d
(x,y). Dis-
continuities larger than ε
d
(x,y) separate two process-
ing elements in the averaging process.
The threshold ε
d
can be set to a fixed value for
the entire image. However, using a locally varying
threshold may be more appropriate. Hence, we also
compute local space average depth. Local space av-
erage depth
˜
d is computed using the same method we
have used to compute local space average color
˜
d
0
(x,y) =
1
|N(x,y)|
(x
0
,y
0
)N(x,y)
a
i
(x
0
,y
0
) (4)
˜
d(x,y) =(1 p
d
)
˜
d
0
(x,y) + p
d
d(x,y) (5)
where d(x,y) is the depth value at position (x,y) and
p
d
is a small value which determines the extent of
the averaging of the depth map. The parameter ¯p
d
is defined in exactly the same way as the parameter
¯p
c
above. Now that we have computed local space
ImprovingColorConstancyinthePresenceofMultipleIlluminantsusingDepthInformation
135
(a) (b) (c)
Figure 1: (a) Sample image with two illuminants. The depth discontinuity separating the two illuminants has been manually
highlighted in red. (b) Estimate of the illuminant. (c) Color constant descriptor.
average depth, we can make the threshold dependent
on local depth. E.g. ε
d
(x,y) = 0.1
˜
d(x,y) means that
we do not average across depth differences larger than
10% of the average depth in the region.
In practice, the alignment between the depth map
and the color map may not be perfect. A non-perfect
alignment between the depth map and the color im-
age may result in artefacts at regions where the il-
luminant from another nearby region is used instead
of the correct illuminant. That’s why we first com-
pute depth edges. A depth edge is located between
two neighboring points (x,y) and (x
0
,y
0
) if we have
|d(x,y) d(x
0
,y
0
)| > ε
d
(x,y). We dilate the resulting
binary edge image using a square structuring element
of size 5 ×5. The size of the structuring element is
set to the size of the uncertainty in the alignment be-
tween the depth map and the color image. The depth
discontinuity is then assumed to be located inside this
enlarged area at a location where we also have a color
edge with a threshold of ε
c
= 0.1. Because of this op-
eration, depth discontinuities are now in perfect align-
ment with color edges in the image. If we have the un-
likely case of a depth edge between two pixels but no
color edge then our method will average across these
pixels. In the human visual system, it is known that
different aspects such as color, shape and motion are
processed by different visual areas (Zeki, 1993; Zeki
et al., 1991). It may be that these aspects are brought
into alignment by a dynamic process similar to the
one shown by Ramachandran (1993).
Figure 2 compares the two threshold methods for
a real scene. Figure 2(a) shows the input image. Fig-
ure 2(b) shows local space average color computed
with a fixed threshold for the entire image while Fig-
ure 2(c) shows local space average color computed
with a spatially varying threshold as described above.
The delineation of the border between the different
illuminants is more accurate with a spatially varying
threshold.
In the human vision system, binocular disparity
can be used to estimate the distance of object points
relative to the observer. For our experiments, we have
used the Kinect to obtain a depth map for each input
image.
4 OBTAINING A KINECT IMAGE
ALIGNED DEPTH MAP
The Kinect sensor has been developed by Microsoft
for the Xbox 360 video game console (Microsoft Cor-
poration, 2011). It is a sensor which can be used for
motion tracking and also sound position tracking. It
consists of a horizontal bar with a RGB camera, a
depth sensor, a multi-array microphone which rests on
a motorized tilt unit as shown in Figure 3(a). Figure
3(b) shows the RGB image obtained with the Kinect
sensor for a sample scene. The corresponding depth
map is shown in Figure 3(c). A detailed description of
the Kinect sensor is given by Kofler (2011). To date, it
has been used in numerous different research projects.
For instance, Newcombe et al. (2011), have shown
how to perform dense surface mapping and tracking
with the Kinect sensor. Gabel et al. (2012) have used
it for full body gait analysis. A detailed evaluation of
the Kinect sensor for computer vision applications is
given by Andersen et al. (2012).
In order to use this depth map, we need to estab-
lish a correspondence between each image pixel of the
input image and each pixel of the depth map. Color
and depth sensors are a small distance apart from each
other and they do not necessarily point into the same
direction. The intrinsics and extrinsics of the sensors
differ. The depth sensor covers a significantly smaller
area than the color sensor. In addition, the depth sen-
sor outputs data with a non-linear correspondence to
distance.
We align the RGB input image and the depth map
by performing a stereo calibration, i.e. computing
intrinsic and extrinsic parameters of the two cam-
eras and then transform the depth value to distance
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
136
(a) (b) (c)
Figure 2: (a) Input image. (b) local space average color with a fixed threshold ε
d
(x,y) = ε
d
. (c) local space average color with
a spatially varying threshold ε
d
(x,y) = 0.1
˜
d(x, y). In addition, depth discontinuities are aligned with color edges.
XBOX 360
RGB Camera
Depth Sensor
Tilt Unit
(a) (b) (c)
Figure 3: (a) Kinect sensor. (b) RGB image (c) Depth map.
in meters. This approach is also described by Burrus
(http://openkinect.org) and Kofler (2011). The Kinect
sensor is not able to compute depth information for all
pixels due to occlusion. Due to the arrangement of the
infrared camera and the infrared laser which produces
the laser grid for depth computation, the grid may not
be visible for certain areas seen by the camera. This
always happens to the left side of an edge. In order
to accurately detect such edges, we need to compute a
dense depth map from the Kinect output. We do this
by iteratively filling in data from the left hand side.
We call pixels for which the Kinect was able to esti-
mate a depth value “a valid depth value” and we call
all other pixels “invalid depth values”. Before we ap-
ply our algorithm, we filter the depth map by remov-
ing isolated valid depth values which are surrounded
by invalid depth values. These depth values are as-
sumed to be incorrect. We then iterate n
f
times over
the image. Within each row with invalid pixels, we
start from the left hand side and loop over all pixels
with invalid depth values from left to right. Each pixel
with an invalid depth value is updated by interpolat-
ing depth values from the top, upper left, left, lower
left and the bottom side. The values from the top, left
and bottom side use a weight of 1 while the diagonal
pixels from the upper left and lower left use a weight
of 1/
2. We end up with a dense depth map which
we can use for our algorithm.
5 EXPERIMENTS AND RESULTS
The algorithm is tested on a number of different im-
ages. Unfortunately, the Kinect only offers a rela-
tively small field of view. The depth sensor provides
data in the range from 0.8 to 3.5 meters with a depth
resolution of 1cm at a distance of 2m (Andersen et al.,
2012). This constrains the types of scenes that we can
shoot. We have taken care to avoid shiny surfaces,
such as mirrors, polished metals or brilliant varnishes
in the scene. Such surfaces irritate the depth sensor.
For dark scenes, noise can be removed by taking mul-
tiple images and then averaging the output.
The Kinect computes a depth map of size 640 ×
480. The alignment algorithm corrects for the dif-
ferences between the RGB image and the depth map.
Since the RGB image and the depth map are not per-
fectly registered, we obtain a border around the im-
age where depth is undefined. Hence, we crop the
images to size 569 ×428 for further processing. The
parameters were set to ¯p
c
= 0.25, ¯p
d
= 0.1, ε
d
(x,y) =
0.1
˜
d(x,y), n
t
= 15 and ε
c
= 0.1.
Figure 4 shows the results. For comparison Fig-
ure 5 through Figure 8 shows the estimate of the il-
luminant, i.e. local space average color for images 1
through 4, computed with three other color constancy
algorithms: the gray world assumption (GW), stan-
dard local space average color (LSA) (Ebner, 2009),
ImprovingColorConstancyinthePresenceofMultipleIlluminantsusingDepthInformation
137
1 2 3 4
input imagedepth mapedges
ill. estimate
output
Figure 4: Results of our algorithm for 4 sample images.
Depth Map-LSA GW LSA Iso-LSA
ill. estimate
output
Figure 5: Comparison with three other color constancy algorithms for image 1: gray world assumption (GW), local space
average color (LSA), computation of anisotropic local space average color along iso-illumination lines (Iso-LSA).
and computation of anisotropic local space average
color along iso-illumination lines (Iso-LSA) (Ebner,
2007b). None of these methods use depth informa-
tion. When we compare the results we see that depth
information allows us to obtain a better estimate of
the illuminant in the vicinity of depth edges.
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
138
Depth Map-LSA GW LSA Iso-LSA
ill. estimate
output
Figure 6: Comparison with three other color constancy algorithms for image 2: gray world assumption (GW), local space
average color (LSA), computation of anisotropic local space average color along iso-illumination lines (Iso-LSA).
Depth Map-LSA GW LSA Iso-LSA
ill. estimate
output
Figure 7: Comparison with three other color constancy algorithms for image 3: gray world assumption (GW), local space
average color (LSA), computation of anisotropic local space average color along iso-illumination lines (Iso-LSA).
Depth Map-LSA GW LSA Iso-LSA
ill. estimate
output
Figure 8: Comparison with three other color constancy algorithms for image 4: gray world assumption (GW), local space
average color (LSA), computation of anisotropic local space average color along iso-illumination lines (Iso-LSA).
6 CONCLUSIONS
We have shown how depth information can help in
improving illumination estimates. A well known al-
gorithm for color constancy based on local space av-
erage color has been updated to also include depth
information. For our experiments, we have used the
Kinect sensor to provide an RGB image with a cor-
ImprovingColorConstancyinthePresenceofMultipleIlluminantsusingDepthInformation
139
responding depth map. Depth discontinuities in the
depth map are assumed to separate different illumi-
nants from each other. Our algorithm was tested on
several different sample images. Comparison results
with three other color constancy algorithms are also
shown.
REFERENCES
Agarwal, V., Abidi, B. R., Koschan, A., and Abidi, M. A.
(2006). An overview of color constancy algorithms.
Journal of Pattern Recognition Research, 1(1):42–54.
Andersen, M. R., Jensen, T., Lisouski, P., Mortensen, A. K.,
Hansen, M. K., Gregersen, T., and Ahrendt, P. (2012).
Kinect depth sensor evaluation for computer vision
applications. Technical Report ECE-TR-6, Aarhus
University, Denmark.
Barnard, K., Cardei, V., and Funt, B. (2002). A comparison
of computational color constancy algorithms part I
and II. IEEE Trans. on Image Processing, 11(9):972–
996.
Barnard, K., Finlayson, G., and Funt, B. (1997). Color con-
stancy for scenes with varying illumination. Computer
Vision and Image Understanding, 65(2):311–321.
Buchsbaum, G. (1980). A spatial processor model for object
colour perception. Journal of the Franklin Institute,
310(1):337–350.
Ebner, M. (2007a). Color Constancy. John Wiley & Sons,
England.
Ebner, M. (2007b). Estimating the color of the illumi-
nant using anisotropic diffusion. In Kropatsch, W. G.,
Kampel, M., and Hanbury, A., eds., Proc. of the 12th
Int. Conf. on Computer Analysis of Images and Pat-
terns, Vienna, Austria, pp. 441–449, Berlin. Springer-
Verlag.
Ebner, M. (2007c). How does the brain arrive at a color con-
stant descriptor? In Mele, F., Ramella, G., Santillo, S.,
and Ventriglia, F., eds., Proc. of the 2nd Int. Symp. on
Brain, Vision and Artificial Intelligence, Naples, Italy,
pp. 84–93, Berlin. Springer.
Ebner, M. (2009). Color constancy based on local space av-
erage color. Machine Vision and Applications Journal,
20(5):283–301.
Ebner, M. (2012). A computational model for color per-
ception. Bio-Algorithms and Med-Systems, 8(4):387–
415.
Ebner, M. and Hansen, J. (2013). Depth map color con-
stancy. Bio-Algorithms and Med-Systems (accepted).
Finlayson, G. and Hordley, S. (2001). Colour signal pro-
cessing which removes illuminant colour temperature
dependency. UK Patent Application GB 2360660A.
Gabel, M., Gilad-Bachrach, R., Renshaw, E., and Schuster,
A. (2012). Full body gait analysis with Kinect. In
Proc. of the Annual Int. Conf. of the IEEE Engineering
in Medicine and Biology Society, San Diego, CA, pp.
1964–1967. IEEE.
Gilchrist, A. L. (1977). Perceived lightness depends on per-
ceived spatial arrangement. Science, 195:185–187.
Horn, B. K. P. (1986). Robot Vision. The MIT Press, Cam-
bridge, MA.
Jain, R., Kasturi, R., and Schunck, B. G. (1995). Machine
Vision. McGraw-Hill, Inc., New York.
Kofler, M. (2011). Inbetriebahme und untersuchung des
Kinect sensors. Master’s thesis, FH Ober
¨
osterreich,
¨
Osterreich.
Land, E. H. and McCann, J. J. (1971). Lightness and retinex
theory. Journal of the Optical Society of America,
61(1):1–11.
Livingstone, M. S. and Hubel, D. H. (1984). Anatomy and
physiology of a color system in the primate visual cor-
tex. The Journal of Neuroscience, 4(1):309–356.
Microsoft Corporation (2011). Programming with the
Kinect for Windows SDK.
Newcombe, R. A., Izadi, S., Hilliges, O., Molyneaux, D.,
Kim, D., Davison, A. J., Kohli, P., Shotton, J., Hodges,
S., and Fitzgibbon, A. (2011). Kinectfusion: Real-
time dense surface mapping and tracking. In Proc. of
the 10th IEEE Int. Symp. on Mixed and Augmented
Reality, pp. 127–136. IEEE.
Ramachandran, V. S. (1993). Filling in gaps in perception:
Part II. Scotomas and phantom limbs. Current Direc-
tions in Psychological Science, 2(2):56–65.
van de Weijer, J., Gevers, T., and Gijsenij, A. (2007). Edge-
based color constancy. IEEE Trans. on Image Pro-
cessing, 16(9):2207–2214.
Zeki, S. (1993). A Vision of the Brain. Blackwell Science,
Oxford.
Zeki, S., Watson, J. D. G., Lueck, C. J., Friston, K. J.,
Kennard, C., and Frackowiak, R. S. J. (1991). A
direct demonstration of functional specialization in
human visual cortex. The Journal of Neuroscience,
11(3):641–649.
BIOSIGNALS2014-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
140