The Dress: a brief analysis of colors staticsticsI

Hossein Baktasha , Mohammad Amin Nazerzadehb , Kamyar Maroufic

a
CE Department, Sharif University of Technology, 94109119
b
CE Department, Sharif University of Technology, 94102263
c
CE Department, Sharif University of Technology, 94105139

Abstract
”the dress” started as a social network phenomena, as a result of argument on it’s perceived color. To perform solid analysis
on such ambiguities, more than one sample like ”the dress” is needed, but no other sample has been found yet. In this project
we have tried to produce another similar stimulus by manipulating other photos. our algorithms are based on some initial
analysis on ”the dress”’s color space and performing clustering and random color generation based on such clusters. we
hypothesize that the ambiguities are a result of changing photo’s lighting (Value in HSV system) and we have produced some
algorithms that could result in very satisfying stimulus.

Keywords: the dress, RGB space, color clustering, HSV analysis

1. Introduction 2. Results
The dress is a photograph taken on 26 February 2015
2.1. color space analysis
became a social network phenomenon that viewers dis-
agree on the color of the clothing item depicted. some Our approach towards tackling this problem was ana-
stated it as black and bluish while others called it white lyzing colors of ”the dress” photo. As the first question
and gold. It has been a debate since then and several coming up to any person who’s thinking about this prob-
researches have been done on the problem. lem would be ”what are the real colors of this dress?”.
In a free-response survey (N = 1401) including those The first step was to analyze the color of all pixels in
naive to the image, reported white/gold or blue/black the RGB space; x,y and z axis respectively showing the
but some said blue/brown(1). Reports of white/gold red, green and blue density of the corresponding pixel.
over blue/black were higher among older people and The hypothesis was that positioning of colors in RGB
women. some subjects reported a switch in perception. space would be highly correlated with people’s percep-
They hypothesized that some people favor a cool illu- tion of the dress color, even though this approach puts
minant (blue sky), discount shorter wavelengths, and aside many other statistics that might have been very
perceive white/gold; others favor a warm illuminant (in- important; such as positioning of pixels with respect to
candescent light), discount longer wavelengths and see each other. After plotting pixel colors in RGB space,
blue/black. Some other researchers stated that the dif- a plane is observed, which all the points are lying on
ferences between two groups arose mainly from differ- it. The plane coordinates were calculated using linear
ences in lightness, rather than chromaticity of the color regression on ”the dress” pixel colors. This plane, as
they adjusted to match the dress(2). They say that shown in figure 2, includes the 2 white and black points
the ambiguity arises because the distribution of colors of the RGB space, (0,0,0) and (255,255,255), and the line
within the dress closely matches the distribution of natu- connecting these two corners of RGB cube. Our first
ral daylights. Another group conclude that the asymme- hypothesis was that this plane would be the key to am-
tries in blue-yellow color perception is the main cause of biguity of ”the dress”. So our attempt was to change
the ambiguity(3). some photos using this plane. In order to cause ambi-
We are going to investigate the main attributes important guity, we tried projecting RGB of every point of a picture
in such ambiguity as it follows. Our main result was to onto our ”special plane”. The produced samples were
produce at least one other picture with such ambiguity not satisfying, since no ambiguity was detected in sub-
or a spectrum of those. jects perceptions. (figure 3 is one of the produced sam-
ples)
I Instructor:
Dr. Ali Ghazizadeh, EE Department Sharif University of In our observations we realized that replacing G and B
Technology
value of pixels would generate ambiguous results (also
Email addresses: hossein.bktsh@gmail.com (Hossein Baktash),
manazerzadeh@yahoo.com (Mohammad Amin Nazerzadeh), mentioned in previous researches). This kind of trans-
kmaroufi76@gmail.com (Kamyar Maroufi) formation is actually a rotation in RGB space. This gave
Preprint submitted to Project Report-Foundations in Systems and Computational Neuroscience July 1, 2018
 Figure 1: the dress pixels colors drawn in RGB space

Figure 3: the plane fitted to colors in RGB space

and in any normal lighting condition, a spectrum rang-

ing from white(light source) to black(fully shaded) is ex-
pected to exist.
Second deduction is that depending on how many col-
ors exist in a photo with normal lighting, the spectrum
mentioned above will be attracted to the those colors
and form clusters around them. E.g. a black and white
photo will form a simple line from (0,0,0) to (255,255,255)
with no curves and the same photo with a yellow fil-
ter, would show a cluster of points around yellow that
spreads slowly to these two corners of RGB cube. More
the diversity of colors, more the number of clusters. So
the fact that points lay on a plane is a special attribute of
”the dress” considering the fact that it has 3 or 4 main
different colors (black, gold, blue, white) and they could
have easily drawn the points away from being a plane
Figure 2: the plane fitted to colors in RGB space
but they haven’t.

us the idea of rotating our ”special plane” with respect

to R or G or B axis. The photos generated by projecting 2.2. color abundances
a new image on the plane and rotating the plane, were So the plane is a valuable observation, but we needed
not ambiguous. but some photos generated by rotating to take in consideration some of the statistics that we
the dress itself, showed a little ambiguity. This lead us had previously removed. The next step was to take into
to the hypothesis that either the plane is not as special account the abundance of the colors. To do so, we used
as we thought, or we have removed a valuable statistics K-means clustering algorithm in RGB cube; using 1000
from ”the dress” by only analyzing the pixels colors. centroids and iterating the algorithm 5 times. In the end
To check whether the plane is important in causing am- of the algorithm we assign a weight to each centroid
biguity or not, we plotted color of some other photos pix- which shows the number of points that are clustered
els in the RGB cube (figures 4 and 5). Observing these around this centroid. After this, the weight is used for
plots, lead us to two deductions: plotting the centroids and only the centroids are plotted
First deduction is that any photo having a light source instead of all pixel colors (figure 6); Size of a centroid in
and some shadings would look like a spectrum ranging space is relative to its weight. With this new plotting
from (0,0,0) to (255,255,255) in RGB cube. which makes algorithm new observations were made. Points seem to
sense, since (0,0,0) is black and (255,255,255) is white be forming 4 clusters around 4 points with colors blue,
ii
Figure 4: a photo with natural lighting and it’s color spectrum. the color
Figure 5: a digital art picture with normal lighting but too many colors acting
spectrum is ranging from (0,0,0) to (255,255,255)
as attractors and color spectrum has drifted very far from a usual spectrum

black, white and yellow(gold). This a very important ob- mated by a normal distribution in RGB space. Next ap-
servation, since these clusters are showing the colors proaches were based on this to produce the desired pic.
that people argue about, and some subjects seem to be
giving more attention to some clusters and neglecting
others. Our next approach for creating a new sample 2.3. final algorithm and results
was altering color of photo’s pixels using these clusters. After using the clustering algorithm , we used a filter-
We fit normal distributions to the clusters. Then a pixel’s ing algorithm to make small changes on the newly gen-
color in the new photo was replaced by a random point erated photo. These changes include reducing noise by
generated by one of the fitted normal distributions; the selecting cluster means as filtering points and getting
one with the closest mean to our point. the colors closer to desired ambiguous colors.
Some samples were generated using this method and The filtering algorithm: a set of points X is given with a
the results were more similar to ”the dress” in terms of set of filtering points F and constant factor alpha. for
color distribution, but because of randomness of algo- every x in X, x is moved closer to the closets f in F,
rithm, they were so noisy. In order to generate smoother with ratio alpha(alpha = 0 means no change and alpha=
photos the algorithm was improved. Lets call the point 1 means every point replaced by a point in F).
that we want to change, X. And mean of the closest Using this filtering method and the last algorithm, some
cluster is A. In the new algorithm a random point using samples were generated. One of these samples is ”the
the corresponding normal distribution is generated and dress” itself after being changed by the algorithm, which
then is projected on AX line. This algorithm provides the held on to its ambiguity, even after considerable change
randomness we need while the new point being closer to (figure 8). Hence, using this method retains all the
it’s previous position. the photos generated by this al- statistics of color space, needed to cause an ambigu-
gorithm were much smoother and our later approaches ity like ”the dress”. But we have neglected all the posi-
were all based on this algorithm. Clusters were esti- tional statistics of pixels. Since generating a photo from
iii
 Figure 7: photo generated with the random algorithms using ”the dress”
clusters after fitting normal distribution to the clusters.

Figure 6: color of ”the dress pixels plotted using centroids of k-means out- Figure 8: running the algorithm on ”the dress” itself. ambiguity still remains
put. the size of each point shows abundance of colors in that area. even though many details of color space are lost and the photo isn’t as
smooth as the original.

scratch ( i.e. coloring all the pixels manually) is very

10). In order to make new samples hue(H) of the last
hard and time consuming, our next approach was to ap-
generated photo was tested for some values between 0
ply our algorithm on many photos and hope for some
and 1, and the same for saturation(S) value and all of
ambiguity! The most promising generated sample
their combinations. The generated photos were not as
is (figure 9). Other samples were showing slight signs
ambiguous as the first one (figure 9) and the slight am-
of ambiguity but they were suspected to be caused by
biguities that were seen in some of them, were mostly
linguistic difficulties in naming the colors. Figure 9 was
due to linguistic issues with naming the colors.
described by some subjects as green-brown and yellow-
redish by others. The test was carried out on 20 sub-
jects and ratio of two kind of answers was 15:5, with the 3. Discussion
green-brown group as the bigger group.
Further samples were generated to check whether the Our analysis is focused on color space of ”the dress”.
only colors causing ambiguity are the ones seen in the Pixel colors are divided into clusters and the clusters
dress and used in the last generated sample. To check play an important role in the ambiguity, since the color
this, HSV of pixels were used instead of RGB, since that subjects see are categorized into a subset of these
changing HSV makes more sense with our natural in- clusters.
tuition of properties of colors. Note that there is a cor- Some of the previous researches indicated that the light-
respondence between these two coloring spaces (figure ing and shading of ”the dress” is the main reason of
iv
Figure 9: a promising result! some subjects seeing green-brown and others
yellow-redish.

Figure 10: color of pixels plotted in HSV space using k-means clustering.
the observed ambiguity(2). In our analysis, the clus-
clusters can be seen in the HSV cylinder. the difference in the bottom and
ters seen in the color space, can be replaced by each top pair of clusters is in their Value (lighting). the points lie on a plane
other (e.g. black-blue cluster moving to white-yellow perpendicular to the cylinder’s top face. different colors can be produced,
cluster) by changing the lighting of the colors; i.e. mov- while keeping same positioning of points in the cylinder, by changing the
hue of all points( i.e. rotating the observed plane around cylinder’s Value
ing the colors along the line connecting (0,0,0) and
axis).
(255,255,255). This can also be seen as increasing or de-
creasing the Value of points in HSV color space, which
was the method we used in our last algorithm.
We hypothesize that the ambiguity is a result of subjects
changing the lighting (Value in HSV system) of photo
and the photo should not have too many clusters in or-
der to cause ambiguity. Smoothness is also an impor-
tant factor, since it could lead the subjects into thinking
about lighting. We believe that using our algorithm on
more (well-chosen) photos can result in photos as am-
biguous as ”the dress”. We couldn’t consider pixel lo-
cations in our analysis and algorithms, so any approach
combining our algorithms with statistics of pixel loca-
tions should produce very satisfying results.
[1] e. a. Lafer-Sousa, Rosa, Striking individual differences in color percep-
tion uncovered by ’the dress’ photograph, Current Biology vol. 25 (no.
13). doi:10.1016/j.cub.2015.04.053.
[2] e. a. Gegenfurtner, Karl R., The many colours of ’the dress’, Current
Biology vol. 25 (no. 13). doi:10.1016/j.cub.2015.04.053.
[3] e. a. Winkler, Alissa D., Asymmetries in blueyellow color percep-
tion and in the color of ’the dress’, Current Biology vol. 25 (no. 13).
doi:10.1016/j.cub.2015.05.004.

Figure 11: subjects arguing between green and blue. but both groups actu-
ally seeing the same color (deduced by using color pallet).