Disagreeing about Crocs and socks: Creating profoundly ambiguous color displays

Abstract

There is an increasing interest in the systematic disagreement about profoundly ambiguous stimuli in the color domain. However, this research has been hobbled by the fact that we could not create such stimuli at will. Here, we describe a design principle that allows the creation of such stimuli and apply this principle to create one such stimulus set - “the crocs and socks”. Using this set, we probed the color perception of a large sample of observers, showing that these stimuli are indeed categorically ambiguous and that we can predict the percept from fabric priors resulting from experience. We also relate the perception of these crocs to other color-ambiguous stimuli - “the dress” and “the sneaker” and conclude that differential priors likely underlie polarized disagreement in cognition more generally.

Full text

Disagreeing about Crocs and socks: Creating profoundly ambiguous color displays
Pascal Wallisch & Michael Karlovich
Department of Psychology and Center for Neural Science, New York University, 4 Washington
Place, New York, NY 10003, USA.
Corresponding author: Pascal Wallisch - pascal.wallisch@nyu.edu - @pascallisch
Abstract: There is an increasing interest in the systematic disagreement about profoundly
ambiguous stimuli in the color domain. However, this research has been hobbled by the fact
that we could not create such stimuli at will. Here, we describe a design principle that allows the
creation of such stimuli and apply this principle to create one such stimulus set - “the crocs and
socks”. Using this set, we probed the color perception of a large sample of observers, showing
that these stimuli are indeed categorically ambiguous and that we can predict the percept from
fabric priors resulting from experience. We also relate the perception of these crocs to other
color-ambiguous stimuli - “the dress” and “the sneaker” and conclude that differential priors
likely underlie polarized disagreement in cognition more generally.
Keywords: Disagreement, Dress, Color, Perception, Ambiguity, Uncertainty, Priors
Acknowledgments: We would like to thank Cecilia Bleasdale for taking the original dress picture
and anyone who either helped us to get the data, in particular Cecilia Hua and anyone else who
donated their data, shared the link or both.
If you want to contribute your data to a followup project, you can do so here:
https://www.surveymonkey.com/r/crocPerception
Author contributions: PW & MK conceived and designed the study. MK & PW created the
stimuli. PW designed the survey. PW & MK recorded the data. MK analyzed the stimuli. PW
analyzed the data. PW & MK wrote the manuscript.
Competing financial interests and commercial relationships: The authors declare no competing
financial interests, nor any commercial relationships. In particular, we want to proactively affirm
that there was no coordination or communication of any kind with Crocs Inc. or any affiliate,
regarding any aspects of this study. We chose our stimulus materials solely on the basis of color
considerations. We also would like to emphasize that this study was entirely self-funded.

Introduction
Strong perceptual illusions - systematic mismatches between the conditions in the external
world and subjective experience - have been known for a long time. Notoriously, prolonged
exposure to a moving stimulus leads to the vivid subjective experience of motion in the opposite
direction once the stimulus objectively stopped moving, a phenomenon known as the motion
aftereffect (Purkinje, 1825). Similar phenomena exist in the color domain - for instance, the
perceived hue of a colored patch can be shifted dramatically from what it would appear in
isolation, if presented with a suitably patterned background (Shevell & Monnier, 2003) or under
carefully modulated lighting conditions (Lotto & Purves, 1999), suggesting that observers take
both background and illumination into account when judging the appearance of a colored patch.
Importantly, all of these illusions are unidirectional, as they seem to affect all observers similarly,
predictably shifting the percept in the same direction.
This state of affairs changed in February 2015, when a categorically ambiguous stimulus in the
color domain surfaced. Most observers perceived this stimulus - the photo of a wedding dress -
as either black and blue (the appearance of the physical dress under sunlight) or as white and
gold, which could not be attributed to problems with fundamental color vision or a misapplication
of color labels (Lafer-Sousa et al., 2015). The reason this stimulus - “the dress” - is considered
categorically (or profoundly) ambiguous is owed to the fact that different observers perceive
different kinds - not degrees - of hue, and these hues are on opposite sides of the color wheel.
Notably, this is not a bi-stable stimulus - only a few observers (on the order of 1%) - experience
a percept that switches back and forth, akin to genuinely bistable stimuli like the Necker cube
(Rubin, 1921), binocular rivalry (Levelt, 1968) or moving plaids (Hupe & Rubin, 2003), all with
characteristic and known perceptual dynamics. Instead, most observers seem to be “locked in”
to a particular interpretation of the image, consistent with the notion that accumulated
experience - in the form of an illumination prior - drives the perception of “the dress” stimulus
(Witzel et al., 2017). Indeed, the specific kind of life experience that could shape the formation
of differential lighting priors, such as differing degrees of exposure to sunlight predicts both the
perception of “the dress” stimulus as well as the readiness of observers to assume particular
kinds of illuminations of the “the dress” (Wallisch, 2017a). However, while being able to explain
the perception of a specific stimulus display in a coherent and meaningful way is compelling, it
is also possible that we just managed to shoehorn an “explanation” of a very specific stimulus,
without being able to generalize. It would be reassuring if the principles invoked in the
explanation of such stimuli allowed us to create novel ambiguous displays at will. Should this
succeed, we could have more confidence that it is really these principles governing the
perception of such displays. Alas - to date - all other categorically color ambiguous display that
have surfaced since “the dress”, including the Adidas Jacket (Wallisch, 2016), the flip-flops
(Wallisch, 2017b) or the sneaker (Wallisch, 2017c) were not generated purposefully with these
principles in mind, but rather arose spontaneously on the internet, allowing for residual doubt
whether we truly understand this phenomenon to persist. Here, we propose and use principles
that allow for the creation of categorically ambiguous color displays and explore how differential
priors determine the subjective perceptual experience evoked by these displays in human
observers.

Method
Stimulus design
We designed our stimuli according to the following principle, guided by the notion that patterned
backgrounds and lighting conditions powerfully influence the perception of a color display.
Specifically, we first maximize uncertainty by using a monochromatic object that could - in
principle be almost any color. Here, we used two versions of “Classic crocs™”, namely
“Ballerina pink” and “New mint”, out of the 28 different colors on offer. We then put these crocs
on an unpatterned, black background to make the display devoid of visual cues which could
allow observers to calibrate their percept (Land & McCann, 1971). Next, we carefully titrated
complementary light to equate the light reflected off the croc to shades of grey.
A final step relies on another piece of garment - in this case the white socks - to mirror the
properties of the illumination source. Thus, we introduce the concepts of a colored target object
(TCO) - this is the percept that we will ask our objects to judge, corresponding in this case to the
color of the crocs. We also introduce the concept of an anchoring or cueing object, which
reflects the color of the illumination, hence illumination mirroring object (IMO), in this case the
socks, see figure 1.
Figure 1: A schematic of the stimulus design principles. Top row: colored target objects (crocs)
under typical lighting conditions. We define sunlight or bright indoor light (e.g. incandescent or
LED) as “typical” lighting conditions, as these are conditions observers can be expected to
experience in everyday living. Bottom row: The same crocs under altered lighting conditions,

carefully titrated to equalize the reflectance. In this view, the crocs can be conceived of as a
filter. Left panel: Pink crocs as the target object. Right panel: Mint crocs as the target object.
Note that this schematic is highly stylized, with arbitrary units of power (AU) on the y-axis. In
reality, the power of the illumination is continuous, not discrete. In addition, the power of “white”
light is actually not equal across wavelengths; wavelengths in the middle of the visible range are
perceived as much brighter by human observers, due to human spectral sensitivity (Wald,
1964). A: Pink crocs, illuminated by white light appear pink because the croc filters wavelengths
in the middle of the range. B: Mint crocs, illuminated by white light appear mint green because
the croc filters wavelengths in the short and long range. C: Pink crocs (the CTO), illuminated by
light with more power in the middle range, as reflected by the appearance of the IMO, appear as
some shade of grey. D: Mint crocs (the CTO), illuminated by light with more power in the short
and long range, as reflected by the appearance of the IMO appear as some shade of grey. Note
that in all cases, the reflectance of the white sock corresponds to the illumination, mirroring the
properties of the light.
This part of the display - the IMO - is also subtly color ambiguous. Whereas this kind of plain
tube sock is likely to be white - as they usually are, when experienced in everyday life - the
person could possibly be wearing colored socks, in which case the socks provide no information
about the lighting. In other words, we use an object that could be any color, but usually isn’t -
like plain tube socks to serve this mirroring function to create a conflict between color
appearance and typically experienced color. We predict that some people - those who have a
strong prior that socks like these are white - will use this cue to disambiguate the display and
discount the illuminant, but others - those who don’t perceive the socks as white - to take the
appearance of the socks at face value. This is in line with other research suggesting that “priors”
- expectations and assumptions about the properties of stimuli that are typically derived from
experience - play a central role in cognition (Körding & Wolpert, 2004; Stocker & Simoncelli,
2006; Griffiths et al., 2008). To summarize, we predict that these principles will yield a
categorically color-ambiguous display. We conceive of this design principle as “Substantial
Uncertainty combined with Ramified or Forked Priors and Assumptions yields Disagreement”
(SURFPAD). We used the following seven images of crocs and socks as stimuli in this
research, see figure 2.

Figure 2: The crocs and socks images we created. A: crocs and Socks under typical (a
combination of daylight coming through a window and a LED (Feit electric, 60 Watt equivalent,
5000K, 800 Lumen) lighting conditions. WP: “Ballerina pink” crocs. WM: “New mint” crocs. B: The
“New mint” crocs under colored, specifically pink/orange light, yielding two specific stimuli with
distinct properties, as reflected by the socks, M1 and M2. C: The “Ballerina pink” crocs under
colored, specifically green light, yielding three specific stimuli with distinct properties, as
reflected by the socks, namely P0, P1 and P2. The lighting conditions in B and C were created
via a dimmable SMD 5050 LED light strip under blackout conditions and on black LifeSpanⓇ
treadmill matting. Note that the granularity of the dimming settings of the light strip in
combination with the reflectance properties of the crocs (e.g. the “New mint” version was
reflecting more strongly in the short wavelengths than anticipated) did not always allow us to
completely “whiten” the light reflected off the crocs. In those cases, we applied minimal digital
filtering with Adobe® Photoshop® CC 2019 software to close the distance. However, we strongly
believe that the use of a more sophisticated lighting system that allows for a sufficiently fine
titration of the different light channels would obviate the need for any digital post-production.
We first created P0 as a proof of concept. We then created two sets of pairs of stimuli using both
pink and mint crocs - and the same socks - under different lighting conditions, P1/P2 and
M1/M2, respectively.
We combined these croc displays we created with two others - “the dress” and “the sneaker”,
both recognized as categorically ambiguous color displays, to use as stimuli in our study. Figure
3 shows the entire set, in combination with gamuts of the individual pixels in hsv space.

Figure 3: The complete set of stimulus types used in this study. In all cases, we show - for
stylistic reasons - a single croc and sock in the top of the croc displays to symbolize the stimulus
type. In the study, we used the complete images as depicted in figure 2. The bottom row for
each stimulus type depicts all pixels from both crocs and socks (but without black background)
as clouds in 2d projections of hsv space, in which individual pixels are represented as colored
dots. The axes of hsv space are oriented as depicted in the inset (to the right of M2 in B). These
pixel clouds were created using MATLAB’s “colorcloud” function. In general, the parts in the dot
cloud that correspond to the part of the image (croc or sock) are aligned. For example, in P1 the
pixels from the sock are to the right and clearly separated from those of the croc. However, this
separation is not always as clear, e.g. in M2, where the point clouds blend together, but in
general the pixels corresponding to the sock are to the left of those corresponding to the croc.
A: “Ballerina pink” (Wp) and “New mint” (Wm) crocs under typical light, as in figure 2A. B: Wm
crocs under colored light, as in figure 2B. C: Wp crocs under colored light, as in figure 2C. D:
“Sneaker” and “Dress” displays and their corresponding colored pixel clouds. The “Dress” image
is courtesy Cecilia Bleasdale.
Participants
We logged data from participants online between July 5th and August 5th 2019. These were
recruited via social media, specifically Twitter and Facebook. The prompts on Twitter were: “If
you have 5 minutes, want to help us out to get to the bottom of why some people see "the
crocs" differently and want to see some pretty striking images check this out:” and “What colors
do you see, when looking at these #crocs? Last call, if you still want to donate your data. Should
only take a couple of minutes.”, with a similar prompt on Facebook: “If you have 5 minutes, want
to help us out to get to the bottom of why some people see "the crocs" differently and are
curious to see some striking ambiguous images, donate your data by taking this brief survey.”
These invitations yielded data from 5,944 participants. Of these, 108 participants indicated that
they were not taking the survey seriously and 80 participants reported a color vision impairment.
Thus, we retained data from 5,762 (~97%) participants for further analysis. Of these, 938
(16.3%) identified as male, 3,764 (65.3%) as female and the remainder did not state a gender.
The reported median age in our sample was 32 years. All study procedures followed the
principles of the Declaration of Helsinki, but were - according to the “final rule” (Revised
Common Rule) - exempt from review by the New York University Institutional Review Board
(UCAIHS).
Survey
We deployed a survey with 18 questions on surveymonkey.com. Nine of these questions
pertained to the perception of the stimulus displays. Of these, seven were the crocs and socks
displays shown in figure 2. One was “the dress” stimulus and one was a display of “the
sneaker”, shown in figure 3D. Response options to the croc stimuli were “Pink/salmon”,
“Green/mint”, “Grey/White”, “Orange/Beige” and “Other, please specify”, ordered randomly for
each question. The response options for “the dress” and “the sneaker” stimuli were the standard
“white/gold”, “black/blue”, “blue/gold”, “white/black”, “switching” and “other” as well as
“pink/white”, “grey/green”, “pink/green”, “grey/white”, “switching” and “other”, respectively.
Importantly, as observers could use the other images to calibrate their perception, each image
was shown in isolation, on a separate page, with participants unable to go back to a previous

page. Participants were prompted with the question: “What color are the crocs in this image?”
for croc images, and: “What is the color of the dress in this image?” as well as: “What color is
the shoe in this image?” for “the dress” and sneaker images, respectively.
The other questions covered whether observers saw the socks as white, how familiar they are
with socks and crocs, what they believed about the source of the illumination and how confident
they are in the veracity of their beliefs in general. We also asked demographic questions about
age and gender as well as whether observers had a color vision deficit and whether they took
the survey seriously.
Data analysis
All data were analyzed with MATLAB (Mathworks, Inc., Sherborn, MA). Specifically, we use
Monte Carlo methods with 100,000 repetitions to bootstrap confidence intervals (Efron &
Tibshirani, 1994). To guard against potential false positives stemming from multiple
comparisons, we adopted a conservative confidence interval of 99% for all analyses. Because
our data stems from voluntary online survey participation, it is not uncommon that participants
did not respond to some questions, e.g. on age or gender. We have no reason to assume that
these omissions are systematic, so we adopt a column-wise approach to handling missing data.
We report dependent variables in terms of proportions of observers who report a particular
percept. As our stimuli are often ambiguous, we report the proportion of our sample with a
percept that corresponds to the “Color Appearance Under Typical IlluminatiON” (CAUTION).
This avoids the awkwardness of reporting the proportion of participants who report a “veridical”
percept, as this concept is ill-defined in the color domain. Veridical pertains to correspondence
with reality, but - as has often been noted - while this makes sense for quantities of space and
time, the qualitative experience of color is largely generated by the brain and experienced by the
mind (Jackson, 1982). Photons have wavelengths, but not colors, and the wavelength of
photons is only loosely related to the color experienced by the human or animal observer. Here,
“typical illumination” means illumination conditions that observers are likely to encounter in
everyday living, such as daylight - when outdoors - or lighting sources designed to mimic
daylight when indoors.
Results
Are images of these crocs perceptually ambiguous under normal (“white”) lighting conditions?
The first question that usually arises when discussing profoundly color ambiguous stimuli is
whether the disagreement indicates that one of the observers suffers from a color vision deficit.
In addition, we have no control over the particular viewing conditions such as viewing distance,
screen type, screen calibration, color temperature and the like, which one would be able to
control when doing such experiments under laboratory conditions. We cannot even be entirely
sure that participants know what a “croc” is or that an image was delivered each and every time,
for every participant, given that server uptimes are not perfect. In addition, one could reasonably
be skeptical that observers are reliably reporting their percepts in the first place, as we are not
paying them nor know their identity nor have any other interactions of any kind.
Thus, it is important to establish that participants can reliably discern and report the color of the
croc intended by the manufacturer (pink vs. mint) and that they agree with each other in spite of
all of these considerations and possibilities for discrepant reporting of percepts.

Figure 4: Perception of croc color
under typical lighting conditions. A:
Percentage of sample reporting
categorical percept on the x-axis in
response to Wp. x-axis: Color labels.
For reasons of economy, we
abbreviate “pink or salmon” as
“Pink”, “green or mint” as “Green”,
“grey or white” as “Grey” and “beige
or orange” as “Beige”, here and in all
other figures. B: As in “A”, but for WM.
The insets to the left of the y-axes
depict the stimulus that evoked the
response.
We report the results of asking participants about the crocs under typical lighting conditions (WP
and WM) in figure 4.
Remarkably, despite all kinds of theoretical possibilities for misunderstandings and for things to
go wrong - whether due to technical issues or
participant characteristics of any kind or even
interactions between them, about 98% of all
participants who provided a response have no
problems discerning the manufacturer intended color
of the crocs as pink/salmon and green/ mint
respectively, from an image served over the internet,
and agreeing with each other about the category of
the color.
The exact p-value for either outcome obtained with
resampling methods is 0, assuming that people used
the available color categories at random. At no point
did the extremes of the null distribution come
anywhere close to the empirical result. Thus, we
conclude that the color of these crocs is not
ambiguous to our participants under typical lighting
conditions.
Are images of these crocs perceptually ambiguous
under altered lighting conditions?
As we saw in figure 4, just putting the crocs on an
unpatterned black background does not render them
perceptually ambiguous with respect to color, so the
question remains whether they do become ambiguous
under suitably altered lighting conditions - along the
lines we laid out in the Method section.
The results are presented in figure 5. Briefly, we
managed to create five categorically color ambiguous
stimuli with different perceptual balancing points.
First, we created P0 as a proof of concept. We then
created two pairs of stimuli from each colored croc. All
5 images are categorically ambiguous with respect to
color - the modal percept for P0 and P1 is grey (43%
and 63%, respectively), for P2 green (40%). The M
percepts are dominated by beige or orange percepts
at 69% (M1) and 61% (M2), respectively. Overall, P2
was the most evenly balanced stimulus - therefore, P2
can be genuinely considered as having four distinct interpretations that are shared by more than
10% of the sample. Also, we note that the available color labels exhausted the space of
possibilities well - less than 1% of responses fell into “other” responses.

The chance of observing this pattern of results - using Chi-squared tests with the proportions
expected from the croc color perception under normal lighting perceptions is low, with a p-value
of below 1e-300 in each of the five cases. Overall, M1, M2 and P1 are conceptually most similar
to “the sneaker” (modal percept of grey in about 63% of observers) and “the dress” (modal
percept of white/gold in about 60% of observers). Therefore, these stimuli can be considered as
bi-stable with respect to the population (not individual observers) and P0 and P2 can be
considered tri- and quadra-stable with respect to the population, respectively, in contrast to the
unambiguous Wp and Wm.
Are people responding randomly, because participants realize that illumination is not normal?
Whereas these results are striking at face value, it is conceivable that lighting conditions are so
obviously unsuitable that participants cannot reasonably be expected to report their actual
percept. Instead, they might be guessing as to what the color of the croc might have been. Such
a response could be akin to a situation under low light conditions, when vision is dominated by
rod-responses, as rods are more light-sensitive than cones, rendering vision effectively
monochromatic (Stabell & Stabell, 2002). However, we don’t think this is plausible, for several
reasons. First, participants were not shy to point out problems if they arose, and had the option
to skip questions. Yet, both the “other” category is almost un-utilized, and most people did
respond to the questions. Second, close to 50% of participants did not suspect anything to be
Figure 5: Perception of color ambiguous stimuli created with altered lighting conditions. Rows,
from top to bottom: P0, P1, P2, M1, M2, in the order in which we created them. Y-axis:
Percentage of sample reporting this percept. X-axis: Categorical color percept. At the beginning
of each right, we depict the stimulus that evoked the observer response.

amiss with the lighting in the images - guessing that they were either taken under sunlight or
indoor lighting, i.e. typical lighting conditions. Finally, and most compellingly, the responses
between stimuli are closely related - if someone perceived one image in a certain way, they
were far more likely to perceive the others similarly, see figure 6.
Figure 6: Croc display perceptions as a function of how P2 was perceived. A: In the left panel,
we depict the absolute probability of perceiving a croc display in a certain way, as a function of
how P2 was perceived (represented by stylized colored crocs on the x-axis). Rows, from top to
bottom: P0, P1, M1, M2. B: As in A, but the center panel depicts differences from the baserates
depicted in figure 5. C: As in B, but the right panel shows the same information as ratios.
We used P2 perception as a base, because this was the most ambiguous of our stimulus
displays. Note that the conditional probability of seeing a particular color is boosted in all 16
cases, sometimes considerably. The average boost was 21 percentage *points*, relative to the
percentages shown in figure 5. Sometimes, the result of this boost is dramatic, i.e. the
conditional probability of seeing M2 as beige if P2 was seen as beige is 90% (up from 70%).
The conditional probability of seeing P1 as pink if P2 was seen as pink is 86% (up from 18%).
On average, the ratio is boosted almost 2.5 fold. To summarize, there is considerable
consistency within observers and between percepts. This is unlikely to happen accidentally,
such as when observers are unsure of which option to pick, particularly as the response choices
are randomized and distributed over various pages - importantly - as we noted in the Method
section, participants cannot go back and revisit previous stimuli to compare.

Note that the pink percept seems to “gain the most” from having seen the reference stimulus as
pink, but this could be simply because this percept had the most to gain to begin with (lowest
absolute prevalence in the sample). Also note that this pattern of results - 16 positive
differences or 16 ratios larger than 1 - is in itself significant, as we would expect such a result
less than 1 in 65,000 cases by chance alone, translating to a p-value of 1.53e-5. As these are
conditional probabilities, and the individuals involved different people, these are independent
events, so this calculation is valid. Another way to interpret figure 6 is that the P2 percept
represents a strong attractor - in whichever way observers perceive P2 - and for whatever
reason - increases the probability that other croc displays are seen in the same way. In other
words, a propensity to see these stimuli in a certain way cuts across all of these stimuli and
might constitute a characteristic of the individual observer.
What determines the perception of the crocs displays?
After establishing that participants can agree on the color of the stimuli under typical lighting
conditions, but that this consensus breaks down under conditions of SURFPAD lighting, and
that this seems to be an individual characteristic, questions about this individual characteristic
come into focus. What - specifically - predisposes an individual observer to perceive these croc
displays in a certain way - and different from other observers? In the case of “the dress”, a key
factor turned out to be illumination priors - assumptions about the illumination derived from
experience (Witzel et al., 2017; Wallisch, 2017a). We asked observers several questions with
this in mind and explore the relative role of fabric and illumination priors here.
Given the previous result - perception of all 5 crocs is tightly inter-related and given that we
asked the question about sock color right after asking about P2 - which might have drawn
attention to the color of the socks for subsequent stimuli, we focus our discussion of these
matters on the percepts involving P2, see figure 7.

Figure 7: Disagreement about the perception of P2 as a function of responses to relevant
survey questions. Each row represents the responses to a self-report question as the
independent variable. The dependent variable in all cases is the reported perception of P2.
Black vertical bars denote the 99% confidence interval.
The primary driver of P2 perception seems to be how the socks were seen. If the socks were
seen as white, the probability that the crocs were perceived as pink is over 12x that of
participants who did not perceive the socks as white. Dovetailing nicely with this effect is the
observation that the opposite is true for green and grey perception, if the socks were not seen
as white, but in the other direction. Observers were more than 2x as likely to perceive the crocs
as grey and over 4x as likely to perceive them as green if they did not perceive the socks as
white, compared to when they did not. The fact that the confidence intervals are much larger for
the “unsure” group is largely due to this being by far the smallest group. A secondary driver of
disagreement seems to be lighting conditions. Consistent with our interpretation, if participants
did perceive the lighting as abnormal, they were slightly - but significantly - more likely to
perceive P2 as pink and less likely to perceive P2 as grey or green, perhaps because they
questioned the prima facie appearance of the crocs.
Finally, answers to arbitrary questions do not systematically distinguish participants responding
in systematic ways. For instance, we did not expect self-reported croc experience to have any
perceptual effects, mostly due to the fact that even being familiar with crocs does not allow to
disambiguate which of the 28 classic crocs is on display. In sum, unspecific experience without
a clear mechanism does nothing to reduce the uncertainty of the croc color, whereas
assumptions about lighting or socks do.
What determines whether socks were seen as white?
Along the lines of Wallisch (2017a), we suspect that experience is a primary predictor of
assumptions about lighting or sock color, and lifestyle choices a primary predictor of experience.
We did not ask a question about lifestyle choices regarding light, as - unlike in “the dress” case,
where chronotypes have an obvious impact on sunlight exposure, everything else being equal,
there is no clear and reasonably common predictor of exposure to colored LED lights. However,
one can ask about direct experience with wearing plain tube socks, so we did ask that question.
Figure 8 summarizes the results with regard to the perceptual sequelae of answering this
question.

Figure 8: Perceptual consequences of wearing white tube socks. Left panel: Proportion seeing
socks as white as a function of tube socks being worn. Right panel: Same, but seeing P2 as
pink as the dependent variable. Black vertical bars represent the 99% confidence interval.
The effects in figure 8 are obviously less strong than what would be needed to comfortably
claim that we understand this aspect of the phenomenon well, but then again it is both
remarkable and interesting to see these systematic and significant effects in the first place,
given how much random variability is likely involved in these self-reported responses. In
addition, there are certainly other ways to be exposed to tube socks than wearing them, i.e. one
might see them in sports, on TV, in ads, or in stores, but none of these are easy to probe in self-
reports. Put differently, we would not expect strong effects on the basis of this simple self-report
question, as wearing the socks is clearly not a very close proxy to actual overall exposure.
Interestingly, the perceptual effects on P2 trade off clearly, i.e. all of the increased pink
perception comes from a decrease in perceiving P2 as green, with the proportions of grey and
beige entirely unaffected (not pictured in figure 8), just as one would expect if the socks are
reflecting the light, allowing observers to discount it.
How broad is the effect of these priors on perception more generally?
Now that we established the power of these priors on the perception of P2 with respect to color,
one can wonder how far these priors extend. Does P2 perception predict perception of “the
sneaker” and “the dress”? An answer in the affirmative could hint at common underlying
mechanisms and observer characteristica. Does a strong sock prior also predict perception of
“the sneaker” or “the dress”? For answers to these questions, see figure 9.

Figure 9: Priors and non-croc perception. A: Descriptive statistics for sneaker perception. Y-
axis: Proportion of the sample reporting seeing Pink/White (P/W), Grey/Green (G/G),
Pink/Green (P/G), Grey/White (G/W), Switching (S) or Other. B: Descriptive statistics for dress
perception. Y-Axis: Proportion of the sample reporting seeing White/Gold (W/G), Blue/Black
(B/B), Blue/Gold (B/G) and Black/White (B/W) or Switching (S). C: Probability of seeing the
sneaker as pink (pink bars) or grey (grey bars) as a function of seeing P2 as pink, green, grey or
beige (as indicated by the croc pictograms on the x-axis). D: Probability of seeing the dress as
White/Gold (Yellow bars) or Black/Blue (Blue bars) as a function of seeing P2 as pink, green,
grey or beige (as indicated by the croc pictograms on the x-axis). E: Probability of seeing the
sneaker as pink (pink bars) or grey (grey bars) as a function of sock perception. F: Probability of
seeing the dress as White/Gold (Yellow bars) or Black/Blue (Blue bars) as a function of sock
perception. Black vertical bars represent the 99% confidence interval.
The first thing to note is that “the sneaker” is most perceptually similar to P1 - grey perception
strongly dominates, with the remainder mostly taken up by pink, as could be anticipated from
the point clouds in figure 3. Second, we broadly replicate the proportions described in the
perception of “the dress” as reported in Lafer Sousa et al. (2015), or Wallisch (2017a), even four
years later. Third, P2 perception strongly determines sneaker perception. Almost 90% of those
who saw P2 as pink saw the sneaker as pink, whereas just over 10% of those who saw P2 as
green did. Almost 90% of those who saw P2 as green saw the sneaker as grey, whereas just
over 10% of those who saw P2 as pink did. The numbers for seeing P2 as grey is almost the
same as seeing P2 as green, whereas seeing P2 as beige falls somewhere in between grey
and pink. In contrast, P2 perception had little to no influence on the perception of “the dress”,
suggesting that crocs and sneaker perception share a common mechanism - perhaps fabric
priors, specifically white sock priors - whereas the perception of “the dress” is primarily driven by
a different mechanism - perhaps lighting priors, specifically sunlight priors. Fourth, the effects of
assuming the socks are white reach across percepts, predicting not only P2 - and indeed the
perception of the other croc displays - but also the sneaker. This could be due to a number of
reasons. For instance, this could be a manifestation of a tendency to apply fabric priors across
the board. Or this pattern of results could simply reflect the fact that people who think socks are
white are the same people who think that shoelaces - and maybe other objects like that - are
white. It could also reflect the workings of another, not yet well understood mechanism or
personal characteristic, of which sock perception is simply a marker of. Notably, these socks are
often worn in athletic contexts - it would not be too far of a stretch to suggest that they might be
worn with - sneakers - that will tend to happen to have white shoelaces. Finally, consistent with
what we already observed here, how the sock is perceived seems to have very little impact on
how the dress is perceived, again suggesting that the perception of “the dress” is not governed
by fabric priors, but primarily by lighting priors. In addition, the lack of an impact of sock
perception on dress perception suggests that there isn’t simply some overall bias in color
perception that affects all of perception affecting the color perception of socks, crocs, sneakers
and dresses alike. The complete lack of an impact on dress perception rules out this possibility -
priors and how they manifest in perception seem to be fairly specific.

What is the cognitive penetrability of these effects?
Most perceptual effects are notoriously immune to awareness and cognitive interventions
(Firestone & Scholl, 2016). Telling people that the physical “dress” is blue and black does not
change one’s percept, should it be white and gold. However, there is reason to believe that the
effect driven croc perception might be different. For instance, if one were to take one’s percept
at face value, a pink croc under typical lighting conditions, as in Wp would likely be perceived as
grey or green under altered lighting conditions, as these are the colors of the pixels that make
up the croc (see figure 3). However, if one were to recognize that these lighting conditions are
irregular, one might scrutinize them and give more weight to one’s priors, importantly one’s
white sock prior, if such a prior exists, allowing an observer to color calibrate the image and
perhaps recover the color under typical light. To account for this possibility, we asked one
question along these lines, namely whether observers consider it possible that their beliefs
could be wrong. This question is somewhat difficult to interpret, as a lot of factors might go into
answering it, including self-awareness, meta-cognition, modesty, as well as unequal efforts
among people to make sure that their beliefs are accurate. In hindsight, we should have asked
something more closely related to perception, i.e. beliefs in naive realism, but we do think it is
fair to assume that this question can be taken as a proxy for confidence in the veracity of one’s
own cognition, given everything else is equal. Of course, everything else won’t be equal on the
level of the individual, which is why a large sample size is critical. See figure 10 for a summary
of these effects.
Figure 10: Cognitive penetrability of percepts. In each panel, we report the proportion of
participants who report a color perception that corresponds to the color appearance under
typical illumination conditions (CAUTION, see Method) as a function of whether participants
allowed for the possibility that some of their beliefs might be wrong, for a given stimulus display.
Black vertical bars represent the 99% confidence interval.

The effects illustrated in this figure are consistent and perhaps even ironic. The more confident
observers are in their beliefs being correct, the less likely they are to recover CAUTION, which
is not implied by the pixel clouds in figure 3. In other words, these observers trust the
appearance of the shoes, without second-guessing this appearance, perhaps due to confidence
in the general veracity of one’s cognition. This extends to CAUTION perceptions of socks and
sneakers, which are significantly modulated by the answer to this question. As we noted above -
the perception of “the dress” is not influenced by meta-cognition. We have confidence that the
confidence intervals in P0 would shrink considerably with more participants - we had by far the
least respondents to this question. Whether this is ironic or not - echoing other meta-cognitive
effects (Dunning, 2011) depends on one’s conception of what color the crocs “really” are -
CAUTION or the color of the individual pixels as seen in the point clouds of figure 3, or if that
question even makes sense, as the concept of veridicality is not well defined for color in the first
place.
What is the effect of demographics?
Finally, we wondered about the effects of demographics on the perception of these stimuli. Prior
research on “the dress” (e.g. Lafer-Sousa et al., 2015; Wallisch, 2017a) suggest that gender
effects are miniscule, but age effects could be considerable, although a relative dearth of old
observers is a concern. See figure 11 for an overview of these demographic effects.
Figure 11: Demographic effects. A: Stimuli where CAUTION decreases with age. Circles: Dress.
Right-pointed triangles: Sneaker. Left-pointed-triangles: P1. Down-pointed triangles: P2. Red
dotted lines represent SEM. B: Stimuli where there is no clear age-effect. Squares: M1.
Diamonds: M2. Up-pointed triangles: P0. Red dotted lines represent SEM. C: Effects of gender

on CAUTION. Each pair of bars corresponds to the stimulus displays labeled on top of it. Black
vertical bars represent the 99% confidence interval.
To summarize, most pink croc stimuli, as well as the dress and sneaker exhibit strong age-
effects, with a gradual decrease in CAUTION as a function of age. This contrasts with a sharp
decline of the White/Gold modal percept with retirement age for the dress (Wallisch, 2017a),
which we replicated here and a sharp rise in Blue/Gold percepts (not pictured). As noted in
Wallisch (2017a), it is hard to meaningfully interpret these effects, as age is a carrier variable of
so many other effects, including physiological effects and cohort effects. We also observe a lack
of an age related decline in CAUTION for the mint croc stimuli, which could possibly be due to a
floor effect - there is little CAUTION for these stimuli to begin with. We attribute the lack of an
age effect for P0 to a lack of participants, as in figure 10 - the error bars are simply too wide to
interpret this trajectory meaningfully. Finally, as already discussed, gender effects are miniscule,
if they exist at all. There seems to be marginal gender effect on the perception of P1 and its
related stimulus, the sneaker, but if they exist, these effects are small and hard to interpret. This
addresses one of the concerns related to this study, namely that the majority of our sample
identified as female. This might be owed to the fact that we framed the recruitment for our study
in terms of donating data, and women have been shown to be more proactive in charitable
giving in general (Piper & Schnepf, 2008; Mesch et al., 2011). As figure 11 illustrates, whereas
this might have biased the gender-composition of the sample, it is unlikely that this imbalance
affected the perceptual effects reported in this study.

Discussion
We managed to create visual displays such that observers categorically disagree about the
colors of the objects depicted, even though the same objects evoke near complete consensus
under typical lighting conditions. We did so by systematically avoiding, minimizing and reducing
cues that observers typically use to determine the color appearance of an object, such as
patterned backgrounds (Jenness & Shevell, 1995), color memory (Witzel et al., 2011) or familiar
lighting (Maloney, 1999), while at the same time introducing a part of the object that mirrors the
properties of the lighting, but for which some - but not all - participants have a prior that they use
to anchor their color percept, a process which we term SURFPAD. We were also able to show
that the perception of the different color displays was systematically linked within a given
observer, extending even to other color-ambiguous stimuli, such as the “sneaker”. We take this
to mean that a disposition to see these stimuli in a certain way amounts to a personality
characteristic. This personality characteristic is - in this case - best characterized as a “fabric
prior” for white socks, which strongly determines the perception of the displays we created. This
fabric prior - in turn - seems to have been determined by experience with this type of sock, akin
to the illumination priors that determine the perception of the “dress” due to differential exposure
to sunlight (Wallisch, 2017a).
We believe that these effects are consistent with the operation of known cognitive mechanism
involved in color vision, most importantly color constancy (McCann et al., 1976). In this special
case, color constancy seems to be derived from a kind of color memory (Jin & Shevell, 1996;
Hansen et al., 2006), in a subset of our observers. Remarkably, this mechanism allows these
observers to recover the color that corresponds to CAUTION, even though there are virtually no
individual pixels in the object that suggest CAUTION, a striking empirical instance of the whole
being different from the sum of its parts, as suggested theoretically by Koffka (1922). Put
differently, instead of alerting the organism to the profound uncertainty inherent in these
displays, the brain seems to err on the side of arriving at actionable conclusions by operating
inadvertent mechanisms akin to autocorrect - educated guesses to recover the stimulus
situation by making assumptions about the color of lighting (in the case of the dress) or fabric
(here). It is increasingly well recognized that perceptual mechanisms in the brain utilize
information beyond the present stimulus in the form of priors, assumptions, expectations and
predictions, in order to cope with and act in an environment that is inherently more uncertain
than one would prefer (Press & Yon, 2019). This line of reasoning also suggests that while we
understand the color appearance of objects under typical lighting conditions relatively well - to
the point of being able to reasonably describe color appearance in terms of a linear system
(Shevell, 2003), the same might be true under carefully altered lighting conditions, where
differential priors play a much larger role in determining the subjective color experience.
There are several limitations inherent in this study that potentially threaten the generality of
these conclusions. Most of these pertain to our specific implementation. First, in order to rapidly
obtain data from many observers, we logged the data online, instead of in a lab. This will
necessarily introduce considerable uncertainty as to the stimulus situation that the participants
actually encountered, such as screen size or color calibration, which would not be a concern
under controlled laboratory conditions. Similarly, we would have wanted to take more complete
control of the lighting conditions when creating these stimuli, involving both a high-end LED
lighting system and a photometer. Then again, that we readily found these effects despite the

noise introduced by presenting the stimuli online speaks to both their strength and robustness,
akin to all the other color-ambiguous stimuli (the “dress”, “adidas jacket”, “flip-flops”, etc.) that
surfaced and thrived “in the wild”, without relying on carefully controlled laboratory conditions. In
addition, our observers were able to agree on the color of these objects under typical lighting
conditions, so it is unlikely that this ambiguity is introduced as an artifact of different screen
settings. Similarly, whereas we strongly suggest future studies in this domain to more carefully
control the lighting conditions when creating the ambiguous displays, our efforts still have merit
as a proof of concept. What seems to matter critically is the relative position of the color clouds
from the two objects - the colored target object (CTO), the croc in this case and the cueing or
anchoring illumination-mirroring object (IMO), in this case the sock - in hsv space. Displays
seem to evoke particularly strong disagreement between observers if there is a separation
between the two clouds and if they form parallel columns in this space, whether this
configuration is achieved by a careful control of lighting conditions or digital filtering. The
creation of such stimuli in hsv space will also allow us to address the issue of beige perception.
Are participants perceiving these stimuli as beige because they tried to color-correct their
percept, but failed (ending up at beige instead of pink), or are the actual pixels - as suggested
by the point clouds in figure 3, particularly for M1 and M2 far enough in the beige range to make
a beige croc color plausible a priori? Cleaner stimuli - that are clearly separated and mostly
avoid the beige range should cut down on beige perception, increasing the perception of pink
(for the Ps) and grey (for M).
Another concern is that we illustrated the SURFPAD principle only with one type of object -
crocs and socks, while claiming that it is a general design principle. This is a clear case of a
concern that can readily be addressed in future studies, using any number of colored stimuli,
such as colored candles as the CTO and the wick as the IMO or colored tissue boxes as the
CTO and white tissues as the IMO. With some imagination, this principle can likely be extended
to other visual domains, such as distance or motion, and perhaps even to auditory stimuli.
Based on our experiences of applying SURFPAD to crocs and socks, we predict that a suitable
application of these principles would yield stimulus displays with similar levels of disagreement
between observers.
However, the biggest concern regarding this kind of work is that we overlooked some low level
explanation for the phenomena we observed, like so many others in perception (Firestone &
Scholl, 2016). It would not be the first time that a high profile and high level explanation is better
explained by other factors, which is usually the case when the experimenter inadvertently fails
to notice or rule out covariates that correlate and drive the dependent variable even better. A
prominent example of this would be Mischel who thought that willpower determines the ability to
resist the temptation of marshmallows (Mischel et al., 1989), a phenomenon better explained by
trust (Watts et al., 2018). This can happen even in experimental work, when the independent
variables created by the experimenter mask a confound. For instance, it has been proposed that
observers perceiving a rotating line to lead a simultaneously flashed line suggests the operation
of motion extrapolation mechanisms to guide dynamic action (Nijhawan, 1994). However, this
phenomenon can be more readily explained in terms of a low-level feature - the flashed line has
lower effective contrast than the rotating one, and lower contrast is associated with longer
neural delays (Raiguel et al., 1999). This suggests that the rotating line is seen as leading
because its signal arrives in the brain before that of the flashed line, so it should be possible to

negate this advantage - and the lead - by increasing the relative contrast of the flashed line,
which is what can be shown empirically (Purushothaman et al., 1998).
While we concede this theoretical possibility, we do not think it likely in this case. First, the
SURFPAD principle leverages a combination of known mechanisms and principles of visual
processing, without relying on exotic postulates or remote theoretical possibilities. Second,
given how rare such color ambiguous displays are - with the advent of smartphones, there are
now well over a trillion images taken per year (Richter, 2017), and only a handful of these are
known to be color-ambiguous, whereas all the displays we created according to this principle
are color ambiguous, creating such displays inadvertently is about as plausible as suggesting to
go to the moon without a solid understanding of gravity or to build a nuclear reactor without a
reasonable understanding of the basic physics governing nuclear reactions involved.
Finally, suggesting that there are more powerful low level features at work that we somehow
overlooked strains credulity, given the sheer size of the effects at play - the white sock prior
alone yields a 12-fold range in the pink percept of P2, which in turn determines close to 90% of
the sneaker percept. One would be hard-pressed to envision more powerful predictors of these
percepts.
Of course, it is true that - in principle - a final determination on this issue won’t be possible
without future work. If we claim that socks are critical to bring about disagreement of the croc
color, it stands to reason that we should create and ask observers about crocs altogether
without socks. As this might look somewhat strange, a fair comparison would be to use the
same exact crocs, but making sure that the socks do not reflect the light. Using perfectly white
socks should not allow observers to use the socks to color calibrate the rest of the display.
Similarly, it might be of interest to ask observers directly about the color of shoelaces to further
elucidate the link between sneaker perception and croc perception, and the role of fabric priors
and their specificity as they apply to perception.
Our results do suggest that disagreements due to differential illumination priors (as in the dress)
and disagreements due to differential fabric priors (as in the crocs) are similar, but separate
phenomena. We believe that this is due to the fact that the “dress” phenomenon is situated on
the daylight axis between yellow and blue, whereas our crocs displays leverage the opposite
axis between pink and green, which is likely why there is so little cross-display information. In
other words, we can’t predict how someone saw the dress based on how they saw the crocs or
vice versa, because these displays operate on different axes through color space. We propose
to create a display involving both fabric priors and daylight priors involving grey, yellow and blue
crocs as well as yellow and blue lights, using SURFPAD principles.
A bigger point of this study - beyond crocs, socks and color - is that this principle pertains to the
nature of disagreement and polarization. Given some of the feedback we received, there is no
question that some people are annoyed by these kinds of ambiguous displays, as being
confronted with these displays threatens their self-perception as being accurate observers of
and effective operators in the external world. Of course, this kind of self-perception is validated
all the time - under typical conditions, people’s interpretation of the situation does correspond to
reality, and the interpretation of others. What makes the operation of these cognitive
“autocomplete” or “autocorrect” mechanisms so pernicious is that they do not flag to the
observer when they had to be leveraged to bring about the perceptual interpretation. They are

simply presented to the observer as “the percept”, regardless of how it came about, perhaps in
an attempt to facilitate and encourage action.
Note that we would expect these effects to be least worrisome in perception, for several
reasons. First, the brain devotes considerable resources to the processing of perceptual
information (Wallisch & Movshon, 2008). Second, there are - under natural conditions - multiple
redundancies both within and across modalities (Landy & Kojima, 2001; Hillis et al., 2002;
Körding et al., 2007). Third, there was likely immense evolutionary pressures to get the right
answer. If the brain is indeed epistemically immodest - biased towards deriving actionable
conclusions despite profound uncertainty - as suggested by our results, it seems important that
these conclusions are not catastrophic to the integrity of the organism. So whereas these
mechanisms are particularly tractable in the perceptual domain, we suspect that they play a
larger role in cognition beyond perception, particularly as we now live in environments and
systems that are complex enough that there are almost always considerable degrees of
freedom for observers to come up with their own interpretation (Wallisch & Whritner, 2017). Put
differently, we are now - cognitively - living in a SURFPAD world. We are increasingly
encountering conditions of tremendous and polarized differences in the interpretations of current
events. This is somewhat surprising, as it is happening in spite of all the information that is now
available about the state of the world, which could drive consensus. But it is important to note
that given the complexity of the world, we are almost always in a regimen of high SU (due to the
complexity of reality and the social systems we inhabit itself). So if there is fundamental
disagreement (D), it is likely not due to disagreements about the evidence - shrouded in SU -
but rather the cognitively invisible RFPA. Where do these ramified or forked priors and
assumptions come from?
It is no secret that media and social media have effectively created increasingly divergent
differential priors, leading to starkly differing interpretations of events, regardless of what
actually happens. Alarmingly, both media and social media seem to have created cognitive
versions of the socks, anchors that appear politically correct to a particular echo chamber
(Brady et al., 2019). To explore reductions in the difference in interpretations, future work on
SURFPAD could explore the malleability of these differential priors - does priming with suitable
stimuli just before the study update the prior? Is it possible to override the prior by introducing
other display elements that could restore consensus, e.g. a colored background that also
reflects the illumination?
Thus, in terms of fostering conflict resolution, we advise neither to highlight the disagreement
itself - which typically garners all the attention, nor the SU, which is inevitable in a complex
world, but rather to focus squarely on the roots of disagreement - the RFPA. Differential priors,
created in a strong and polarized fashion and under conditions of profound uncertainty will make
dramatically differing interpretations of the same events and situations inevitable.
It is appropriate to address this issue with the urgency that it necessitates if one is to avoid the
adverse outcomes associated with mounting and polarized disagreement.

References
Brady, W. J., Crockett, M., & Van Bavel, J. J. (2019). The MAD Model of Moral Contagion: The
role of motivation, attention and design in the spread of moralized content online.
Dunning, D. (2011). The Dunning–Kruger effect: On being ignorant of one's own
ignorance. In Advances in experimental social psychology. Academic Press.
Efron, B., & Tibshirani, R. J. (1994). An introduction to the bootstrap. CRC press.
Firestone, C., & Scholl, B. J. (2016). Cognition does not affect perception: Evaluating the
evidence for “top-down” effects. Behavioral and Brain Sciences, 39.
Griffiths, T. L., & Tenenbaum, J. B. (2006). Optimal predictions in everyday cognition.
Psychological Science, 17(9), 767-773.
Hansen, T., Olkkonen, M., Walter, S., & Gegenfurtner, K. R. (2006). Memory modulates color
appearance. Nature Neuroscience, 9(11), 1367.
Hillis, J. M., Ernst, M. O., Banks, M. S., & Landy, M. S. (2002). Combining sensory information:
mandatory fusion within, but not between, senses. Science, 298(5598), 1627-1630.
Hupé, J. M., & Rubin, N. (2003). The dynamics of bi-stable alternation in ambiguous motion
displays: a fresh look at plaids. Vision Research, 43(5), 531-548.
Jackson, F. (1982). Epiphenomenal Qualia. The Philosophical Quarterly, 32(127)
Koffka, K. (1922). Perception: an introduction to the Gestalt-Theorie. Psychological Bulletin,
19(10), 531.
Jenness, J. W., & Shevell, S. K. (1995). Color appearance with sparse chromatic context. Vision
Research, 35(6), 797-805.
Jin, E. W., & Shevell, S. K. (1996). Color memory and color constancy. JOSA A, 13(10), 1981-
1991.
Körding, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature,
427(6971), 244.
Körding, K. P., Beierholm, U., Ma, W. J., Quartz, S., Tenenbaum, J. B., & Shams, L. (2007).
Causal inference in multisensory perception. PLoS one, 2(9), e943.
Lafer-Sousa, R., Hermann, K. L., & Conway, B. R. (2015). Striking individual differences in color
perception uncovered by “the dress” photograph. Current Biology, 25 (13), R545–R546.
Land, E. H., & McCann, J. J. (1971). Lightness and retinex theory. Josa, 61(1), 1-11.

Landy, M. S., & Kojima, H. (2001). Ideal cue combination for localizing texture-defined edges.
JOSA A, 18(9), 2307-2320.
Levelt, W. J. (1965). On binocular rivalry (Doctoral dissertation, Van Gorcum Assen).
Lotto, R. B., & Purves, D. (1999). The effects of color on brightness. Nature Neuroscience,
2(11), 1010.
Maloney, L. T. (1999). Physics-based approaches to modeling surface color perception. Color
Vision: From genes to perception, 387-416.
McCann, J. J., McKee, S. P., & Taylor, T. H. (1976). Quantitative studies in retinex theory a
comparison between theoretical predictions and observer responses to the “color mondrian”
experiments. Vision Research, 16(5), 445-IN3.
Mesch, D. J., Brown, M. S., Moore, Z. I., & Hayat, A. D. (2011). Gender differences in charitable
giving. International Journal of Nonprofit and Voluntary Sector Marketing, 16(4), 342-355.
Mischel, W., Shoda, Y., & Rodriguez, M. I. (1989). Delay of gratification in children. Science,
244(4907), 933-938.
Nijhawan, R. (1994). Motion extrapolation in catching. Nature.
Piper, G., & Schnepf, S. V. (2008). Gender differences in charitable giving in Great Britain.
Voluntas: International Journal of Voluntary and Nonprofit Organizations, 19(2), 103-124.
Purkinje, J. E. (1825). Beobachtungen und Versuche zur Physiologie der Sinne. Neue Beitrage
zur Kenntniss des Sehens in subjektiver Hinsicht. Berlin, Germany: Reimer.
Press, C., & Yon, D. (2019). Perceptual Prediction: Rapidly Making Sense of a Noisy World.
Current Biology, 29(15), R751-R753.
Purushothaman, G., Patel, S. S., Bedell, H. E., & Ogmen, H. (1998). Moving ahead through
differential visual latency. Nature, 396(6710), 424.
Monnier, P., & Shevell, S. K. (2003). Large shifts in color appearance from patterned chromatic
backgrounds. Nature Neuroscience, 6(8), 801.
Raiguel, S. E., Xiao, D. K., Marcar, V. L., & Orban, G. A. (1999). Response latency of macaque
area MT/V5 neurons and its relationship to stimulus parameters. Journal of Neurophysiology,
82(4), 1944-1956.
Richter, F. (2017). https://www.statista.com/chart/10913/number-of-photos-taken-worldwide/

Rubin, E. (1921). Visuell wahrgenommene Figuren: Studien in psychologischer Analyse.
Gyldendalske Boghandel.
Shevell, S. K. (Ed.). (2003). The science of color. Elsevier.
Stabell, B., & Stabell, U. (2002). Effects of rod activity on color perception with light adaptation.
JOSA A, 19(7), 1249-1258.
Stocker, A. A., & Simoncelli, E. P. (2006). Noise characteristics and prior expectations in human
visual speed perception. Nature Neuroscience, 9(4), 578.
Wald, G. (1964). The receptors of human color vision. Science, 145(3636), 1007-1016.
Wallisch, P., & Movshon, J. A. (2008). Structure and function come unglued in the visual cortex.
Neuron, 60(2), 195-197.
Wallisch, P. (2016): https://slate.com/technology/2016/03/the-science-of-the-black-and-blue-
dress-one-year-later.html
Wallisch, P. (2017a). Illumination assumptions account for individual differences in the
perceptual interpretation of a profoundly ambiguous stimulus in the color domain:“The dress”.
Journal of Vision, 17(4), 5-5.
Wallisch, P. (2017b): https://slate.com/technology/2017/04/heres-why-people-saw-the-dress-
differently.html
Wallisch, P. (2017c) https://slate.com/technology/2017/10/the-shoe-is-the-new-dress.html
Wallisch, P., & Whritner, J. A. (2017). Strikingly low agreement in the appraisal of motion
pictures. Projections, 11(1), 102-120.
Watts, T. W., Duncan, G. J., & Quan, H. (2018). Revisiting the marshmallow test: A conceptual
replication investigating links between early delay of gratification and later outcomes.
Psychological Science, 29(7), 1159-1177.
Witzel, C., Valkova, H., Hansen, T., & Gegenfurtner, K. R. (2011). Object knowledge modulates
colour appearance. i-Perception, 2(1), 13-49.
Witzel, C., Racey, C., & O'Regan, J. K. (2017). The most reasonable explanation of “the dress”:
Implicit assumptions about illumination. Journal of Vision, 17(2), 1-1.