This is the accepted version of the journal article: Jonauskaite, Domicele; Dael, Nele; Parraga, Carlos Alejandro; [et al.]. «Strip- ping The Dress : the importance of contextual information on inter-individual differences in colour perception». Psychological Research, Vol. 84, Issue 4 (June 2020), p. 851-865. DOI 10.1007/s00426-018-1097-1 This version is available at https://ddd.uab.cat/record/275086 under the terms of the license 1 1 2 3 Stripping #The Dress: The importance of contextual information on inter-individual 4 5 differences in colour perception 6 7 8 9 Domicele Jonauskaite1, Nele Dael1, C. Alejandro Parraga2, Laetitia Chèvre1, Alejandro García Sánchez2 10 11 and Christine Mohr1 12 13 14 15 16 17 18 1 19 Institute of Psychology, University of Lausanne, Lausanne, Switzerland 20 21 22 2 Computer Vision Center / Comp. Science Dept., Universitat Autònoma de Barcelona, Spain 23 24 25 26 27 28 29 Corresponding author: 30 31 32 33 Domicele Jonauskaite 34 35 36 37 Institute of Psychology 38 39 40 University of Lausanne 41 42 43 44 Quartier Mouline, Bâtiment Géopolis 45 46 47 CH-1015, Lausanne, Switzerland 48 49 50 51 +44 21 692 32 52 52 53 54 55 Domicele.Jonauskaite@unil.ch 56 57 58 ORCID: 0000-0002-7513-9766 59 60 61 62 63 64 -1- 65 #The Dress: decontextualized 2 Acknowledgements: The research was supported by the Institute of Psychology, University of 1 2 Lausanne and the Swiss National Science Foundation (Doc.CH fellowship grant to DJ, number 3 4 5 P0LAP1_175055). We wish to thank Loïc Gigandet and Simona Scopazzini for collecting part of the 6 7 data. We also wish to thank AkzoNobel, Imperial Chemical Industries (ICI) Limited, in particular Dr 8 9 David Elliott, Dr Tom Curwen and Peter Spiers, Color R&I team, Slough, UK, and Stephanie Kraneveld, 10 11 12 Sassenheim, the Netherlands, for having supported our empirical work on colours and emotions. No 13 14 conflicts of interest are declared. 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 -2- 65 #The Dress: decontextualized 3 Abstract 1 2 3 4 In 2015, a picture of a Dress (henceforce the Dress) triggered popular and scientific interest; some 5 6 reported seeing the Dress in white and gold (W&G) and others in blue and black (B&B). We aimed to 7 8 describe the phenomenon and investigate the role of contextualization. Few days after the Dress 9 10 11 had appeared on the Internet, we projected it to 240 students on two large screens in the 12 13 classroom. Participants reported seeing the Dress in B&B (48%), W&G (38%), or blue and brown 14 15 (B&Br; 7%). Amongst numerous socio-demographic variables, we only observed that W&G viewers 16 17 18 were most likely to have always seen the Dress as W&G. In the laboratory, we tested how much 19 20 contextual information is necessary for the phenomenon to occur. Fifty-seven participants selected 21 22 colours most precisely matching predominant colours of parts or the full Dress. We presented, in 23 24 25 this order, small squares (a), vertical strips (b), and the full Dress (c). We found that i) B&B, B&Br, 26 27 and W&G viewers had selected colours differing in lightness and chroma levels for contextualized 28 29 30 images only (b, c conditions) and hue for fully contextualized condition only (c), and ii) B&B viewers 31 32 selected colours most closely matching displayed colours of the Dress. Thus, the Dress phenomenon 33 34 emerges due to interindividual differences in subjectively perceived lightness, chroma, and hue, at 35 36 37 least when all aspects of the picture need to be integrated. Our results support previous conclusions 38 39 that contextual information is key to colour perception; it should be important to understand how 40 41 this actually happens. 42 43 44 45 46 47 Abstract word count: 248 48 49 50 51 Total word count: 8,192 52 53 54 55 Keywords: #TheDress, Internet dress, context, visual colour illusion, colour perception 56 57 58 59 60 61 62 63 64 -3- 65 #The Dress: decontextualized 4 1 2 3 4 Introduction 5 6 7 The online picture of a dress (henceforth, the “Dress”) evoked a phenomenal interest in early 2015. 8 9 10 Online, people could not agree whether the colour combination of the Dress was white and gold 11 12 (W&G) or blue and black (B&B) (Holderness, 2015). The actual dress from which the Dress picture 13 14 was taken was manufactured as royal blue with black horizontal stripes (Wallisch, 2017). The 15 16 17 ambiguous Dress picture, on the other hand, displayed very different colours to royal blue and black 18 19 (for CIE Lab coordinates of the Dress see Melgosa, Gomez-Robledo, Isabel Suero, & Fairchild, 2015). 20 21 The B&B perceivers could not believe that the W&G perceivers saw the same picture and vice versa. 22 23 24 Yet, these two perceptions emerged even when people were in the same place at the same time 25 26 looking at the same monitor (Gegenfurtner, Bloj, & Toscani, 2015). Furthermore, although the colour 27 28 combination perception seemed to generally remain stable across time, for some individuals, the 29 30 31 perception switched between B&B and W&G colour combinations (Chetverikov & Ivanchei, 2016; 32 33 Lafer-Sousa & Conway, 2017; Lafer-Sousa, Hermann, & Conway, 2015; Vemuri, Bisla, Mulpuru, & 34 35 36 Varadarajan, 2016). 37 38 39 The interest in the Dress phenomenon has somehow abated in the public domain, but remains 40 41 42 significant to scientists as the scientific contributions including but not limited to a special issue in 43 44 the Journal of Vision highlight. These authors were facing a new visual (colour) illusion (Brainard & 45 46 Hurlbert, 2015) – an experience when a true source of a stimulus differs from what is perceived 47 48 49 (Corney & Lotto, 2007). The ambiguous Dress picture is an interesting visual illusion to study. Unlike 50 51 some other visual illusions (e.g., simultaneous brightness contrast illusion, Lotto & Purves, 1999), the 52 53 ambiguous Dress picture does not evoke the “misleading” perception of colours in the same way for 54 55 56 all people. The ambiguous Dress picture has been argued to resemble a specific class of visual 57 58 illusions called bi-stable (or multi-stable) visual illusions (Chetverikov & Ivanchei, 2016; Lafer-Sousa 59 60 61 62 63 64 -4- 65 #The Dress: decontextualized 5 & Conway, 2017; Lafer-Sousa et al., 2015), because it can be perceived in different, changing colour 1 2 combinations (B&B, W&G). In the case of some bi-stable illusions, perceivers can train to more or 3 4 5 less spontaneously switch between different perceptions or even see both of them simultaneously 6 7 (e.g. see a duck and a rabbit; McManus, Freegard, Moore, & Rawles, 2010). This is not the case for 8 9 the ambiguous Dress picture, where the change in perception arises spontaneously and only to 10 11 12 some viewers (e.g., Lafer-Sousa et al., 2015). In the case of other visual illusions, perceivers may see 13 14 no solution at all or see several contradictory solutions at the same time (e.g., Mooney faces, 15 16 17 Carbon, Grüter, & Grüter, 2013). Many of these visual illusions arise due to prior perceptual 18 19 experience, knowledge or expectations, which may lead perceivers to seeing one, several, or no 20 21 possible interpretations. 22 23 24 25 In the initial Internet survey of about three million respondents, 68% of the respondents indicated 26 27 seeing the ambiguous Dress picture as W&G and 32% of the respondents as B&B in a forced-choice 28 29 30 situation (Holderness, 2015). Subsequent studies reported a higher proportion of W&G viewers 31 32 (Dixon & Shapiro, 2017; Mahroo et al., 2017; Moccia et al., 2016; Wallisch, 2017; Witzel, Racey, & 33 34 O’Regan, 2017), a higher proportion of B&B viewers (Chetverikov & Ivanchei, 2016; Lafer-Sousa et 35 36 37 al., 2015), or an equal split between the two viewer types (Aston & Hurlbert, 2017; Chetverikov & 38 39 Ivanchei, 2016; Hesslinger & Carbon, 2016; Karlsson & Allwood, 2016; Schlaffke et al., 2015; Vemuri 40 41 et al., 2016; Winkler, Spillmann, Werner, & Webster, 2015) using forced-choice and free naming 42 43 44 paradigms. In studies that went beyond the possibility of two viewer types, mainly using free- 45 46 naming paradigms to assess the colours of the ambiguous Dress picture, an intermediate variant of 47 48 49 the Dress perception emerged – participants reported seeing the Dress as blue and brown/gold 50 51 (B&Br) (Aston & Hurlbert, 2017; Lafer-Sousa et al., 2015; Mahroo et al., 2017; Wallisch, 2017; Witzel 52 53 et al., 2017). Subsequently, it has been suggested that the Dress perception follows a continuum 54 55 56 ranging from white to blue and gold to black rather than resulting in discrete categories of either a 57 58 B&B or W&G perception (e.g. Gegenfurtner et al., 2015). 59 60 61 62 63 64 -5- 65 #The Dress: decontextualized 6 The possibility that all people perceive the same colours on the ambiguous Dress picture, but name 1 2 them differently (e.g., very light blue may be named “white” or “blue”) has largely been ruled out by 3 4 5 colour matching studies (Aston & Hurlbert, 2017; Chetverikov & Ivanchei, 2016; Gegenfurtner et al., 6 7 2015; Lafer-Sousa et al., 2015; Toscani, Gegenfurtner, & Doerschner, 2017; Witzel et al., 2017). 8 9 Common to these studies, participants matched the ambiguous Dress picture colours to colours on 10 11 12 computerized or physical colour-matching tools. A consistent result across studies indicated that 13 14 B&B viewers overall match darker colours than W&G viewers. These perceptual differences support 15 16 17 the notion that the differences in reported Dress colours go beyond naming differences and reflect 18 19 genuine perceptual differences. Colour can be perceptually defined as having three components 20 21 (colour appearance parameters) – hue (what a laymen refer to as colour: “red”, “yellow”, “green”), 22 23 24 chroma (how vivid the colour is), and lightness (how light the colour is; Hunt & Pointer, 2011). These 25 26 colour parameters are intuitively comprehensible and interpretable. Previous studies did not 27 28 interpret colour matches to the ambiguous Dress picture from the angle of hue, chroma, and 29 30 31 lightness (Aston & Hurlbert, 2017; Chetverikov & Ivanchei, 2016; Lafer-Sousa et al., 2015; Witzel et 32 33 al., 2017), but they could bring certain insight into how viewers’ perception differs when confronted 34 35 with the ambiguous Dress image. 36 37 38 39 Contextualization might play a role in the colour perception of the ambiguous Dress picture. By 40 41 showing the same picture information, but in a decontextualized way (e.g. cutting the picture in 42 43 44 pieces, scrambling), information processing about the surroundings (and, by inference, illumination 45 46 source) would be disrupted. As the Dress illusion is likely to emerge from the integration of the 47 48 49 contextual information on the ambiguous Dress picture, decontextualization should reduce 50 51 differences in perceived colours on the ambiguous Dress picture between viewer types. Hesslinger 52 53 and Carbon (2016) provided first evidence for this possibility. These authors cut the ambiguous 54 55 56 Dress picture into squares with varying size and scrambled them. They then asked participants to 57 58 report how blue each scrambled image as well as the original ambiguous Dress picture appeared 59 60 61 62 63 64 -6- 65 #The Dress: decontextualized 7 from looking not blue at all to very blue. They showed that B&B viewers reported the ambiguous 1 2 Dress picture being very blue while W&G viewers reported it being not blue at all. As the scrambled 3 4 5 square sizes decreased, the difference between B&B and W&G viewers decreased in terms of 6 7 reported “blueness” of the stimulus. Hence, low degree of contextual information of the ambiguous 8 9 Dress picture made everyone perceive the colours of the ambiguous Dress picture in a similar way. 10 11 12 13 Taking into account the published literature, we here report on two studies we performed right after 14 15 the appearance of the ambiguous Dress picture online. When we spotted this picture, we were, as 16 17 18 probably most of our colleagues, intrigued by this phenomena. We tested whether individual 19 20 differences in the ambiguous Dress image perception persisted in stimuli progressively lacking the 21 22 contextual richness to understand how much contextual information is necessary for the 23 24 25 phenomenon to occur. If the illusion persisted with little contextual information, differences in the 26 27 ambiguous Dress perception might be driven by lower level factors rather than the different 28 29 30 assumptions about illumination. Firstly, four days after the picture’s appearance, we conducted a 31 32 classroom survey to assess the prevalence of the different colour perceptions as a function of 33 34 demographic variables and subjective illumination interpretations. In a subsequent laboratory study, 35 36 37 we wanted to test if and how participants’ colour perception changes when systematically adding 38 39 contextual information of the Dress. We were interested in such potential changes as a function of 40 41 the different subjective “perception” groups. 42 43 44 45 Participants were invited to partake in a study on colour perception (the ambiguous Dress was not 46 47 mentioned). In this fixed order, participants saw parts of the ambiguous Dress picture and then the 48 49 50 full, contextualized image. We asked participants to use a computerized colour picker and match 51 52 their prevalent colour impression for three Dress-related stimuli that varied in the amount of 53 54 55 contextual information: isolated patches taken from the blue/white regions and the black/gold 56 57 regions of the ambiguous Dress picture, strips along the vertical axis of the ambiguous Dress picture 58 59 and the ambiguous Dress picture shown in full (Figure 1). While the patches condition provided 60 61 62 63 64 -7- 65 #The Dress: decontextualized 8 reduced information about the surrounding context of the ambiguous Dress as well as eliminated 1 2 information about the pattern of the ambiguous Dress picture, the strips condition preserved 3 4 5 information about the ambiguous Dress pattern but reduced surrounding contextual information. 6 7 Thus, we could investigate: 1) colour matches of the Dress colours and how they differed between 8 9 viewer types; 2) the role of degree of decontextualization of the ambiguous Dress picture; and 3) the 10 11 12 differences in colour parameters between displayed and matched colours of the ambiguous Dress 13 14 picture. 15 16 17 18 To be able to report results conveying subjective perception, we broke down participants’ colour 19 20 matches into three perceptual attributes of colour (i.e., hue, chroma, and lightness) and tested 21 22 whether any of these attributes differed between self-reported colour perception when seeing the 23 24 25 full ambiguous Dress picture. Accordingly, we used participants’ reported colour perception to 26 27 subsequently allocate them to Blue and Black (B&B) and White and Gold (W&G) viewers, and also to 28 29 30 the repeatedly described group of Blue and Brown (B&Br) viewers. Indeed, in contrast to notions 31 32 that the ambiguous Dress picture perception is bimodal (Chetverikov & Ivanchei, 2016; Drissi 33 34 Daoudi, Doerig, Parkosadze, Kunchulia, & Herzog, 2017; Vemuri et al., 2016), the emergence of the 35 36 37 B&Br viewer group supports independent reports of a continuous phenomenon (Aston & Hurlbert, 38 39 2017; Gegenfurtner et al., 2015; Witzel et al., 2017). It seems that the perceptions of the ambiguous 40 41 Dress picture ranges from white to blue, and from gold to black. In line with this continuum idea, we 42 43 44 analysed results as a function of a linear variation between viewer types and expected that B&Br 45 46 viewers’ colour matches would fall between the matches of B&B viewers and W&G viewers. 47 48 49 50 Based on previous studies, we hypothesised that lightness of matched colours would differ between 51 52 self-reported viewer types: B&B viewers would select darker colours than W&G viewers (Aston & 53 54 55 Hurlbert, 2017; Chetverikov & Ivanchei, 2016; Gegenfurtner et al., 2015; Lafer-Sousa et al., 2015; 56 57 Toscani et al., 2017; Witzel et al., 2017). We also expected to find some differences in hue and 58 59 chroma between viewer types (Aston & Hurlbert, 2017; Chetverikov & Ivanchei, 2016; Lafer-Sousa et 60 61 62 63 64 -8- 65 #The Dress: decontextualized 9 al., 2015; Witzel et al., 2017). Because these previous studies analysed colour matches using 1 2 different colour models, we could not use their results to formulate more precise predictions. For 3 4 5 contextualization, we could expect that decontextualization disrupts the Dress illusion because 1) 6 7 different viewer types have different interpretation of the background on the ambiguous Dress 8 9 picture (Chetverikov & Ivanchei, 2016) and 2) scrambling of the ambiguous Dress picture reduces the 10 11 12 difference in perceived “blueness” on the ambiguous Dress picture between viewer types 13 14 (Hesslinger & Carbon, 2016). If background information was more important, we could expect the 15 16 17 differences between viewer types to disappear in both decontextualized conditions (patches and 18 19 strips). If information about the Dress pattern was more important, we could expect the differences 20 21 between viewer types to disappear in the patches condition but remain present (although 22 23 24 potentially reduced) in the strips condition. Finally, when comparing the colorimetric values of the 25 26 colours matched by participants and displayed on the ambiguous Dress picture, we could test which 27 28 viewer types determined colours with the colour picker that were the closest to the displayed 29 30 31 colours on the ambiguous Dress picture. In other words, we identified which viewer types 32 33 overestimated or underestimated any of the perceptual attributes of colours as compared to the 34 35 colorimetric signal of the ambiguous Dress picture. Based on the colour measurements of the 36 37 38 ambiguous Dress image (Melgosa et al., 2015), we could expect that the colour matches will be the 39 40 closest of B&Br viewers while the colour matches of two more extreme perceptions (B&B and W&G) 41 42 43 will deviate from the displayed colours. 44 45 46 47 48 49 50 Study One: Demographic and other variables predicting the Dress perception 51 52 53 54 55 56 57 Participants 58 59 60 61 62 63 64 -9- 65 #The Dress: decontextualized 10 In the classroom, we tested a convenience sample of 240 undergraduate psychology students (52 1 2 males). They had a mean age of 21.91 years (SD = 4.61, range = 18-54 years). Three participants self- 3 4 5 reported being colour blind, so they were excluded from the analyses. 6 7 8 Materials and Procedure 9 10 11 12 In a large lecture theatre (up to 500 places), students attended their weekly cognitive psychology 13 14 lecture. Right at the start, just four days after the ambiguous Dress picture appeared on the Internet, 15 16 17 they were invited to voluntarily take part in a survey on the Dress illusion. On two large projection 18 19 screens (next to each other), we showed the ambiguous Dress picture. The ambiguous Dress picture 20 21 was not calibrated to the screen because a) all participants were tested at the same time looking at 22 23 24 the same monitors, and b) there was no clear way to calibrate the presentation of the ambiguous 25 26 Dress image due to the lack of established perceptually relevant colorimetric values of the Dress 27 28 image. We distributed a short questionnaire asking 1) the colour(s) of the Dress; 2) whether they 29 30 31 have seen the Dress before (if yes they were labelled non-naïve), 3) whether the colours have 32 33 changed since previous viewing (if yes they were labelled unstable); 4) where the light source seems 34 35 36 to come from; 5) what type of light there seems to be (natural or artificial); 6) whether they had 37 38 more or fewer hours of sleep than usual (how many) the night before; 7) whether they had 39 40 consumed more or less coffee/black tea than usual on the day of testing; 8) what is their eye colour; 41 42 43 9) whether they are wearing glasses/contact lenses; and 10) demographic questions: age, gender, 44 45 ethnical background. Participants took around 10 minutes to complete the questionnaire while 46 47 looking at the ambiguous Dress image, upon which they were thanked and debriefed. We presented 48 49 50 them the results of the survey about 4 weeks later. 51 52 53 Design and analysis 54 55 56 57 We analysed the data according to the different viewer types (B&B, W&G, B&Br, and Other). We 58 59 used a chi-square test to compare the distribution of people in each viewer type group, and for 60 61 62 63 64 - 10 - 65 #The Dress: decontextualized 11 different categories: 1) naïve vs. non-naïve participants, 2) stable vs. unstable perception. Then, we 1 2 created a logistic regression model with several categorical and linear variables as predictors and 3 4 5 viewer type (B&B or W&G) as an outcome variable. In this model, only the number of B&B viewers 6 7 was compared to the number of W&G viewers to avoid biases due to low numbers of responses in 8 9 the other categories. These were the predictor variables: 1) gender (male, female), 2) ethnicity 10 11 12 (Caucasian, Asian, African, Latin American, or mixed), 3) type of light (natural, artificial, both), 4) light 13 14 source in the image (behind, side, other), 5) sleep deprivation (linear: the number of hours slept the 15 16 17 night before which were over or under the usual number), 6) caffeine consumption (more than 18 19 usual, as usual, less than usual), 7) eye colour (blue, green, brown, mixed), 8) corrected vision 20 21 (glasses or lenses vs. no correction), 9) age (linear variable). Data can be publicly accessed following 22 23 24 this link: https://forsbase.unil.ch/project/study-public-overview/15066/0/ 25 26 27 Results 28 29 30 31 The majority of participants saw the ambiguous Dress picture in B&B (47 %) or W&G (38%), and a 32 33 few participants reported other colour combinations, namely B&Br and other (see Table 1, column 34 35 36 Number of participants). Thus, the distribution of participants across all different viewer types varied 37 38 significantly, 2(3) = 121.84, p < .001, V = .414. When comparing B&B and W&G viewers only, the 39 40 number of participants in the B&B and W&G groups did not differ, 2(1) = 2.40, p = .121, V = .109. 41 42 43 44 45 46 47 48 [Insert Table One around here] 49 50 51 52 53 54 55 Most participants had seen the ambiguous Dress picture already (89% of non-naïve viewers, Table 1, 56 57 column Seen previously). The number of non-naïve viewers did not differ per viewer type; 2(3) = 58 59 60 3.51, p = .319, V = .122. Around one third of non-naïve viewers reported that the colours they saw 61 62 63 64 - 11 - 65 #The Dress: decontextualized 12 during the experiment were different from the colours they had seen initially (Table 1, column 1 2 Change colour). Around half of the unstable perceivers belonging to B&B, B&Br or Other viewer 3 4 5 types had experienced a change in the perception of the ambiguous Dress colours, while only 16% of 6 7 W&G viewers had experienced such change, and this difference was significant; 2(3) = 22.83, p < 8 9 .001, V = .333 (Table 1). By self-report (Table 1, column Previous colour), the majority of the unstable 10 11 12 viewers had previously seen the ambiguous Dress picture as W&G but it changed to B&B or B&Br 13 14 before or during the current testing session. Consequently, the change between W&G to B&B or 15 16 17 B&Br was more common than from B&B or B&Br to W&G; 2(1) = 15.78, p < .001, V = .478 18 19 (comparing the change from W&G vs. B&B and B&Br). The logistic regression with all the 20 21 demographic predictors was not significant overall (2(17) = 14.20, p = .653) and none of the 22 23 24 predictors were significant on their own (see Table 2). 25 26 27 28 29 30 31 [Insert Table Two around here] 32 33 34 35 36 37 38 Study Two: Context effects on the Dress colour reproduction 39 40 41 42 Participants 43 44 45 46 We tested 57 participants (10 males) with a mean age of 21.30 years (SD = 4.43, range 17-42 years). 47 48 We performed a sample size power analysis (Mayr, Buchner, Erdfelder, & Faul, 2007) for a 2 x 3 x 3 49 50 mixed measures ANOVA (for more information, see Data and Analysis section). This analysis 51 52 53 suggested that at an alpha level of 0.050 and beta level of 0.950, and assuming a correlation 54 55 between repeated measures of 0.5 and epsilon of 1, the total sample size of 54 is sufficient to detect 56 57 a medium effect size of 0.25. None of the participants was colour blind as confirmed using the 58 59 60 Ishihara’s Colour Blindness Test (Ishihara, 1993). Some of the participants received course credit for 61 62 63 64 - 12 - 65 #The Dress: decontextualized 13 participation while others were remunerated in gift vouchers. Participants were free to choose a 1 2 preferred form of gratification. All participants provided written informed consent prior to 3 4 5 experimentation. 6 7 8 Materials 9 10 11 12 To test effect of decontextualization, we presented three types of stimuli taken from the Dress: (a) 13 14 small patches cut out of different parts of the ambiguous Dress picture, evenly distributed (Figure 1, 15 16 17 left), (b) vertical strips cut out of the ambiguous Dress picture (Figure 1, middle), and (c) the full 18 19 ambiguous Dress picture (Figure 1, right). They varied from the least to the most recognisable, and 20 21 for this reason were always presented in the same order: patches, strips, and the full ambiguous 22 23 24 Dress picture. Also, surrounding context information was eliminated in the patches and strips 25 26 condition but preserved in the full ambiguous Dress picture condition. Contrast information was 27 28 reduced in the patches condition but preserved in the strips and the full ambiguous Dress picture 29 30 31 conditions. The exact locations from which the patches and the strips were cut out are presented in 32 33 Figure S1. 34 35 36 37 The least recognizable type of stimulus was the patch. Ten small patches were selected from the 38 39 ambiguous Dress picture at various locations to correspond to white/blue and gold/black regions of 40 41 42 the ambiguous Dress picture (see Figure 1, left, and Figure S1). The selected regions were similar to 43 44 those investigated by Melgosa and colleagues (2015). The patches were displayed in the centre of 45 46 the screen, one at a time. Four of these patches were cut from the lighter part of the ambiguous 47 48 49 Dress picture (white or blue, depending on perception) – ”Light”, and six of these patches were cut 50 51 from the darker part of the ambiguous Dress picture (gold or black) – “Dark”. We used six additional 52 53 patches of comparable size taken from the pictures of two salamanders as control (see Figure 1 left 54 55 56 and Figure 2). We chose salamanders instead of homogeneous single colour patches because these 57 58 salamander patches were perceptually complex and acted as lure patches. We decided to use such 59 60 61 62 63 64 - 13 - 65 #The Dress: decontextualized 14 “lure” patches to make the goal of the study (testing the ambiguous Dress picture) less obvious at 1 2 this initial stage. All patches (test and control) were treated such that they encompassed the space 3 4 5 of 2 x 2 cm (14,000 pixels) and a visual angle of 1.64 (see Table S1 in supplementary material). 6 7 8 The second type of stimulus, which we presented after the patches, was the strip. Two strips were 9 10 11 vertically cut out from two ambiguous Dress locations (see Figure 1 middle and Figure S1), 12 13 potentially recognizable as belonging to the ambiguous Dress picture by non-naïve viewers (similar 14 15 stimuli used in Drissi Daoudi et al., 2017). The strips were not stretched. Their size was 19.1 x 1.3 cm, 16 17 18 and they were presented in the centre of the screen. The vertical viewing angle was 15.6 and the 19 20 horizontal viewing angle was 0.01. The third type of stimulus, presented after the strips, was the 21 22 23 full ambiguous Dress picture (see Figure 1, right). It was not manipulated in any way and presented 24 25 exactly as it circulated the Internet. The picture size was 19.1 x 12.6 cm, the vertical viewing angle 26 27 was 15.6 and the horizontal viewing angle was 10.2. After the three experimental conditions, we 28 29 30 further presented a picture of the same dress in unambiguous lightening conditions as control (“the 31 32 real dress”, see Figure 2). All the stimuli appeared on a neutral grey background in the centre of the 33 34 35 screen. Participants were seated in a dark room, lit by a computer monitor only. 36 37 38 Colour picker 39 40 41 42 The colour picker provides a structured representation of colour patches showing eight variations 43 44 (lighter, darker, more or less chromatic, more yellow, more blue, more red and more green) of the 45 46 47 centrally presented colour patch shown on white background on the computer screen (Jonauskaite 48 49 et al., 2016). It starts with the presentation of nine (red, orange, yellow, yellow-green, green, green- 50 51 blue, blue, purple and grey) square colour patches on a white background (14.5 x 13.0 cm). In the 52 53 54 current study, the eye-screen distance was kept constant at 70 cm. Thus, the colour picker extended 55 56 to a vertical visual angle of 11.8 and a horizontal visual angle of 10.6. After making an initial choice, 57 58 participants had to narrow down their colour selection by clicking on a patch that most closely 59 60 61 62 63 64 - 14 - 65 #The Dress: decontextualized 15 resembles the colour of the patch they were trying to match. The colour selection process as such is 1 2 sequential so that with each selection the target (central) colour varies along various colour 3 4 5 dimensions to a lesser degree. Once the variations reach a threshold, the central colour is singled- 6 7 out as the final selection. It is then displayed in the centre of the screen on a white background; its 8 9 size is 4.7 x 4.0 cm, making a vertical visual angle of 3.8 and a horizontal visual angle of 3.4. The 10 11 12 colour picker records the colour in RGB values. The key feature of this colour picker tool is that it 13 14 allows a user-friendly, intuitive, and fast selection of any colour that the computer screen can 15 16 17 produce. 18 19 20 Apparatus 21 22 23 24 The task was performed on three different monitors with identical specifications (Colour Edge 25 26 CG243W 24.1" Widescreen LCD display). The white point of monitor one was (0.326, 0.345, 107.95); 27 28 29 monitor two (0.326, 0.347, 102.64); and monitor three (0.326, 0.345, 98.96) in CIE xyY colour space. 30 31 Similarly, the empirical primaries of monitor one were Red (R) = (0.642, 0.333, 23.9), Green (G) = 32 33 (0.305, 0.581, 78.60), and Blue (B) = (0.140, 0.051, 5.45), monitor two – R = (0.641, 0.332, 22.1), G = 34 35 36 (0.304, 0.584, 75.5), and B = (0.140, 0.049, 5.04), and monitor three – R = (0.641, 0.331, 21.9), G = 37 38 (0.304, 0.583, 72.1), and B = (0.140, 0.050, 4.96). The viewing distance was approximately 70 cm in 39 40 all cases. 41 42 43 44 Conversion of colour parameters 45 46 47 48 The colour selections were recorded with the colour picker tool as device-dependent RGB values. 49 50 We converted the RGB values into CIE Lab and then CIE LCh values, which are device-independent, 51 52 and thus more realistically describe actually perceived colours. This kind of conversion cannot be 53 54 55 achieved purely arithmetically because the relationship between the device emitted luminance and 56 57 the perceiver observed lightness is not linear. This non-linear relationship can be described by the 58 59 gamma curve (Robson, 1999). In order to calculate the gamma curve, we used a Konica Minolta 60 61 62 63 64 - 15 - 65 #The Dress: decontextualized 16 chromameter to measure Yxy values of the emitted colour from red, green and blue guns at 1 2 different voltages for each of our three monitors and convert them into CIE LCh values. When 3 4 5 performing the conversions, we took into account the monitor settings of the monitor that each 6 7 participant performed the experiment on. 8 9 10 11 12 13 14 [Insert Figure One around here] 15 16 17 18 19 20 21 Procedure 22 23 24 25 The main task of the experiment was to reproduce the colours of each type of stimulus using the 26 27 computerised colour picker program. Participants were not aware that the experiment was designed 28 29 30 to test perception of the ambiguous Dress picture. The only information they had was that the 31 32 experiment tested “colour perception”. To keep the participants naïve as long as possible, we 33 34 35 presented the conditions regarding contextualisation of the ambiguous Dress picture in this order: 36 37 patches, strips and the full ambiguous Dress picture. The 16 patches were individually shown in the 38 39 same order with dark, light and control patches inter-mixed (see the order in Figure 1, left). 40 41 42 Participants were asked to use the colour picker (displayed on the right side of the screen next to the 43 44 stimuli) and select the predominant colour of the stimuli displayed on the left side of the screen. 45 46 Reference white was always available, which was considered to be the adapting monitor white- 47 48 49 point. The presentation of the left strip, the right strip and the full ambiguous Dress picture followed 50 51 the patches. Participants used the colour picker to first match the colours of the left and right strips 52 53 54 cut from the ambiguous Dress picture (Figure 1, middle) and then of the full ambiguous Dress 55 56 picture (Figure 1, right). 57 58 59 60 61 62 63 64 - 16 - 65 #The Dress: decontextualized 17 Afterwards, participants completed a short demographics questionnaire and were asked what 1 2 colour(s) they saw on the ambiguous Dress picture, whether they had seen the picture before, 3 4 5 whether its colour changed from previous viewing (specify), and what colour(s) they saw on the real 6 7 dress. Based on the answer to the first question, participants were assigned into four self-reported 8 9 viewer types: B&B, W&G, B&Br, or other. The experiment took around 50 minutes. Afterwards, 10 11 12 participants received a voucher (10 CHF value) or two experimental participation points. Finally, they 13 14 were thanked and fully debriefed. All participants were recruited from the same first year 15 16 17 psychology student pool. They were free to choose their preferred method of remuneration. Around 18 19 two thirds of the participants chose experimental participation points as remuneration. 20 21 22 Design and analysis 23 24 25 26 We had three independent variables (IVs) in this dataset. The first IV, between-subjects, was self- 27 28 reported VIEWER TYPE with three levels: B&B, W&G, and B&Br. The second IV, within-subjects, was 29 30 31 the PATCH COLOUR with two levels: dark (i.e. black or gold) and light (i.e. blue or white). The third 32 33 IV, within-subjects, was CONTEXT with three levels: squares (i.e. out of context), strips (i.e. partially 34 35 36 contextualized), and full Dress (full context of the ambiguous Dress picture). The colour selections 37 38 for patches and strips were averaged across the different selections (i.e. six dark patch selections, 39 40 four light patch selections, two dark strip and two light strip selections). We excluded two 41 42 43 participants because of missing data. We further excluded three participants whose self-reported 44 45 perception of the ambiguous Dress picture could not be categorised into one of the three viewer 46 47 types – B&B, W&G or B&Br. Thus, all analyses were performed on 52 participants (9 males). 48 49 50 51 We broke down colour selections into three colour parameters making three dependent variables 52 53 (DVs): CIE LCh hue (range 0°-360°), CIE LCh chroma (1-141) and CIE LCh lightness (1-100). Hue was a 54 55 56 circular variable while lightness and chroma were linear variables. Differences in hue were analysed 57 58 using one criterion analysis of variance (ANOVA) for circular data (Agostinelli & Lund, 2017). This 59 60 61 62 63 64 - 17 - 65 #The Dress: decontextualized 18 method assumes that the samples are drawn from the von Mises distribution and tests whether all 1 2 groups (here B&B, B&Br, and W&G) have the same mean direction (null hypothesis). We could 3 4 5 implement circular statistics, because the hue data was unimodally distributed. We compiled six 6 7 ANOVA models for circular data, Bonferroni corrected (i.e., p values were multiplied by the number 8 9 of tests – i.e., six), to compare hue matches between the viewer types for dark and light patches, 10 11 12 dark and light strips, and dark and light parts of the full ambiguous Dress image. Additionally, to test 13 14 how colour categories match the circular hue angles, we binned hues into nine perceptually relevant 15 16 17 categories (Jonauskaite et al., 2016; Parraga & Akbarinia, 2016): red (346°-40°], orange (40°-72°], 18 19 yellow (72°-105°], yellow-green (105°-130°], green (130°-166°], green-blue (166°-220°], blue (220°- 20 21 275°], and purple (275°-346°]. All these categories had chroma values above 5, while the last 22 23 24 category – achromatic – had chroma values below 5 of any hue angle. We compared hue categories 25 26 matched to the dark parts of the Dress and the light parts of the Dress using the marginal model for 27 28 correlated multinomial responses (Bergsma, Croon, & Hagenaars, 2009; Bergsma & Van der Ark, 29 30 31 2015). 32 33 34 Differences in chroma and lightness levels were analysed using respectively a 2 x 3 x 3 mixed ANOVA 35 36 37 with PATCH COLOUR and CONTEXT as within-subjects measure and VIEWER TYPE as between- 38 39 subject measure. To account for the notion that the Dress phenomenon is continuous (Gegenfurtner 40 41 et al., 2015), whenever significant interactions included VIEWER TYPE, the interactions were broken 42 43 44 down with planned linear polynomial contrasts by arranging the viewers as B&B – B&Br – W&G. 45 46 Finally, we used one-sample t-tests to compare the colorimetric values of hue, chroma and lightness 47 48 49 of the selected colours (colour picker) and what was displayed from the ambiguous Dress image. All 50 51 analyses were performed with R (R Core Team, 2018) and SPSS (IBM Corp, 2013) statistical software 52 53 programs. Data can be publicly accessed following this link: https://forsbase.unil.ch/project/study- 54 55 56 public-overview/15066/0/. 57 58 59 60 61 62 63 64 - 18 - 65 #The Dress: decontextualized 19 Results 1 2 3 4 In analogy to study one, we show the distribution of viewer types, naïve participants and those 5 6 having reported unstable perceptions of the ambiguous Dress picture (Table 3). Statistically, all three 7 8 viewer types occurred at the same frequency, 2(2) = 3.50, p = .174, V = .259. That is, participants 9 10 11 were equally likely to perceive the ambiguous Dress picture colours as B&B, W&G, or B&Br. Other 12 13 comparisons were not made due to an insufficient number of cases in each bin. 14 15 16 17 18 19 20 [Insert Table Three around here] 21 22 23 24 25 26 27 Colour selections as a function of hue, lightness, and chroma values 28 29 30 31 Figure 3 shows the colours that participants matched to the darker and the lighter parts of the 32 33 ambiguous Dress picture, its patches and strips. Simply descriptive, colour matches to the 34 35 36 ambiguous Dress image seem more heterogeneous than colour matches to control patches or the 37 38 real dress image (Figure 2). For comparison with other research studies, we further represent colour 39 40 41 selections in CIE Lab colour space (see Figure S2). 42 43 44 45 46 47 48 [Insert Figure Two around here] 49 50 51 52 53 54 55 [Insert Figure Three around here] 56 57 58 59 60 61 62 63 64 - 19 - 65 #The Dress: decontextualized 20 ANOVA for circular data on hue showed no difference between VIEWER TYPES for patch condition 1 2 (dark: F(2, 49) = 0.83, p = 1.000,  p2 = .031; light: F(2, 49) = 0.25, p = 1.000,  p2 = . 010), strips 3 4 5 condition (dark: F(2, 49) = 0.57, p = 1.000, p 2 = .023; light; F(2, 49) = 0.24, p = 1.000, p 2 = .009), or 6 7 the light parts of the ambiguous Dress image (F(2, 49) = 2.53, p = .539, p 2 = .090). The difference 8 9 10 between VIEWER TYPES emerged for the dark parts of the ambiguous Dress image (F(2, 49) = 9.51, p 11 12 = .002, p2 = .257), whereby B&B viewers matched redder hues (hues closer to 0°) compared to 13 14 colour matches by B&Br (p = .004) or W&G (p = .009) viewers. Colour matches were statistically 15 16 17 indistinguishable between B&Br and W&G viewers (p = 1.000; see Figure 4). 18 19 20 21 22 23 24 [Insert Figure Four around here] 25 26 27 28 29 30 31 Additional analyses with the marginal model for correlated multinomial responses showed that hue 32 33 34 categories differed between PATCH COLOURS (χ² (7) = 296.59, p < .001) and CONTEXT conditions (χ² 35 36 (14) = 46.01, p < .001) but not VIEWER TYPES (χ² (14) = 20.63, p = .111). Standardised residuals 37 38 indicated that the dark parts were matched to orange (p < .001) and yellow (p < .001) hues 39 40 41 significantly more often than the light parts, while the light parts were matched to blue (p < .001) 42 43 and purple (p < .001) hues significantly more often than the dark parts across all three context 44 45 conditions of the ambiguous Dress picture (Figure S3). Also, orange hue was significantly more often 46 47 48 used for PATCH colours in patches than for strips or full ambiguous Dress picture colours (p < .001) 49 50 while achromatic colours were significantly more often used for the full ambiguous Dress picture 51 52 53 colours than for patch or strip colours. 54 55 56 The 2x2x3 ANOVA on lightness levels showed a main effect of VIEWER TYPE (F(2, 49) = 16.73, p < 57 58 59 .001, p2 = .406). Polynomial linear contrast indicated that B&B viewers’ matched colours were the 60 61 62 63 64 - 20 - 65 #The Dress: decontextualized 21 darkest while W&G viewers’ matched colours were the lightest (p < .001). Importantly, there was a 1 2 significant three-way interaction between VIEWER TYPE, CONTEXT and PATCH COLOUR pointing to 3 4 5 the importance of context in which the ambiguous Dress was originally embedded; F(4, 98) = 292.95, 6 7 p < .001,  p2 = .165 (see Figure 5, panel A). We compared the differences in lightness of the matched 8 9 colours between B&B, B&Br, and W&G viewers separately for each context condition (i.e., patch, 10 11 12 strip, and the ambiguous Dress image) using linear polynomial contrasts. There were no differences 13 14 between VIEWER TYPES for patch condition, whether matching dark (p = .922) or light (p = .940) 15 16 17 parts of the patches cut out of the ambiguous Dress picture. There were, however, differences 18 19 between VIEWER TYPES for strip (dark: p = .028; light: p = .021) and the ambiguous Dress image 20 21 (dark: p < .001; light: p < .001) conditions. In both strip and the Dress conditions, B&B viewers 22 23 24 matched the darkest colours and W&G viewers matched the lightest colours while colour matches of 25 26 B&Br viewers fell in between the two (see Figure 5, panel A). 27 28 29 30 31 32 33 [Insert Figure Five around here] 34 35 36 37 38 39 40 The 2x2x3 ANOVA on chroma levels did not show a main effect of VIEWER TYPE (F(2, 49) = 1.99, p = 41 42 43 .148,  p2 = .075). However, as for lightness, there was a significant three-way interaction between 44 45 VIEWER TYPE, CONTEXT and PATCH COLOUR; F(4, 98) = 1305.61, p < .001,  p 2 = .408 (see Figure 5, 46 47 48 panel B). Similar to results obtained for lightness, there were no differences between VIEWER TYPES 49 50 for patch condition when matching dark (p = .401) and light (p = .977) parts of the patches cut out of 51 52 the ambiguous Dress picture. In the strip condition, a linear contrast was significant between 53 54 55 VIEWER TYPES for light (p = .007) but not dark (p = .134) parts of the strips cut out of the ambiguous 56 57 Dress picture. B&B viewers matched the most chromatic light colours and W&G viewers matched 58 59 60 the least chromatic light colours, while colour matches of B&Br viewers fell in between. In the 61 62 63 64 - 21 - 65 #The Dress: decontextualized 22 ambiguous Dress condition, a linear contrast was significant between VIEWER TYPES for dark (p < 1 2 .001) and light (p < .001) parts of the ambiguous Dress picture. B&B viewers matched the least 3 4 5 chromatic colours for the dark parts and the most chromatic colours for the light parts, W&G 6 7 viewers matched the most chromatic colours for the dark parts and the least chromatic colours for 8 9 the light parts, while colour matches of B&Br viewers fell between matches of B&B and W&G 10 11 12 viewers (see Figure 5, panel B). 13 14 15 16 17 18 19 Deviation of selected colours from presented (reference) colours 20 21 22 23 We used the Konica Minolta CS-100A chroma meter to measure the colorimetric signal of the 24 25 displayed colours on the ambiguous Dress picture. We averaged across six points of the dark parts of 26 27 the ambiguous Dress picture and six points of the light parts of the ambiguous Dress picture. After 28 29 30 having done this separately on the three monitors, we averaged the CIE LCh values (obtained from 31 32 xyY values) of the three monitors and established the reference colour for the dark parts of the 33 34 ambiguous Dress picture and the reference colour for the light parts of the ambiguous Dress picture 35 36 37 (see Table 4 column Reference value). For averaging hue values, we implemented circular statistics 38 39 (Agostinelli & Lund, 2017). We compared lightness, chroma, and hue values of the matched colours 40 41 42 with the reference colours, using one-sample t-tests for lightness and chroma, and one-sample 43 44 ANOVA for circular data for hue. The comparisons were Bonferroni corrected, for lightness, chroma, 45 46 and hue separately. This comparison gave us an indication of the extent of over- or underestimation 47 48 49 of colour values by the viewers. 50 51 52 The majority of the colours matched by the viewers to the colours displayed on the ambiguous Dress 53 54 55 picture deviated from the reference colours (Table 4). In general, B&B viewers matched lightness, 56 57 chroma and hue the closest to the reference colorimetric values, except from chroma of the light 58 59 parts of the ambiguous Dress picture (“blue / white”), which was slightly overestimated. The other 60 61 62 63 64 - 22 - 65 #The Dress: decontextualized 23 two viewer types (i.e., W&G and B&Br) largely overestimated lightness of the light and the dark 1 2 parts of the ambiguous Dress picture, and overestimated chroma of the dark parts. B&Br viewers, 3 4 5 but not W&G viewers also overestimated chroma of the light parts of the ambiguous Dress picture. 6 7 Finally, B&Br and W&G viewers matched hue of the dark parts of the ambiguous Dress picture as 8 9 more yellow – their matches were shifted clockwise from the reference values. The observations are 10 11 12 in line with the above reported ANOVA results. These results highlight the extent of the Dress 13 14 illusion, as viewers overestimated colour parameters by 30% - 410% when compared to the 15 16 17 displayed colorimetric values. 18 19 20 21 22 23 24 [Insert Table Four around here] 25 26 27 28 29 30 31 Discussion 32 33 34 35 The Dress picture went viral in early 2015 with people dividing into two perceptual camps – Blue and 36 37 Black (B&B) and White and Gold (W&G) viewers. Subsequently, the scientific and popular media 38 39 tried to explain this new visual illusion (Brainard & Hurlbert, 2015). We performed two studies right 40 41 42 after the appearance of the ambiguous Dress picture online. Here, we investigated what variables 43 44 can predict the colours that people see on the ambiguous Dress picture and how colour perception 45 46 of different viewer types changes with increase of contextual information of the ambiguous Dress 47 48 49 picture. In the first study, we investigated the proportion of different viewer types when seeing the 50 51 same ambiguous Dress picture at the same time in a classroom setting and whether inter-individual 52 53 54 variables could explain the differences in the perception of the ambiguous Dress picture. In the 55 56 subsequent laboratory study, participants used a computer-based colour picker program to match 57 58 perceived colours for light and dark parts of the ambiguous Dress picture, in that order, for patches, 59 60 61 62 63 64 - 23 - 65 #The Dress: decontextualized 24 cut out of the ambiguous Dress picture, strips also cut out of the ambiguous Dress picture, and the 1 2 whole ambiguous Dress picture. Here, we could test 1) whether perceptual attributes of colour (hue, 3 4 5 chroma, and lightness) differed between viewer types; 2) whether the illusion persisted when the 6 7 ambiguous Dress picture was decontextualized; and 3) which viewer type matched colours the 8 9 closest to the displayed colours on the ambiguous Dress picture. 10 11 12 13 Results from the classroom study yielded three self-reported viewer types: the expected B&B and 14 15 W&G viewers, but also the previously mentioned group of B&Br viewers. The frequency of particular 16 17 18 viewer types varies between studies (Chetverikov & Ivanchei, 2016; González Martín-Moro et al., 19 20 2018; Karlsson & Allwood, 2016; Wallisch, 2017). Using an unconstrained colour term choice 21 22 methodology in the current study, we report an approximately even split between B&B and W&G 23 24 25 viewers (similar to Aston & Hurlbert, 2017; Chetverikov & Ivanchei, 2016; Hesslinger & Carbon, 26 27 2016; Karlsson & Allwood, 2016; Schlaffke et al., 2015; Vemuri et al., 2016; Winkler et al., 2015). We 28 29 30 further confirm that several individuals reported seeing the ambiguous Dress picture as B&Br (Aston 31 32 & Hurlbert, 2017; Lafer-Sousa et al., 2015; Mahroo et al., 2017; Wallisch, 2017; Witzel et al., 2017). 33 34 While being conjectural, a bimodal distribution of Dress viewers might have arisen due to using 35 36 37 forced-choices when testing participants’ colour perception (i.e., asking them to choose between 38 39 B&B and W&G) (see also Aston & Hurlbert, 2017; Lafer-Sousa et al., 2015; Wallisch, 2017; Witzel et 40 41 al., 2017). Thus, our results support previous notions of a perceptual continuum (e.g., Gegenfurtner 42 43 44 et al., 2015) instead of a bimodal distribution of B&B and W&G viewers (e.g., Chetverikov & Ivanchei, 45 46 2016). 47 48 49 50 B&B and B&Br viewers in the classroom study were more likely than W&G viewers to have 51 52 experienced a change in the perception of the ambiguous Dress colours (see also Lafer-Sousa et al., 53 54 55 2015). During the classroom study as compared to prior viewing(s), we found that B&B and B&Br 56 57 viewers reported a change in perception, i.e. not seeing the ambiguous Dress picture as W&G 58 59 anymore, but as B&B or B&Br. This observation could explain why the initial Internet survey with 3 60 61 62 63 64 - 24 - 65 #The Dress: decontextualized 25 million participants (Holderness, 2015) reported a larger number of participants seeing the 1 2 ambiguous Dress picture as W&G while later studies reported a more even split between B&B and 3 4 5 W&G viewers (e.g., Karlsson & Allwood, 2016). Some initial W&G viewers might have become B&B 6 7 or B&Br viewers after repeated exposure to this image. 8 9 10 11 In the laboratory study, one of the questions we wanted to know was whether perceptual attributes 12 13 of colour (hue, chroma, and lightness) differed between viewer types. We answer yes, at least 14 15 partially. Our data supports the notion of a continuum of Dress viewer types: B&B viewers 16 17 18 reproduced darker ambiguous Dress picture colours than B&Br viewers, which reproduced darker 19 20 ambiguous Dress picture colours than W&G viewers (see also Aston & Hurlbert, 2017; Chetverikov & 21 22 Ivanchei, 2016; Gegenfurtner et al., 2015; Lafer-Sousa et al., 2015; Toscani et al., 2017; Witzel et al., 23 24 25 2017). When matching colours to dark (i.e., black/gold) and light (i.e., blue/white) parts of the 26 27 ambiguous Dress image, B&B viewers reproduced the dark ones as more chromatic and light ones as 28 29 30 less chromatic when compared to B&Br and then to W&G viewers. While the differences between 31 32 lightness and chroma were rather large, we found fewer differences in reproduced hue. 33 34 35 36 For hue, the only difference between the viewer types emerged for the reproduced hue of the dark 37 38 parts of the ambiguous Dress image. B&B viewers matched hues that were closer to a hue angle of 39 40 0°, i.e. redder, when compared to B&Br or W&G viewers. All viewers together, almost always 41 42 43 selected the dark parts of the ambiguous Dress picture as having a yellow or an orange hue, and the 44 45 light parts of the ambiguous Dress picture as having a blue or a purple hue. While hue differences 46 47 were reported for reproduced Dress colours using various colour models (Aston & Hurlbert, 2017; 48 49 50 Chetverikov & Ivanchei, 2016; Lafer-Sousa et al., 2015; Witzel et al., 2017), the current study shows 51 52 that differences in hue perceptions are less detectable and obvious regarding the subjective report 53 54 55 than in chroma and lightness. 56 57 58 Chroma and lightness differences between the viewer types can be easily compared with their 59 60 61 62 63 64 - 25 - 65 #The Dress: decontextualized 26 subjective perceptions. Viewers who reported the lighter parts of the ambiguous Dress picture as 1 2 being blue (i.e., B&B and B&Br viewers) perceptually matched darker colours and so accentuated the 3 4 5 blue/purple hue. W&G viewers, who reported the lighter parts of the ambiguous Dress picture as 6 7 being white, perceptually matched blue/purple hues as well but chose a very light version of them. 8 9 This choice explains why the very light blue was named as white. Viewers who reported the darker 10 11 12 parts of the ambiguous Dress picture as being gold or brown (i.e., B&Br and W&G viewers) 13 14 perceptually matched lighter colours and so accentuated the yellow/orange hue. B&B viewers 15 16 17 matched much darker and redder hues to the same parts of the ambiguous Dress picture and called 18 19 them black. The redder hues might potentially be an artefact of choosing a very dark colour. Due to 20 21 a cylindrical representation of the LCh colour space, colours of low lightness and chroma have poorly 22 23 24 defined hue. A perfect black may by default be represented by the hue angle of 0°. However, this 25 26 does not explain why B&B viewers matched a wide range of reddish hues, ranging from almost 27 28 purplish (anti-clockwise from 0°) to red-orange (clockwise from 0°). In any case, the matched colours 29 30 31 were very dark and perceptually appeared as almost black. 32 33 34 In the laboratory study, we also wanted to know whether the Dress illusion persisted when the 35 36 37 ambiguous Dress picture was decontextualized. We again answer yes, at least partially. The 38 39 differences in matched colours were most apparent for the contextualized ambiguous Dress picture 40 41 (i.e., the original stimulus), somewhat less apparent for the strips and not apparent for the patches 42 43 44 cut out of the ambiguous Dress picture. We observed a linear arrangement in the matched colours 45 46 from B&B to B&Br to W&G viewers when the ambiguous Dress picture was presented as a strip as 47 48 49 well as the whole ambiguous Dress picture. Hence, results indicate that differences between 50 51 perceptions of the ambiguous Dress picture colours (i.e., viewer types) already appear in minimal 52 53 context conditions. In other words, a reduced section of the ambiguous Dress picture is enough for 54 55 56 the illusion to emerge. Emergence (although weaker) of the Dress illusion in the strip condition (the 57 58 background was occluded) demonstrated that the surrounding context is not the essential driving 59 60 61 62 63 64 - 26 - 65 #The Dress: decontextualized 27 force for the Dress illusion to occur (Chetverikov & Ivanchei, 2016; Drissi Daoudi et al., 2017; 1 2 Wallisch, 2017; Witzel et al., 2017). A priori, our study was not designed to explain how context 3 4 5 affects colour perception of the ambiguous Dress image. To this goal, different hypotheses have 6 7 been proposed to explain the Dress illusion (e.g., Dixon & Shapiro, 2017; Hesslinger & Carbon, 2016; 8 9 Witzel et al., 2017). Nevertheless, our results would invite for further investigation of contrast 10 11 12 information (i.e., contrast between light and dark parts of the ambiguous Dress picture) as a 13 14 potential explanatory factor for the Dress illusion. Such future studies should be designed to address 15 16 17 the mechanisms of the Dress illusion and why individual differences in the ambiguous Dress 18 19 perception persist in scenes hardly recognisable as the ambiguous Dress image (as shown here and 20 21 in Hesslinger & Carbon, 2016). Potentially, the texture structure rather than the illuminated 22 23 24 naturalistic scene is crucial in eliciting perceptual differences that disappear for uniform patches of 25 26 similar chromaticities. 27 28 29 30 In the laboratory study, we finally wanted to know which viewer type matched colours the closest to 31 32 the actual colours displayed on the monitor. Although we had expected B&Br viewers’ colour 33 34 matches to be the closest to the displayed colours, since their matches always fall between the 35 36 37 other two viewer types (Aston & Hurlbert, 2017), we observed that B&B viewers’ colour matches 38 39 were the closest to the displayed colours. We compared lightness, chroma, and hue values of the 40 41 colours that participants matched with the colours of the ambiguous Dress picture that were 42 43 44 displayed on the monitors. Comparing these, we had an indication of how much subjective 45 46 perception of the colours of the ambiguous Dress picture deviated from the displayed colorimetric 47 48 49 values. We could determine which viewer types over- or underestimated the colorimetric 50 51 parameters of the ambiguous Dress picture. The colorimetric values of the B&B viewers’ 52 53 reproductions were the closest to the displayed values. Both W&G and B&Br viewers considerably 54 55 56 overestimated lightness and chroma of the ambiguous Dress picture in the matched colours. They 57 58 also matched colours shifted towards the yellow hue (i.e., away from 0°) compared to the displayed 59 60 61 62 63 64 - 27 - 65 #The Dress: decontextualized 28 colours. Considering that the ambiguous Dress picture displays bluish and brownish colours 1 2 (Melgosa et al., 2015), one could expect that these displayed colours (in terms of their lightness and 3 4 5 chroma) are most similar to what B&Br perceivers reproduce. Nonetheless, our current data 6 7 indicates that B&B perceivers most accurately match the displayed colours as compared to both 8 9 W&G and B&Br perceivers who over-estimate colorimetric parameters. The question opens as to 10 11 12 why the majority of people (i.e., those who perceive W&G and B&Br) cognitively resolve the 13 14 ambiguous stimulus so that it leads to an illusionary percept? The extent of the Dress illusion is 15 16 17 large; it affects the majority of people, and deserves further attention. 18 19 20 To summarize, we performed two studies on the ambiguous Dress picture to contribute to on-going 21 22 research efforts aimed at understanding this new visual illusion. In our first classroom study, 23 24 25 participants’ self-report resulted in the clustering of individuals into B&B, B&Br, or W&G viewers. 26 27 These groups did not differ according to subjectively perceived light source or inter-individual 28 29 30 factors. The laboratory study indicated that contextualization is key for the Dress illusion to occur 31 32 and contrast or texture information might be the driving force. Moreover, results on lightness and 33 34 chroma highlighted previous notions that the illusion occurs at a continuum from W&G, to B&Br, to 35 36 37 B&B viewers, with the latter being most closely in their colour matching to what has been actually 38 39 displayed on the monitor. Our results contribute to an increasing knowledge base on the Dress 40 41 illusion, but remains insufficient explaining why so many individuals (W&G, B&Br) over-estimate 42 43 44 colour parameters of the ambiguous Dress picture and only few individuals (B&B) perceive colours 45 46 close to what has actually been displayed. Even knowing that contextualization, contrast, and 47 48 49 illumination are important for the illusion to occur, these factors are not yet able to explain why 50 51 individuals differ so strongly in their subjective perception. While currently light source (Chetverikov 52 53 & Ivanchei, 2016) and one-shot learning (Drissi Daoudi et al., 2017) are some suggestions, we are 54 55 56 certain additional suggestions will be communicated in the years to come. 57 58 59 60 61 62 63 64 - 28 - 65 #The Dress: decontextualized 29 Compliance with Ethical Standards 1 2 3 4 Funding: The current study did not receive any specific funding. 5 6 7 Conflict of interest: None declared. 8 9 10 11 Ethical approval: All procedures performed in studies involving human participants were in 12 13 accordance with the ethical standards of the institutional and/or national research committee and 14 15 with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. No 16 17 18 specific ethical clearance was received for this study, as it was not required by the Canton of Vaud in 19 20 Switzerland law. 21 22 23 24 Informed consent: Informed consent was obtained from all individual participants. 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 - 29 - 65 #The Dress: decontextualized 30 References 1 2 3 4 Agostinelli, C., & Lund, U. (2017). R package “circular”: Circular statistics (version 0.4-93). Retrieved 5 6 from https://r-forge.r-project.org/projects/circular/ 7 8 9 Aston, S., & Hurlbert, A. (2017). What # theDress reveals about the role of illumination priors in color 10 11 perception and color constancy. Journal of Vision, 17(9), 1–18. 12 13 http://doi.org/10.1167/17.9.4.doi 14 15 16 Bergsma, W. P., Croon, M. A., & Hagenaars, J. A. (2009). Marginal models: For dependent, clustered, 17 18 19 and longitudinal categorical data. Springer. 20 21 Bergsma, W. P., & Van der Ark, L. A. (2015). cmm: Categorical marginal models. R package version 22 23 24 0.8. 25 26 27 Brainard, D. H., & Hurlbert, A. C. (2015). Colour vision: Understanding #TheDress. Current Biology, 28 29 25(13), R551–R554. http://doi.org/10.1016/j.cub.2015.05.020 30 31 32 Carbon, C. C., Grüter, M., & Grüter, T. (2013). Age-dependent face detection and face categorization 33 34 performance. PLoS ONE, 8(10), 1–7. http://doi.org/10.1371/journal.pone.0079164 35 36 37 Chetverikov, A., & Ivanchei, I. (2016). Seeing “the Dress” in the right light: Perceived colors and 38 39 inferred light sources. Perception, 45(8), 910–930. http://doi.org/10.1177/0301006616643664 40 41 42 Corney, D., & Lotto, R. B. (2007). What are lightness illusions and why do we see them? PLoS 43 44 45 Computational Biology, 3(9), 1790–1800. http://doi.org/10.1371/journal.pcbi.0030180 46 47 Dixon, E. L., & Shapiro, A. G. (2017). Spatial filtering, color constancy, and the color-changing dress. 48 49 50 Journal of Vision, 17(3), 1–20. http://doi.org/10.1167/17.3.7 51 52 53 Drissi Daoudi, L., Doerig, A., Parkosadze, K., Kunchulia, M., & Herzog, M. H. (2017). The role of one- 54 55 shot learning in #TheDress. Journal of Vision, 17(3), 1–7. http://doi.org/10.1167/17.3.15 56 57 58 Gegenfurtner, K. R., Bloj, M., & Toscani, M. (2015). The many colours of “the dress.” Current Biology, 59 60 25(13), R543–R544. http://doi.org/10.1016/j.cub.2015.04.043 61 62 63 64 - 30 - 65 #The Dress: decontextualized 31 González Martín-Moro, J., Prieto Garrido, F., Gómez Sanz, F., Fuentes Vega, I., Castro Rebollo, M., & 1 2 Moreno Martín, P. (2018). Which are the colors of the dress? Review of an atypical optic 3 4 5 illusion. Archivos de La Sociedad Española de Oftalmología (English Edition), 93(4), 186–192. 6 7 http://doi.org/10.1016/j.oftale.2018.02.003 8 9 10 Hesslinger, V. M., & Carbon, C.-C. (2016). #TheDress: The role of illumination information and 11 12 individual differences in the psychophysics of perceiving white-blue ambiguities. I-Perception, 13 14 15 7(2), 1–10. http://doi.org/10.1177/2041669516645592 16 17 Holderness, C. (2015). What colours are this dress? Retrieved January 30, 2018, from 18 19 20 https://www.buzzfeed.com/catesish/help-am-i-going-insane-its-definitely- 21 22 blue?utm_term=.qwZEEk7wM2#.jrPpp4lo9a 23 24 25 Hunt, R. W. G., & Pointer, M. R. (2011). Colour order systems. In Measuring Colour (Fourth edi, pp. 26 27 155–195). John Wiley & Sons, Ltd. 28 29 30 IBM Corp. (2013). IBM SPSS statistics for Macintosh. Armonk, NY: IBM Corp. 31 32 33 Ishihara, S. (1993). Album-test pour la recherche des dyschromatopsies congenitales, 38 plates edn. 34 35 [Ishihara’s Test for Colour Deficiency: 38 Plates Edition]. Tokyo: Kanehara (Original work 36 37 38 published 1917). 39 40 41 Jonauskaite, D., Mohr, C., Antonietti, J.-P., Spiers, P. M., Althaus, B., Anil, S., & Dael, N. (2016). Most 42 43 and least preferred colours differ according to object context: New insights from an 44 45 unrestricted colour range. Plos One, 11(3), 1–22. http://doi.org/10.1371/journal.pone.0152194 46 47 48 Karlsson, B. S. A., & Allwood, C. M. (2016). What is the correct answer about the dress’ colors? 49 50 Investigating the relation between optimism, previous experience, and answerability. Frontiers 51 52 53 in Psychology, 7, 1808. http://doi.org/10.3389/FPSYG.2016.01808 54 55 Lafer-Sousa, R., & Conway, B. R. (2017). # TheDress: Categorical perception of an ambiguous color 56 57 58 image. Journal of Vision, 17(12), 1–30. http://doi.org/10.1167/17.12.25 59 60 61 62 63 64 - 31 - 65 #The Dress: decontextualized 32 Lafer-Sousa, R., Hermann, K. L., & Conway, B. R. (2015). Striking individual differences in color 1 2 perception uncovered by “the dress” photograph. Current Biology, 25(13), R545–R546. 3 4 5 http://doi.org/10.1016/j.cub.2015.04.053 6 7 Lotto, R. B., & Purves, D. (1999). The effects of color on brightness. Nature Neuroscience, 2(11), 8 9 10 1010–1014. http://doi.org/10.1038/14808 11 12 13 Mahroo, O. A., Williams, K. M., Hossain, I. T., Yonova-Doing, E., Kozareva, D., Yusuf, A., … Hammond, 14 15 C. J. (2017). Do twins share the same dress code? Quantifying relative genetic and 16 17 environmental contributions to subjective perceptions of “‘ the dress ’” in a classical twin 18 19 20 study. Journal of Vision, 17(1), 1–7. http://doi.org/10.1167/17.1.29 21 22 23 Mayr, S., Buchner, A., Erdfelder, E., & Faul, F. (2007). A short tutorial of GPower. Tutorials in 24 25 Quantitative Methods for Psychology, 3(2), 51–59. http://doi.org/10.1037/0096-1523.32.4.932 26 27 28 McManus, I. C., Freegard, M., Moore, J., & Rawles, R. (2010). Science in the making: Right hand, left 29 30 hand. II: The duck–rabbit figure. Laterality: Asymmetries of Body, Brain and Cognition, 15(1–2), 31 32 33 166–185. http://doi.org/10.1080/13576500802564266 34 35 Melgosa, M., Gomez-Robledo, L., Isabel Suero, M., & Fairchild, M. D. (2015). What can we learn from 36 37 38 a dress with ambiguous colors? Color Research and Application, 40(5), 525–529. 39 40 http://doi.org/10.1002/col.21966 41 42 43 Moccia, M., Lavorgna, L., Lanzillo, R., Brescia Morra, V., Tedeschi, G., & Bonavita, S. (2016). The 44 45 Dress: Transforming a web viral event into a scientific survey. Multiple Sclerosis and Related 46 47 48 Disorders, 7, 41–46. http://doi.org/10.1016/j.msard.2016.03.001 49 50 Parraga, C. A., & Akbarinia, A. (2016). NICE: A computational solution to close the gap from colour 51 52 53 perception to colour categorization. PLoS ONE, 11(3). 54 55 http://doi.org/10.1371/journal.pone.0149538 56 57 58 R Core Team. (2018). R: A language and environment for statistical computing. R Foundation for 59 60 Statistical Computing. Vienna, Austria. Retrieved from https://www.r-project.org/ 61 62 63 64 - 32 - 65 #The Dress: decontextualized 33 Robson, T. (1999). Topics in computerized visual-stimulus generation. In R. H. S. Carpenter & J. G. 1 2 Robson (Eds.), Vision Research: A Practical Guide to Laboratory Methods (pp. 81–105). Oxford: 3 4 5 Oxford University Press. 6 7 Schlaffke, L., Golisch, A., Haag, L. M., Lenz, M., Heba, S., Lissek, S., … Tegenthoff, M. (2015). The 8 9 10 brain’s dress code: How the dress allows to decode the neuronal pathway of an optical illusion. 11 12 Cortex, 73, 271–275. http://doi.org/10.1016/j.cortex.2015.08.017 13 14 15 Toscani, M., Gegenfurtner, K. R., & Doerschner, K. (2017). Differences in illumination estimation in 16 17 #thedress. Journal of Vision, 17(1), 22. http://doi.org/10.1167/17.1.22 18 19 20 Vemuri, K., Bisla, K., Mulpuru, S., & Varadarajan, S. (2016). Does normal pupil diameter differences in 21 22 23 population underlie the color selection of the #dress? Journal of the Optical Society of America 24 25 A, 33(3), 137–142. http://doi.org/10.1364/JOSAA.33.00A137 26 27 28 Wallisch, P. (2017). Illumination assumptions account for individual differences in the perceptual 29 30 interpretation of a profoundly ambiguous stimulus in the color domain: “The dress.” Journal of 31 32 33 Vision, 17(4), 5. http://doi.org/10.1167/17.4.5 34 35 Winkler, A. D., Spillmann, L., Werner, J. S., & Webster, M. A. (2015). Asymmetries in blue-yellow 36 37 38 color perception and in the color of “the dress.” Current Biology, 25(13), R547–R548. 39 40 http://doi.org/10.1016/j.cub.2015.05.004 41 42 43 Witzel, C., Racey, C., & O’Regan, J. K. (2017). The most reasonable explanation of “‘the dress’”: 44 45 Implicit assumptions about illumination. Journal of Vision, 17(2), 1–19. 46 47 48 http://doi.org/10.1167/17.2.1 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 - 33 - 65 15 16 17 18 #The Dress: decontextualized 34 19 20 21 22 Tables 23 24 25 Viewer type Number of Seen previously – Changed colour – Colour seen previously (% of unstable viewers) 26 27 28 participants non-naïve (% from unstable (% from 29 30 (% from total) the number in the those who saw 31 32 category) previously) 33 34 35 B&B W&G B&Br Other 36 37 B&B: Blue & Black 112 (47.3%) 99 (88.4%) 44 (44.4%) 2 (4.5 %) 40 (90.9 %) 2 (4.5 %) 0 (0 %) 38 39 40 W&G: White & Gold 90 (38.0%) 82 (91.1%) 13 (15.9%) 7 (53.8 %) 1 (7.7 %) 3 (23.1 %) 2 (15.4 %) 41 42 B&Br: Blue & Brown 16 (6.8%) 12 (75.0%) 6 (50%) 1 (16.7 %) 5 (83.3 %) 0 (0 %) 0 (0 %) 43 44 Other or missing data 19 (8.0%) 17 (89.5%) 9 (52.9%) 1 (11.1 %) 5 (55.5 %) 2 (22.2 %) 1 (11.1 %) 45 46 47 Total number 237 (100%) 210 (88.6%) 72 (34.3%) 11 (15.3%) 51 (70.8%) 7 (9.7%) 3 (4.2%) 48 49 Table 1. The descriptive data of the participants in relation to viewer type in Study One. 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 - 34 - 65 #The Dress: decontextualized 35 1 2 3 4 Predictors of the viewer 2 statistic 5 6 type (B&B or W&G) 7 8 Gender 1.51 9 10 11 Age 0.54 12 13 Ethnicity 3.68 14 15 Eye colour 2.31 16 17 18 Corrected vision 0.01 19 20 Sleep deprivation 2.29 21 22 23 Caffeine 0.48 24 25 Type of light 0.19 26 27 Light source 1.87 28 29 30 Table 2. Predictors of the ambiguous Dress viewer type (B&B or W&G) in the Study One included 31 32 in the logistic regression. None of the predictors were significant. 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 - 35 - 65 #The Dress: decontextualized 36 1 2 3 4 Viewer type Number of participants Seen previously (% Changed colour (% 5 6 (% from total) from the number in from those who saw 7 8 the category) previously) 9 10 11 Blue & Black 12 (21.82%) 12 (100%) 3 (25.00%) 12 13 White & Gold 17 (30.91%) 15 (88.24%) 5 (33.33%) 14 15 Blue & Brown 23 (41.82%) 17 (73.91%) 5 (29.41%) 16 17 18 Other 3 (5.45%) 2 (66.66%) 2 (100%) 19 20 Total number 55 (100%) 47 (85.45%) 16 (34.04%) 21 22 23 Table 3. The descriptive data of the participants in relation to viewer type in Study Two. 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 - 36 - 65 #The Dress: decontextualized 37 Colour Dress Viewer Mean SD Reference Statistic 1 2 parameter part type value (significance) 3 4 5 Lightness Light B&B 59.46 7.59 53.6 2.68 6 7 B&Br 73.83 11.59 53.6 8.37*** 8 9 10 W&G 85.91 11.98 53.6 11.12*** 11 12 Dark B&B 25.47 16.10 37.8 -2.65 13 14 B&Br 60.40 11.24 37.8 11.77*** 15 16 17 W&G 59.52 9.20 37.8 7.97*** 18 19 Chroma Light B&B 34.58 15.21 17.9 3.80* 20 21 B&Br 23.20 6.96 17.9 3.67** 22 23 24 W&G 12.10 9.34 17.9 -2.56 25 26 Dark B&B 13.04 15.68 9.2 .85 27 28 29 B&Br 38.32 17.97 9.2 7.77*** 30 31 W&G 46.83 10.16 9.2 15.28*** 32 33 Hue Light B&B 270.20 0.18 279.1 0.67 34 35 36 B&Br 269.34 0.13 279.1 1.53 37 38 W&G 256.68 0.54 279.1 0.54 39 40 Dark B&B 16.50 1.33 36.7 0.11 41 42 43 B&Br 79.40 0.12 36.7 34.20*** 44 45 W&G 81.10 0.08 36.7 80.11*** 46 47 48 Table 4. One-sample t-tests to compare the deviation of the selected colours from the colours 49 50 displayed by the monitor (i.e., reference value). * p < 0.050, ** p < 0.010, *** p < 0.001 (Bonferroni 51 52 correction applied separately for lightness, chroma, and hue). For hue variable, circular statistics 53 54 55 were implemented to calculate mean, standard deviation (SD) and deviation from displayed colour 56 57 values. 58 59 60 61 62 63 64 - 37 - 65 #The Dress: decontextualized 38 Figure captions 1 2 3 4 Figure 1. The ambiguous Dress picture related stimuli tested in Study Two. Context conditions were 5 6 always presented in the same order: patch (1-16), strip (17-18), and Dress (19). Left panel presents 7 8 16 patches cut out from the ambiguous Dress picture (D = dark part; L = light part) or control lure 9 10 11 stimuli (C = control). Middle panel presents two strips cut out of the left side (17) and the right side 12 13 (18) of the ambiguous Dress picture. The right panel present the full ambiguous Dress picture (19). In 14 15 the strip and the Dress conditions (17-19), participants matched a predominant colour of dark and 16 17 18 light parts (D&L). 19 20 21 Figure 2. Colour matches of dark and light parts of the ambiguous Dress picture displayed by 22 23 24 viewer types and context. Please note that the displayed colours are only approximate 25 26 representations of the participant colour matches. 27 28 29 30 Figure 3: Colour matches for control patches. Control selections are displayed to exemplify a small 31 32 individual variation in colour selections of stimuli that are other than the ambiguous Dress picture. 33 34 Control patches were cut out from the specified locations of the yellow-green salamander and the 35 36 37 blue frog. The black-and-blue dress was presented as a whole. For all images, participants were 38 39 instructed to match colours of the most predominant colour(s) of the image. Please note that the 40 41 42 displayed colours are only approximate representations of the participant colour matches. 43 44 45 Figure 4: Circular histogram of hue matches. Hues were matched to the dark (top) and the light 46 47 48 (bottom) parts of the ambiguous Dress picture, displayed by context condition (patch, strip, and the 49 50 ambiguous Dress image). The histogram bins are 15° each (radial light grey separators). Each bin is 51 52 divided into three parts to encompass selections of each viewer type. Viewer types are colour coded 53 54 55 (B&B = darkest, B&Br = medium, W&G = lightest). The y axis was square root transformed to make 56 57 the area of each bar correspond to the number of hue choices. 58 59 60 61 62 63 64 - 38 - 65 #The Dress: decontextualized 39 Figure 5. Mean lightness (A) and chroma (B) levels of the matched colours. Light and dark colours 1 2 were matched by Blue and Black (B&B), Blue and Brown (B&Br) and White and Gold (W&G) viewers 3 4 5 in the three context conditions, displayed in fixed order: patch, strip, or the ambiguous Dress image. 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 - 39 - 65 Figure 1 Figure 3