The Many Colours of 'The Dress'

Current Biology Magazine

Correspondence dress, but a photograph in which Different viewing sizes certainly add the automatic white balance setting to the variability [2]. However, even The many colours of of the camera did not match the when viewing the image on the same true illumination of the scene. Once device, from the same distance ‘the dress’ the image was taken, the colours at the same angle, differences that would be perceived by most emerge. These must be due to the Karl R. Gegenfurtner1, Marina Bloj1,2, observers when viewing the dress visual system of different observers and Matteo Toscani1 in real life (blue and black) are no performing different computations. longer perceptually available to the What are these differences then, and There has been an intense discussion majority of observers. Different people how might they arise? among the public about the colour see different colours when viewing In the first few days following the of a dress, shown in a picture posted the photograph; and that opens posting, we made measurements originally on Tumblr (http://swiked. many interesting questions that have in our lab of the colour percepts tumblr.com/post/112073818575/ fascinated the public and scientists of 15 observers. The observers guys-please-help-me-is-this-dress- alike. viewed the image of the dress on a white-and; accessed on 10:56 am Clearly, physical factors play a well-calibrated colour display under GMT on Tue 24 Mar 2015). Some role. When viewing the image on LCD controlled lighting conditions. They people argue that they see a white screens at different viewing angles, had the task of adjusting the colour of dress with golden lace, while others vastly different colours emerge. a disc, displayed on the same screen, describe the dress as blue with black lace. Here we show that the question A 0.4 B 0.5 C “what colour is the dress?” has more than two answers. In fact, there is a 0.3 0.4 continuum of colour percepts across 0.3 different observers. We measured 0.2 colour matches on a calibrated screen 0.2 for two groups of observers who 0.1 0.1 had reported different percepts of the dress. Surprisingly, differences 0 0 between the two groups arose mainly -0.1 from differences in lightness, rather -0.1 DKL luminance DKL blue-yellow than chromaticity of the colours they -0.2 adjusted to match the dress. We -0.2 speculate that the ambiguity arises -0.3 -0.3 in the case of this particular image -0.4 because the distribution of colours -0.4 -0.5 within the dress closely matches -0.2 0 0.2 -0.2 00.2 the distribution of natural daylights. DKL red-green DKL blue-yellow This makes it more difficult to D E F G disambiguate illumination changes from those in reflectance. As Newton remarked, colour is not a property of an object. It arises when a surface is illuminated and light is reflected into the eye of an observer, who interprets the light distribution of the whole scene and assigns a colour to the object. Remarkably, humans and animals are very good at assigning constant colours to objects, even though the retinal light stimulus is the ever-changing product Figure 1. Color matches and manipulated images. of illumination and reflectance. A Distribution of coloured pixels within the dress in (A) chromaticity and (B) luminance, together simple adaptation mechanism can with the colour matches of 15 observers. Dress matches (triangles) are shown in blue and lace explain this colour constancy to a matches in black for the observers perceiving the dress as ‘blue and black’, and in white and large extent, but there are numerous orange for the observers perceiving the dress as ‘white and gold’. Red dots illustrate the variabil- other factors at work [1]. How can ity of grey point matches, data replotted from [7]. The red curve indicates the daylight locus [4]. constancy then fail so badly in the (C) Photograph of the Munsell selections for 14 of the same observers in the order of their luminance screen matches. The bottom five rows show selections of observers who reported seeing case of this dress? the dress as blue and black. (D) Original image. (E) All colour pixels within the dress were rotated Constancy fails in the first place 90 degrees in colour space, and (F) by 180 degrees. (G) Greyscale image. Further details available because the stimulus is not the real in Supplemental Information. Dress image reproduced with permission from Cecilia Bleasdale.

so that it would match for them the We are left with the open question has shown impressively that our colour of the dress [3]. We also had how different people arrive at different perception of the world is not just a them match the colour of the lace conclusions when interpreting the result of physical properties recorded parts of the dress. In a separate task, same sensory data. The distribution by our senses. Rather, we make we asked them to select from the of dress pixels along the daylight assumptions about the world that Munsell Glossy collection the chip locus might be coincidental, but there guide the interpretation of sensory that best matched their recollection is some evidence that this would data, and these assumptions can be of the dress and lace colours (see make it much harder for the observers quite different for different individuals. Supplemental Information). to disentangle illumination colour from Displaying the participants’ matches object reflectance [7,8]. The bright SUPPLEMENTAL INFORMATION in colour space, we can infer that blue tones present in the image could there is a continuous distribution of equally well be due to a dark bluish Supplemental Information includes experi- colour percepts, rather than a bimodal illumination on a white dress, or to a mental procedures, one figure and one table one, which might have been expected blue dress under a neutral bright light. and can be found with this article online at from the two labels colloquially used Indeed, we have shown in a recent http://dx.doi.org/10.1016/j.cub.2015.04.043. (‘white and gold’ versus ‘blue and study [7] that observers differ mainly black’). Secondly, the dress matches along this direction when they have for the two different groups of to adjust the colour of a surface to ACKNOWLEDGEMENTS observers overlap to a large degree appear neutral grey (Figure 1A). Under in Figure 1A, where only chromaticity conditions of high uncertainty, as This work was supported by the Deutsche Forschungsgemeinschaft SFB TRR 135 and is considered (t13 = 0.06, n.s.). They found in the photograph, observers separate well when the luminance of may differ quite substantially in by BMBF-NSF grant 01GQ1304. We are the matches is taken into account in their assumptions about the colour grateful to Katja Dörschner, Roland Fleming, Figure 1B (t = 4.56, p < 0.001). This temperature and intensity of the Christoph Witzel, Maria Olkkonen and 13 Matteo Valsecchi for helpful comments and is also borne out by their choices light source. This in turn affects their suggestions. of Munsell chips (Figure 1C) that perception of the surface colours largely overlap in Chroma and Hue within the scene. for both groups, but differ in Value If the particular colour direction REFERENCES (Supplemental Information). Thirdly, is indeed of importance, then the the distributions of colours within both uncertainty should vanish if different 1. Hurlbert, A. (2007). Colour constancy. Curr. Biol. 17, R906–R907. the dress and the lace fall near the colours are chosen for the dress, as 2. Poirson, A.B., and Wandell, B.A. (1993). same line through the origin of colour for example in Figure 1D–G. When Appearance of colored patterns: pattern— color separability. J. Opt. Soc. Am. A Opt. space (Figure 1A). This line is very viewing the dress with the rotated Image Sci. Vis. 10, 2458–2470. close to the daylight locus, the set of colour distribution (E), none of our 3. Toscani, M., Valsecchi, M., and all illuminant colours from yellow to observers kept naming the dress Gegenfurtner, K.R. (2013). Optimal sampling of visual information for lightness judgments. blue that occur during the course of a ‘white’. It was seen as ‘pink’ or ‘red’, Proc. Natl. Acad. Sci. USA 110, 11163–11168. day [4,5]. presumably because there is no 4. Wyszecki, G., and Stiles, W.S. (1982). Color Science, second edition (New York: Wiley). We can conclude from these results uncertainty anymore about reflectance 5. Granzier, J.J., and Valsecchi, M. (2014). that different observers indeed perceive and illumination. This is also the case Variations in daylight as a contextual cue for different colours when looking at the when the colours in the image are estimating season, time of day, and weather conditions. J. Vis. 14, 22. picture of the dress. However, the rotated by 180 degrees. In this case, 6. Olkkonen, M., Witzel, C., Hansen, T., and differences do not arise with respect the chromaticities still fall on the Gegenfurtner, K.R. (2010). Categorical color constancy for real surfaces. J. Vis. 10, 1–16. to hue or saturation, but are mainly daylight locus, but the luminances 7. Witzel, C., Valkova, H., Hansen, T., and due to the perceived differences in no longer correspond to the natural Gegenfurtner, K.R. (2011). Object knowledge lightness. The question should thus variations of sunlight. During the modulates colour appearance. i-Perception 2, 13. 8. Pearce, B., Crichton, S., Mackiewicz, M., not be whether the dress is blue or course of a day, more yellowish Finlayson, G.D., and Hurlbert, A. (2014). white, but whether it is light blue or sunlight goes along with lower Chromatic illumination discrimination ability reveals that human colour constancy is dark blue. Despite the continuous intensities [4,5]. Asymmetries between optimised for blue daylight illuminations. PLoS choice of matching colours, observers bluish and yellowish illuminations One 9, e87989. are consistent in calling the dress have been reported before [8,9]. Thus, 9. Delahunt, P.B., and Brainard, D.H. (2004). Does human color constancy incorporate the ‘white’ when their match lies above it seems that observers do use this statistical regularity of natural daylight? J. Vis. a certain luminance, and ‘blue’ when correlation to disentangle illumination 4, 1. it lies below. We can thus exclude and surface reflectance. Interestingly, the possibility that observers would most of the variation is also lost in the simply differ in their colour naming grayscale image to the right, where 1Department of Psychology, Giessen conventions and use different labels all our observers name the dress as University 35394 Giessen, Germany. 2 for identical percepts. This finding is in ‘light grey’ or ‘silver’, but not ‘white’. Bradford School of Optometry and Vision Science, University of Bradford, BD7 1DP, agreement with previous colour naming It seems to be the covariation of UK. studies where remarkably high levels luminance and colour that is required E-mail: [email protected], of consistency were observed between to elicit ambiguity about the dress. [email protected], and within participants [6]. The popular image of this dress [email protected]