Journal of Vision (2017) 17(12):25, 1–30 1 #TheDress: Categorical perception of an ambiguous color image Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Rosa Lafer-Sousa Cambridge, MA, USA Laboratory of Sensorimotor Research, National Eye Institute, and National Institute of Mental Bevil R. Conway Health, National Institutes of Health, Bethesda, MD, USA We present a full analysis of data from our preliminary surfaces or objects. Despite being underdetermined, report (Lafer-Sousa, Hermann, & Conway, 2015) and test most retinal images are resolved unequivocally. It is not whether #TheDress image is multistable. A multistable known how the brain resolves such ambiguity, yet this image must give rise to more than one mutually process is fundamental to normal brain function exclusive percept, typically within single individuals. (Brainard et al., 2006; Conway, 2016). Multistable Clustering algorithms of color-matching data showed images are useful tools for investigating the underlying that the dress was seen categorically, as white/gold (W/ neural mechanisms. The two defining properties of G) or blue/black (B/K), with a blue/brown transition multistable stimuli are that they give rise to more than state. Multinomial regression predicted categorical one plausible, stable, percept within single individuals labels. Consistent with our prior hypothesis, W/G and that the alternative percepts are mutually exclusive observers inferred a cool illuminant, whereas B/K (Leopold & Logothetis, 1999; Long & Toppino, 2004; observers inferred a warm illuminant; moreover, Schwartz, Grimault, Hupe, Moore, & Pressnitzer, 2012; subjects could use skin color alone to infer the Scocchia, Valsecchi, & Triesch, 2014). Multistable illuminant. The data provide some, albeit weak, support images are similar to binocular rivalrous stimuli, for our hypothesis that day larks see the dress as W/G and night owls see it as B/K. About half of observers who although in binocular rivalry the competition is were previously familiar with the image reported between two different images rather than alternative switching categories at least once. Switching probability interpretations of a single image. Although the first increased with professional art experience. Priming with account of binocular rivalry involved color (Dutour, an image that disambiguated the dress as B/K biased 1760), to date there are no striking examples of reports toward B/K (priming with W/G had negligible multistable color images. Of course, not all colored impact); furthermore, knowledge of the dress’s true stimuli are unambiguous: Consider turquoise, which colors and any prior exposure to the image shifted the might be called blue or green by different people. Such population toward B/K. These results show that some ambiguous color stimuli typically retain their ambigu- people have switched their perception of the dress. ity even when labeled categorically, unlike multistable Finally, consistent with a role of attention and local shape images (Klink, van Wezel, & van Ee, 2012). To image statistics in determining how multistable images date, the best example of something approximating a are seen, we found that observers tended to discount as multistable color phenomenon is the colored Mach achromatic the dress component that they did not card, in which the color of a bicolored card folded attend to: B/K reporters focused on a blue region, along the color interface and viewed monocularly can whereas W/G reporters focused on a golden region. vary depending on whether one perceives the card receding or protruding (Bloj, Kersten, & Hurlbert, 1999). But it is not clear that the color perceptions of the Mach card are categorical. Moreover, the phe- Introduction nomenon is primarily an illusion of 3-D geometry: Without stereopsis, the perspective cues are ambiguous; Most visual stimuli are underdetermined: A given the way the colors are perceived is contingent on how pattern of light can be evidence for many different these cues are resolved. Citation: Lafer-Sousa, R., & Conway, B. R. (2017). #TheDress: Categorical perception of an ambiguous color image. Journal of Vision, 17(12):25, 1–30, doi:10.1167/17.12.25. doi: 10.1167/17.12.25 Received April 13, 2017; published October 31, 2017 ISSN 1534-7362 Copyright 2017 The Authors Downloaded from jov.arvojournals.orgThis work ison licensed 10/02/2021 under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. Journal of Vision (2017) 17(12):25, 1–30 Lafer-Sousa & Conway 2 Could #TheDress be an elusive multistable color images. Here we determined the extent to which the image? Initial reports on social media raised the individual differences in perception of the dress image possibility that the image was seen in one of two are fixed. We characterized the conditions that promote mutually exclusive ways, as white and gold (W/G) or perceptual reversals of the dress, and tested five factors blue and black (B/K). But color-matching data (not known to influence how multistable images are color names) reported by Gegenfurtner, Bloj, and perceived: prior knowledge about the image (Rock & Toscani (2015) concluded that there were many Mitchener, 1992); exposure to disambiguated versions different ways in which the dress’s colors could be seen. (Fisher, 1967; Long & Toppino, 2004); low-level The tentative conclusion was that reports of two stimulus properties (e.g., stimulus size; Chastain & categories arose as an artifact of the two-alternative Burnham, 1975); where subjects look (or attend; Ellis & forced-choice question posed by social media (‘‘Do you Stark, 1978; Kawabata & Mori, 1992; Kawabata, see the dress as W/G or B/K?’’). The implication was Yamagami, & Noaki, 1978); and priors encoded in that the true population distribution is unimodal, genes or through lifetime experience (Scocchia et al., which is inconsistent with the idea that the image is 2014). These experiments were afforded because we multistable. The Gegenfurtner et al. study measured tested people who varied in terms of both prior perceptions of 15 people. It is not known how many exposure to the image and knowledge about the color subjects would be required to reject the hypothesis that of the dress in the real world. the population distribution is unimodal. We addressed By examining the factors that influence perception of these issues through a full, quantitative analysis of the the dress image, we hoped to shed light on how the results that we presented in preliminary form shortly brain resolves underdetermined chromatic signals. after the image was discovered, in which we argued that While low-level sensory mechanisms like adaptation in the dress was seen categorically (Lafer-Sousa, Her- the retina can account for color constancy under simple mann, & Conway, 2015). A side goal was to evaluate viewing conditions (Chichilnisky & Wandell, 1995; the extent to which tests conducted online replicate D’Zmura & Lennie, 1986; Foster & Nascimento, 1994; results obtained under laboratory conditions. Many Land, 1986; Stiles, 1959; von Kries, 1878; Webster & studies of perception and cognition are being con- Mollon, 1995), they fail to explain constancy of natural ducted through online surveys; it remains unclear surfaces (Brainard & Wandell, 1986; Webster & whether results obtained in a lab and online are Mollon, 1997) and real scenes (Hedrich, Bloj, & comparable. Ruppertsberg, 2009; Khang & Zaidi, 2002; Kraft & Popular accounts suggest that people are fixed by Brainard, 1999). We have argued that the competing ‘‘one-shot learning’’ in the way they see the dress image (Drissi Daoudi, Doerig, Parkosadze, Kunchulia, & percepts of the dress are the result of ambiguous Herzog, 2017). These observations have been taken to lighting information. The colors of the pixels viewed in imply that the dress is not like a typical multistable isolation align with the colors associated with daylight image, because it is widely thought that most people (Brainard & Hurlbert, 2015; Conway, 2015; Lafer- experience frequent perceptual reversals of multistable Sousa et al., 2015). The visual system must contend images. But frequent reversals might not be a necessary with two plausible interpretations—that the dress is property of multistability (see Discussion). The per- either in cool shadow or in warm light. In our prior ception of multi-stable shape images at any given report, we tested the idea that illumination assumptions instant was initially thought to depend only on low- underlie the individual differences in color perception level factors, such as where in the image one looked of the dress, by digitally embedding the dress in scenes (Long & Toppino, 2004): since we move our eyes containing unambiguous illumination cues to either frequently, it was assumed that the perception of a warm or cool illumination. Most observers conformed multi-stable shape image would necessarily reverse to a single categorical percept consistent with the frequently. It is now recognized that high-level factors, illumination cued (Lafer-Sousa et al., 2015). Here we including familiarity with the image, prior knowledge, directly tested the hypothesis by analyzing subjects’ personality, mood, attention, decision making, and judgments about the light shining on the dress. We also learning, also play a role in how multi-stable shape tested our hypothesis that the way the dress is seen can images are seen (Kosegarten & Kose, 2014; Leopold & be explained by one’s chronotype: Night