Ethics Statement

All participants gave informed written consent. All experiments were approved by the local ethics committee at University College London.

Participants were seated in front of a computer screen in portrait orientation (Dell model 2007 WFPb, measuring 43.5 cm vertical, 27.5 cm horizontal) which displayed only a small central dot. The position of the dot on the screen was randomised across trials. Participants were instructed to imagine their own face projected frontally, life-size on the screen, with the tip of the nose located at the dot. They used a mouse to indicate the locations corresponding to 11 landmark facial features The figure reproduced as Figure 1A was shown to participants before the experiment to indicate the exact anatomical landmarks intended. Before each trial, a text label (e.g., “botton of chin”, “centre of left eye”) briefly appeared centrally on the screen. Environmental lighting was controlled so that they could not see any reflection of their face on the screen. Each landmark was judged five times in a random order. To quantify errors in perceived position of facial features, responses were later compared to the actual locations of those landmarks, obtained by taking a photograph under standardized conditions and rendering it at life-size on the same screen. The average horizontal (x) and vertical (y) error for attempts to locate each facial landmark were calculated.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Biases in face representation.

A Schematic of feature locations used to instruct participants. B. Actual and mean represented locations. C, Average of 50 female faces reproduced with permission from www.perceptionlab.com. Blue arrows indicate mean judgement error for each feature. D. Average female face adjusted according to the mean represented locations of our participants. E, F: as for C, D with average of 50 male faces.

https://doi.org/10.1371/journal.pone.0076805.g001

Fifty participants (24 female, average age 25 years) took part. The x data from left-sided landmarks (ears, nose and mouth edges, eyes) was reflected in the midline, and averaged with the corresponding right-sided landmark. This imposed an assumption of facial symmetry, but reduced the number of dependent variables and avoided possible confusion regarding the terms left and right in the context of the task. By analysing the pattern of errors, we aimed to investigate the internal stored representation of one’s own face.

Finally, a subset of 10 participants were asked to attend for a second session, in which the screen was rotated to landscape mode.

Configural processing

Models of face perception distinguish between information about individual facial features, and ‘holistic’ or ‘configural’ information about spacing between features [19]. Psychophysical studies, for example using the composite face effect, confirm that configural information plays an important role in face perception [20,21], and that this information is processed ‘holistically’. However, the structure of the underlying Gestalt or face configuration is not known. Most previous studies have focussed on spatial relations between facial features that are either hypothesised a priori, or motivated by general processing considerations independent of face perception. These include relations between the upper and lower and left and right facial features [17]. In contrast, the Face Image Task (FIT) provides a new, hypothesis-free method for investigating how multiple features are combined in configural representations, at least for representation of one’s own face.

In particular, our factor method extracted distinct sets of features whose representations tended to covary, even though we did not impose such a pattern of variation by designing our stimuli, and even though only one feature was ever judged at a time. This grouping of features was not simply defined by proximity (e.g., the edges of the nose grouped with the chin in factor Y2, not with the eyes, despite being closer to the latter than to the former). We suggest that such feature grouping may underlie configural face processing, and could provide a useful data-driven method for identifying what structural information is actually stored in the hypothesised configural representation. Configural processing might reflect precise representation of the spatial relations of features within a group, while spatial relations between groups of features might be less precisely represented. These findings generate testable predictions for future face-recognition experiments. For example, laterally shifting ears relative to eyes should be readily detectable, due to the common high loadings of these features on factor X1. But vertically shifting ears relative to eyes should be less detectable, since these features are not strongly grouped by any important factor.

Perceptual and productive self-representation

Our results show that the structural knowledge about one’s own facial features is remarkably poor. This contrasts with numerous results in visual self-face recognition showing that self-face processing is remarkably good, and superior to processing of other faces (e.g., [7]). Our results suggest that the internal representation of the face is strongly and systematically distorted, but we have no difficulty in recognising much smaller distortions when viewing faces (Figure 1). This points to a dissociation between the processes of matching visual input to a perceptual template, and the processes of accessing structural representations directly for purposes of reproducing them. Artists often improve their face drawing skills by learning geometric rules regarding the spacing of facial features. This may be considered a transfer of training from perceptual representation to productive representation. Interestingly, this process is accompanied by strengthened representation of local featural detail in face perception, at the expense of holistic, configural processing [22,23]. Comparisons of self-face and other-face processing also suggest a dominance of local over configural information for one’s own face [24]. Our data suggest that configural information about one’s own face is also poorly represented because there are systematic biases in judgements about feature locations. Nevertheless, we found grouping of features in virtue of loading on a single factor. This suggests that some configural structure to face representation is present, albeit of limited accuracy.

In addition, our results offer a dramatic example of the asymmetry between fluent, automatic, stimulus-driven access to object representation, and the limited accessibility of such object representations to the kind of deliberate controlled processing involved in our task. Even our own face appears to be impenetrable to controlled cognition. It is well-known from memory research that recognition is superior to recall. In contrast, the everyday concept of self-awareness implies an opposite pattern. We do not need to recognise our thoughts and mental states as ours. Rather, a stable, persistent core self is held to be directly known, and to provide an origin for mental states, attitudes and actions. This account of the self has recently been questioned [25]. Our approach suggests that bodily self-knowledge is poor, even for elements such as the face, which may be important for personal identity. Therefore, if there is a stable core self underlying self-identity, knowledge about the physical structure of one’s own face does not appear to be strongly linked to it.

Specificity

It is unclear whether the distortions reported here are specific to representing one’s own face, or indeed to faces as a category. Identifying suitable objects for a control task is problematic. The quality and quantity of experience we have with other people’s faces, and with non-face objects, is entirely different from the experience of our own face. Controlling for modality, familiarity, prototypicality and other relevant factors is therefore difficult. Further, the features of non-face objects cannot match those of faces in number, salience and configuration, almost by definition. Thus, the representation of information about faces cannot easily be compared to representation of other objects. Many perceptual studies suggest a specialised brain system for face processing [26], consistent with specificity. In addition, processing of one’s own face may involve a specialised network not used, or used to a lesser extent, for processing of other faces [3]. Comparisons between perception of faces and of non-face objects generally focus on neural processes, reflecting the difficulty of comparing the content of information represented [27].

For these reasons, it remains unclear if our effects are specific to representations of one’s own face. However, the bias towards short and wide face representation recalls similar biases for hands [11] and body shape [10]. The literature on visual perception and memory for shape do not suggest similar distortions for other objects. For example, people robustly overestimate vertical visual distances compared to horizontal distances [28], whereas we found a striking 27.7% underestimation of face height with relatively unbiased representation of face width (Figure 1). A previous study reported systematic overestimates of one’s own head size [29]. However, this conclusion was based on drawing outlines rather than locating features, and more specific analyses identified primarily width overestimation rather than height overestimation [30]. Classic studies of memory for feature locations report several Gestalt-type distortions of spatial representation, but do not mention distortions of aspect ratio [31]. The extensive literature on memory representations for complex figures [32] scarcely mentions distortions of shape – yet it seems unlikely that bias and variability as striking as those we have found for face representation would simply be overlooked. Therefore, we tentatively suggest that the effects reported here may be face-specific, but more research is needed.

Alternative explanations

Could the factor structure we identified arise artefactually, from some process other than face representation? One possibility is a simple rotational error. Any head tilt in the facial photographs we used to measure judgement accuracy, or in the internal representation of the face that participants used to locate features, would produce systematic errors in judging the features of positions. The misrepresentation of face shape cannot be explained in this way because shape is invariant under rotation. However, some of the other distortions we noted could potentially be due to rotation. Tilt of the head (canting) is particularly likely [33], and is known to influence face recognition [34]. The pattern of errors would depend on the precise centre of rotation. For example, a tilt of the head around the centre of the face would cause equal and opposite X shifts in the hairline and chin. Crucially, our analyses would place these shifts in the same factor, with equal and opposite loadings, because the two shifts are perfectly correlated. In fact, we found that hairline and chin shifts were associated with orthogonal factors. Therefore errors in feature judgements do not appear to be due to face rotation.

A second alternative explanation would involve the spatial distribution of pointing errors around the fixation/anchor point. For example, regression to the mean might cause people to judge all facial features as closer to the nose-tip anchor point than their true location. On this account, errors should vary strictly geometrically with each feature’s position in the face, but we found several aspects of face representation that were feature-specific and independent of position in the face or on the screen. For example, we found that errors in localising the bridge of the nose were lower than errors in localising the edges of the mouth (Figure 1), even though both are approximately equidistant from the nose-tip anchor. Our factor analyses confirmed that individual features make distinct contributions to face representation, which are not simply explained by the feature’s location within the face. For example, factor Y2 loaded strongly on the mouth, but much less on the nose edges and chin, even though these features are all close together. Further, simple geometric features of our response method cannot readily explain the strong correlations between factors underlying vertical and horizontal errors. In a previous study of hand representation, patterns of distortion were shown to be invariant when the hand was presented rotated by 90 degrees relative to the body. This suggested the distortion arose from an allocentric representation of the hand, rather than from egocentric or screen-based responding. Such tests can rule out response-specific explanations of bodily distortions for the hand. Such a test is more challenging for face representation, because the face cannot be repositioned within egocentric space in the same way as the hand.

Limitations

Finally, we acknowledge several limitations of our study. First, the number of participants is small, though it meets standards for exploratory factor analysis based on detailed simulation studies [12]. Second, our data reduction method enforced symmetry of the face around the midline, so is insensitive to possible asymmetries in representation of lateral face structures. Fluctuating asymmetry is an important facial cue to health, genetic quality, and judgements of attractiveness [35]. Future research should examine facial symmetry systematically by testing larger groups, and by directly comparing laterally inverted (mirror) versus confrontational (photograph) representations of the face [36]. Interestingly, we nevertheless identified factors involving midline shifts, confirming that asymmetry is an important aspect of face representation. Third, we have tested location judgement relative to just one central anchor, the tip of the nose. Using another anchor might, in principle, give different results – although tests of body image were largely unaffected by moving the anchor from the head to the feet [10]. Fourth, we tested only the representation of one’s own face, so we cannot say whether comparable distortions exist for less familiar faces of others, or for faces as a general semantic category. Fifth and finally, we have used factor analysis to identify the general structure of face representations from individual participants’ errors. However, we could not investigate how differences between individuals may influence their face representation, due to limited sample size. In particular, an individual’s face representation might depend on their actual facial structure, on their gender, or on cultural factors such as a desire to play down unusual or “unattractive” features.