Eyewitness identification plays a crucial role in approximately 80,000 criminal cases per year in the United States (Wells et al., 1998). However, research suggests that eyewitness identification can be subject to substantial error (Lampinen, Neuschatz, & Cling, 2012). Field studies have found that 30 %–40 % of identifications are identifications of innocent fillers (Slater, 1994, 37.93 %; Valentine, Pickering, & Darling, 2003, 30.16 %; Wells, Steblay, & Dysart, 2011, 41 %; Wright & McDaid, 1996, 33.73 %). Data from the innocence project indicate that approximately 75 % of all DNA exoneration cases involve a mistaken identification, often from more than one witness (Innocence Project, 2012). Many eyewitness events involve viewing conditions that are impoverished (poor lighting, long distances) or occur for a brief duration. Yet very few empirical studies have directly examined face recognition as a function of these situational variables. The purpose of the present project is to examine the effects of distance on face recognition under carefully controlled conditions.

Distance and face recognition

Percipient witnesses are sometimes a great distance away when they observe criminal activities. For instance, Loftus and Harley (2005) described a case in Fairbanks Alaska in which a witness made an identification after viewing an event from 450 feet away. Basic knowledge of the human perceptual system suggests that long distances such as this will be associated with impaired recognition (Lampinen et al., 2012). At increasing distances, the size of the visual image on the retina decreases, causing face recognition to depend increasingly on lower spatial frequencies, whereas high spatial frequency details will no longer be resolvable (Loftus & Harley, 2005). Because optimal face recognition depends upon both high and low spatial frequency bands (Goffaux, Hault, Michel, Vuong, & Rossion, 2005), this implies that greater distances will be associated with progressively worse recognition.

In one of the few studies that have attempted to address the effect of distance on identification accuracy, Wagenaar and Van der Schrier (1996) provided participants with pictures of faces that were sized so as to approximate the visual angle that would be produced by an actual face viewed at various distances. Each image was displayed for 12 s, and after each image, the participant was shown a target-present or -absent lineup. Wagenaar and Van der Schrier found that there was an abrupt decline in recognition at distances over 15 m. In another study, Loftus and Harley (2005) used photographs of celebrities that had been resized or low-pass filtered to mimic the perceptual experience of viewing a face at various distances. Participants were shown images that were gradually increased in size, or gradually deblurred, until they could name the celebrity. Naming accuracy was approximately 50 % when photographs were sized to produce a visual angle that was equivalent to a face seen from 55 feet away (based on inspection of their Fig. 9). Naming accuracy zeroed out when the photographs were sized to create a visual angle that was equivalent to a face seen from 128 feet away.

Although these studies provide useful insights into the relationship between distance and face recognition, these studies did not examine recognition of actual faces at a distance. Rather, they examined recognition of pictures of faces that had been altered to approximate viewing faces at varying distances. Moreover, Wagenaar and Van der Schrier’s (1996) data were extremely noisy, due to low cell sizes, and their attempt to smooth the results by fitting it to a recognition model was not successful; that is, the model did not provide a very good fit to the data. Loftus and Harley’s (2005) results also provided useful insights, but the naming data were most relevant to familiar face recognition, not recognition of strangers. Only one study has examined recognition of actual faces of strangers at a distance. Lindsay, Semmler, Weber, Brewer, and Lindsay (2008) had participants view a live confederate at distances that were under 15 m or over 15 m. Performance was worse at the longer distance, especially for target-present lineups. However, the study does not provide fine-grained information about the relationship between distance and viewing accuracy, because although Lindsay et al. utilized a variety of different distances, they reported only whether the viewing distance was above or below 15 m.

In addition to examining the effects of distance on recognition of faces, the present article also compares the underlying mechanisms responsible for recognition of faces and how those mechanisms vary as a function of distance. Considerable evidence suggests that recognition memory is mediated by two dissociable processes known as recollection and familiarity (Jacoby, 1991). Recollection involves a conscious reexperiencing of the original event, resulting in high-confidence acceptance of targets on recognition memory tests (Yonelinas, 2001). Familiarity, on the other hand, is associated with a graded feeling of prior experience and typically leads to less definitive acceptances. In modeling work, Yonelinas (1994) demonstrated that recollection-based acceptances were associated with above zero y-intercepts on target/foil ROC curves.

Rajaram’s (1996, 1998) distinctive/fluency account posits that recollection is driven by the distinctiveness of the target, relative to background items. Arguably, because faces seen close up have greater high spatial frequency information, they should lead to more distinctive representations than faces seen from greater distances. This suggests that faces seen up close should lead not only to greater recognition, but also, specifically, greater recollection. We address this issue by examining recognition memory ROC curves for faces seen at different distances.

The above suggests quite strongly that a detailed examination of the effects of distance on face recognition, under free-field viewing conditions, is sorely needed and would be informative.

Method

Participants

One hundred ninety-five university students participated (61 % female). The average age of participants was 19.7 years (SD = 3.80 years). The majority of participants were Caucasian (86.67 %), with the remaining number split between African-American (4.1 %), Asian (5.1 %), Hispanic (2.0 %), Native American (1.0 %), and other races/ethnicities (1.0 %). A power analysis revealed that, assuming Cohen’s f 2 of .06, that a sample of at least 130 would be needed to obtain a desired power of .80. Thus, our sample size was adequate to detect an effect of the size we anticipated.

Materials/procedure

Participants viewed eight target individuals. For each participant, all eight targets were displayed at one of six outdoor distances (5, 10, 15, 20, 30, or 40 yards) during daylight hours. Thirty-two participants were randomly assigned to all of the conditions except the 5-yard condition, which had 33 participants, and the 30-yard condition, which had 34 participants. Participants were led to a marked area within a flat and evenly lit stretch of sidewalk, free from excessive pedestrian traffic. Participants did not know their actual distance from the viewing area, nor was there any numbered marking within the area to betray such information. The participant was instructed to turn their back to the viewing area, during which time the first target individual would assume their place within the viewing area. Once the target individual was in place, the participant was instructed to turn and view the target for 10 s, measured via stopwatch by an experimenter. After 10 s elapsed, the participant was instructed to face away from the viewing area, at which point the target would leave, to be replaced by a second target. This cycle was repeated until the participant viewed all eight targets. The order in which the targets were presented was randomly determined for each participant. The co-experimenter then led the participant to an area outside of the experiment viewing area, at which point the experimenter administered a survey to the participant. Participants then saw 16 photographs presented one at a time in a random order. Eight photographs were of previously viewed targets, and 8 photographs depicted foils, each of which matched a verbal description of a yoked target. The verbal description of each target was generated by the fourth author on the basis of an inspection of the target’s photographs. Foils matching those descriptions were then found in a face database maintained by the first author. The participant rated their confidence in having seen or not having seen the pictured individual on an 8-point scale (1 = absolutely certain, have not seen; 4 = just guessing, have not seen; 5 = just guessing, did see; 8 = absolutely certain, did see).

Results

For each participant, we calculated the proportion of hits and false alarms (see Fig. 1). A linear regression analysis was conducted to analyze the effect of increasing distance on the proportion of hits. The best fitting line intercepted the y-axis at 0.7538, with a slope of −0.0055, meaning that for each yard added to the distance, the hit rate decreased by 0.55 %. The linear regression model accounted for 11.63 % of variance, which was significant, F(1, 193) = 26.54, p < .0001, f 2 = .13. For false alarms, the best-fitting line intercepted at y = 0.1593 with a slope of .0044, indicating that false alarms increased 0.44 % with each added yard between target and participant. The regression accounted for 7.61 % of variance, which was significant, F(1, 193) = 17, p < .0001, f 2 = .08. Thus, as was expected, at increasing distances, the hit rate decreases and the false alarm rate increases, consistent with a mirror effect (Glanzer & Adams, 1985).

Fig. 1
figure 1

Effects of distance on proportion of hits and false alarms

Full size image

In addition to calculating hit and false alarm rates, we calculated the signal detection measure d′ (Green & Swets, 1966). A linear regression analysis indicated that as distance increased, memory strength as measured by d′ decreased (see Fig. 2). The y-intercept for d′ was 1.63, with a decrease of 0.0289 for each added yard. The regression accounted for 18.18 % of variance, which was significant, F(1, 193) = 44.09, p < .0001, f 2 = .22. A linear regression indicated that as distance increased, the response bias measure β decreased, suggesting that participants were becoming increasingly liberal in their response bias. These results are shown in Fig. 3. The regression line for β had a y-intercept at 1.34. For each additional yard added, β decreased by 0.0056. The regression accounted for 17.03 % of variance, which was significant, F(1, 193) = 40.83, p < .0001, f 2 = .21.

Fig. 2
figure 2

Effects of distance on d′

Full size image
Fig. 3
figure 3

Effects of distance on β

Full size image

We next used confidence data to derive empirical receiver operator characteristic (ROC) curves (see Macmillan & Creelman, 1991, Fig. 5). The dashed line of Fig. 4 represents the line of no discrimination. At longer distances, recognition is poor, as reflected by the flatter ROC curves. As distance decreases, the ROC curves become increasingly bowed toward the upper left corner, indicating better memory. Crucially, we fit the data at each distance to the dual-process signal detection model, using procedures outlined by Yonelinas (1994). For recognition data with an 8-point confidence scale, Yonelinas’s model has nine paramaters: recollection (R), d′, and seven response criteria (C1–C7). Recognition of a target can occur either via recollection or, if recollection fails to occur, via familiarity as outlined in signal detection theory. Recognition of foils is based entirely on familiarity.Footnote 1 Graphically, recollection is reflected by ROC fits with a y-intercept that is greater than zero. There was little evidence of recollection at any of the distances, as can be seen in Table 1.

Fig. 4
figure 4

Empirical ROC curves as a function of distance

Full size image
Table 1 Fits of data to Yonelinas’s (1994) dual-process signal detection theory
Full size table

Discussion

The present study was designed to fill a gap in the extant literature on the influence of physical distance between viewer and target on the ability to recognize unfamiliar faces. Namely, most previous work merely simulated the effect of distance photographically (Loftus & Harley, 2005; Wagenaar & Van der Schrier, 1996) or did not systematically report effects of distance on recognition at a variety of distances (Lindsay et al., 2008). Our study, on the other hand, actually manipulated distance between viewer and unfamiliar targets in an ecologically valid outdoor scenario on a college campus. We begin discussion with a reiteration of the results and outline their implication for theoretical accounts of facial perception and the applied realm of eyewitness memory.

Distance: does the “rule of 15” stand on firm ground?

We observed a steady, monotonic decrease in hits, coupled with an increase in false alarms, as the distance between witness and target grew. This, in principle, replicates previous findings (Jong, Wagenaar, Wolters, & Verstijnen, 2005; Loftus & Harley, 2005; Wagenaar & Van der Schrier, 1996), but our major contribution is that we tested the phenomenon with live actors in the field in an ecologically valid scenario. This allowed us to test Wagenaar and Van der Schrier’s “rule of 15,” where the diagnosticity ratio of an identification stays above 15 at distances less than 15 me. The authors stated that the goal of establishing the rule was to create a useful heuristic tool for forensic experts in order to assess the quality of eyewitness identifications. Although their study revealed that, generally speaking, high-quality identifications are found when the rule of 15 is met (but not always), the data do not constrict all good identifications (i.e., ones with a diagnostic ratio > 15) to this zone. This may call into question the wisdom of a “rule of 15” to begin with. Consistent with Lindsay et al.’s (2008) interpretation, the effect of distance on d′ appears to be very close to a linear function, suggesting that while performance certainly decreases after 15 m, there is not a sudden drop in performance at this point. However, it is worth noting that the so called “rule of 15” was not predicated entirely on there being a sharp drop in performance after 15 m, but also on a notion of what error rate would be acceptable to society—that is, the Blackstone rule (Pound, 1908).

A new discovery we have made is that the response bias of witnesses grew increasingly liberal as distances grew greater. This conforms to a principal components analysis (PCA) face-space account of facial recognition (Turk & Pentland, 1991; Valentine, 2001). Face-space theory posits that faces are stored at the intersections of multiple holistic featural dimensions that reflect different ways in which faces can vary. The architecture of such a space implies that, somewhere, there is an “average” face and around it lie multitudes of faces clustered such that faces similar in appearance will be close together and dissimilar faces will lie farther apart. PCA modeling of this space parameterizes the faces such that they themselves are combinations of weighted eigenfaces, which capture oblique quantities of variance within a set of face images. Eigenfaces are ordered such that the first eigenface captures the most amount of variance (external shape and general location of internal features) and the following eigenfaces capture increasingly smaller amounts of variance, such as light skin imperfections and wrinkles. Previous research has shown that adequately reconstructing a recognizable face from its constituent eigenfaces requires very few eigenfaces (Burton, Bruce, & Hancock, 1999; O’Toole, Abdi, Deffenbacher, & Valentin, 1993), although the image appears blurry and lacks fine-grained textural details. This blurriness renders the face closer in appearance to the average face in face-space and also filters out much high spatial frequency information that Loftus and Harley (2005) have found contributes to accurate recognition. Our results conformed to the notion of facial averaging because initially viewing faces from a distance acted as a filter to the fine-grained details that would normally be captured if the faces were viewed close up. This, in turn, made the faces more average-looking and, therefore, made them more similar in appearance to their verbally matched foils than they were in actuality, because the faces were closer to each other in face-space.

Rajaram (1993) argued that conscious recollection is a function of the distinctiveness of the target items. On the basis of face-space theory, we reasoned that faces would be more distinctive at short distances because of the greater prevalence of high spatial frequency information. Yet analysis of ROC curves provided very little evidence of recollection. This finding may reflect that the individuals who acted as confederates in the present experiment were fairly typical looking, so that their faces were not distinctive even at brief distances. This suggests the possibility of a distance × distinctiveness interaction in face recognition.

Conclusions and future directions

The major message to take from the present study is that we found a significant decrease in hits (.55 % per yard), coupled with a significant increase in false alarms (.44 % per yard), as the distance between participant witness and target increased. However, the present study and previous research (Wagenaar & Van der Schrier, 1996) examined the ability to recognize only unfamiliar faces in this paradigm. Another avenue of interest would be the qualitatively different realm of familiar face recognition, which has previously only been studied using photographic simulation of distance (Jong et al., 2005, Loftus & Harley, 2005). This is particularly salient because much research has shown that familiar faces can be identified with far less high spatial frequency information than unfamiliar faces (Burton et al., 1999). This fine-grained detail is precisely the sort of information that is lost as distances grow greater, and although this might preserve hit rates of familiar faces well, no research has examined how false alarms might be affected. Imagine a scenario where a witness from a distance mistakes a stranger perpetrating a crime for someone known to them. Although the witness’s ability to recognize their acquaintance at that distance might be good, their ability to discriminate between the acquaintance and a similar-looking person might be poor. This dissociation between hits and false alarms is especially salient to convictions based on mistaken identity, which has been cited as the cause of 72 % of convictions that have been since overturned due to DNA evidence (Innocence Project, 2012). However, for smaller crimes, such as burglary, robbery, nonsexual assault, or vandalism, DNA is rarely available at the crime scene and not typically appropriate to the case. Although penalties for such crimes are light, as compared with sexual crimes or murders, the future lives of those convicted could still be compromised, making this avenue of study particularly worthwhile for applied forensic purposes.

Yet another avenue of research would address the other major estimator variable that factors into the rule of 15: luminosity. Ambient lighting plays a major part in the amount of visual information that can be derived from a stimulus, and previous work has also found a steady decrease in recognition accuracy as luminosity decreases (Loftus, 1985; Wagenaar & Van der Schrier, 1996). In addition, Loftus found that, rather than acting as a perceptual filter as distance does, luminosity impacts the rate at which information can be processed. In other words, it takes just a bit longer to make sense of a dark stimulus and then to orient to it, as comparedwith a lighter stimulus. This has profound implications for eyewitness research, because many crimes occur at night under less-than-ideal lighting conditions.