Measuring Accuracy Facial Forensics Comparisons - Journal-Paper

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

rspb.royalsocietypublishing.org

Perceptual expertise in forensic facial
image comparison
David White1, P. Jonathon Phillips2, Carina A. Hahn3, Matthew Hill3
and Alice J. O’Toole3

Research
Cite this article: White D, Phillips PJ, Hahn
CA, Hill M, O’Toole AJ. 2015 Perceptual
expertise in forensic facial image comparison.
Proc. R. Soc. B 282: 20151292.
http://dx.doi.org/10.1098/rspb.2015.1292

Received: 31 May 2015
Accepted: 5 August 2015

Subject Areas:
cognition
Keywords:
visual expertise, face recognition,
person identification, biometrics,
forensic science

Author for correspondence:
David White
e-mail: [email protected]

1

School of Psychology, The University of New South Wales, Sydney, New South Wales 2052, Australia
National Institute of Standards and Technology, 100 Bureau Drive, MS 8940, Gaithersburg,
MD 20899, USA
3
The University of Texas at Dallas, Richardson, TX 75080, USA
2

Forensic facial identification examiners are required to match the identity of
faces in images that vary substantially, owing to changes in viewing conditions
and in a person’s appearance. These identifications affect the course and outcome of criminal investigations and convictions. Despite calls for research on
sources of human error in forensic examination, existing scientific knowledge
of face matching accuracy is based, almost exclusively, on people without
formal training. Here, we administered three challenging face matching tests
to a group of forensic examiners with many years’ experience of comparing
face images for law enforcement and government agencies. Examiners outperformed untrained participants and computer algorithms, thereby providing
the first evidence that these examiners are experts at this task. Notably, computationally fusing responses of multiple experts produced near-perfect
performance. Results also revealed qualitative differences between expert
and non-expert performance. First, examiners’ superiority was greatest at
longer exposure durations, suggestive of more entailed comparison in forensic
examiners. Second, experts were less impaired by image inversion than nonexpert students, contrasting with face memory studies that show larger face
inversion effects in high performers. We conclude that expertise in matching
identity across unfamiliar face images is supported by processes that differ
qualitatively from those supporting memory for individual faces.

1. Introduction
Proliferation of CCTV, mobile image capture and face recognition technology
entails a critical role for facial images in modern forensic identification. As a
result, facial image comparison is a major source of evidence in criminal investigations and trials [1,2], and wide deployment of automatic recognition systems
over recent years has been accompanied by substantial gains in reliability [3].
Importantly, forensic applications of this biometric software—as with automatic
fingerprint recognition systems—are configured to provide lists of potential
matches according to the computed scoring metric. For final identity judgements,
such as those provided as evidence in court, trained facial forensic examiners
adjudicate suspected matches [1,2,4]. Given this reliance, and evidence of
DNA-based exonerations owing to errors in forensic judgements [5], there is a
pressing need for research that can benchmark the skills of examiners relative
to untrained humans and computer-based face recognition systems [6].
There is striking evidence that untrained individuals perform poorly on the
apparently straightforward task of matching the identity of an unfamiliar face
across two different images [7– 11]. Even under optimal matching conditions
in laboratory tests conducted using images that are taken on the same day, in
the same neutral pose, and under similar environmental conditions; error
rates for untrained individuals are in the range of 20–30% [8,9]. In suboptimal
capture conditions when environmental factors are unconstrained, such as
when matching between CCTV footage and high-quality mug-shots, performance can approach chance [10]. Moreover, in field tests conducted outside of the

& 2015 The Author(s) Published by the Royal Society. All rights reserved.

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

GFMT

PICT

EFCT

2
inverted

same
identity

different
identity

laboratory, professional police officers and passport officers
make the same number of face matching errors as standard
groups of student participants [10,11], despite performing
face matching as part of their daily work.
Here, we assessed perceptual expertise in facial image comparison by developing and administering a battery of tests to
an international group of forensic facial examiners. Smallscale studies have recently reported higher identification
accuracy in court practicing facial image examiners [12,13],
but are limited by use of unstandardized tests containing
small numbers of image comparisons. Inconsistency across
experiments also raises the possibility that group performance
is specific to the workplaces tested. To ensure that experts represented the highest global standards in forensic image
comparison, we approached organizers of the Facial Identification Scientific Working Group (FISWG; www.fiswg.org).
The May 2014 meeting of this group was held in the FBI academy in Quantico, Virginia, and presented a unique
opportunity to test an international group of facial forensic
examiners with rigorous training and many years of professional experience identifying unfamiliar faces (henceforth
examiners). As comparison groups, we tested the FISWG meeting attendees who do not perform forensic facial examination
as part of their daily work, but were attending the meeting
as managers, technical experts or administrators in biometric
systems (controls). Because people in this group were knowledgeable of the types of training used for forensic examiners
and difficulties associated with unfamiliar face matching [14],
we also tested untrained university students, representing the
most commonly tested cohort in previous research (students).
In all tests, participants decided if pairs of images were of
the same identity or of two different identities. We mapped
expert performance to established human accuracy using
a standard psychometric test of unfamiliar face matching
ability—the Glasgow Face Matching Test (GFMT) [9]—and
created two new tests to benchmark forensic examiners against
both human and state-of-the-art algorithm-based matching performance. Tests were designed to be sufficiently challenging for
examiners, and representative of the types of decisions encountered in daily work (see figure 1 for example images). To create
the Expertise in Facial Comparison Test (EFCT), we selected
pairs of images for identity comparisons that were challenging
for computers and untrained humans based on data from pilot
work and previous evaluations of human and computer face
matching performance [15,16]. Image pairs in the Person Identification Challenge Test (PICT) included body cues and were

selected to have no computationally useful identity information
in the face, as indicated by the fact that leading algorithms make
100% errors on this set [17] (for details of test construction,
please refer to Methods).
The FISWG meeting was a unique opportunity to address
a key theoretical question in the study of face identification.
Decades of research have shown that, relative to other classes
of objects, face recognition is a skill for which humans are
experts. Because face recognition in the general population
is disproportionately impaired by inverted presentation
compared with other objects [18], the face inversion effect
(FIE) has been taken as an index of this expertise [18–21].
This view is bolstered by evidence that FIE (i) increases as
face processing abilities improve with development [19,20],
(ii) is weaker in people with impairments in face identification ability [22,23] and (iii) is stronger in those with
exceptionally good face processing ability [23]. However,
this evidence is almost entirely based on face memory
tasks. In facial image comparison tasks performed by forensic
examiners, it has been proposed that different mechanisms
are recruited [24] and so it is not clear whether expertise of
facial examiners will be indexed by FIEs. To test this, we
included an inverted face matching test in the EFCT.
To probe the nature of expertise in face matching further,
we manipulated exposure duration of image pairs in identity
comparisons. Previous research suggests that 2 s is optimal
for face matching decisions in untrained participants [25],
but training in forensic facial examination emphasizes careful
comparison and effortful analysis of facial images prior
to identification decisions [14]. Thus, we predicted that the
accuracy of the forensic experts would be superior to the controls and students on the longer (30 s), but not the shorter
exposure times (2 s).

2. Results
Aggregated results across tests confirmed expertise of the facial
examiners (figure 2 summarizes GFMT and PICT results and
figure 3 summarizes EFCT). The rank order of identification
performance for examiners, controls and students was stable
for all results. In all experimental conditions, across the three
tests (GFMT; EFCT: 2 s upright, 2 s inverted, 30 s upright, 30 s
inverted; and PICT), accuracy was ranked as follows:
examiners . controls . students. The consistency of this order
proved statistically reliable (examiners . controls, Wilcoxon

Proc. R. Soc. B 282: 20151292

Figure 1. Examples of image pairs from the GFMT, PICT and EFCT. Images in the GFMT were all taken on the same day and under very similar lighting conditions,
whereas images from the PICT and EFCT were captured in diverse ambient conditions and over a longer time period. For details of test construction, please refer to
Methods. (Online version in colour.)

rspb.royalsocietypublishing.org

upright

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

(b)

3
1.0

90

0.9
aROC curve

100

80
70
60

rspb.royalsocietypublishing.org

% correct

(a)

0.8
0.7
0.6
0.5

student

control

examiner

student

control

examiner

Figure 2. (a) Mean percentage correct (+1 s.e.) for experimental groups in the GFMT. (b) PICT performance is plotted as average area under the receiver operating
characteristic (ROC) curve. Mean normative accuracy in the GFMT is indicated by the dashed line.

(a)
2s

1.0

aROC curve

0.9

30 s

examiner
control
student

0.8
0.7
0.6

(b)

group × exposure duration
1.0
0.9

aROC curve

inverted

30 s
2s

upright

inverted

group × orientation
upright
inverted

0.8
0.7
0.6
0.5

student

control

examiner

student

control

examiner

upright
(c)
aROC upright – aROC inverted

0.5

face inversion effect
0.2
0.1
0

student

control

examiner

–0.1
–0.2

Figure 3. Analysis of EFCT results. (a) Mean ROC scores for three groups on the EFCT (+1 s.e.). (b) Significant interactions in EFCT data. Simple main effects
revealed stronger effects of group in 30 s compared with 2 s exposure durations (left). In addition, there was a significant interaction between group and orientation
(right). (c) Contrary to hypotheses based on FIEs in memory tasks, inversion effects are larger for students than for experts. Details of analysis are provided in
the text.

sign test, t5 ¼ 4.85, p ¼ 0.0313; controls . students, t5 ¼ 4.85,
p ¼ 0.0313), showing a general superiority of forensic examiners
across tests.

(a) Glasgow Face Matching Test
To map performance of experimental groups to the population
at large, we compared performance on the GFMT with normative data for the test [9]. Mean GFMT scores for the three
groups are shown in figure 2. Accuracy of examiners exceeded normative, control and student accuracy (t219 ¼ 6.35,
p , 0.0001, Cohen’s d ¼ 0.858; t39 ¼ 2.34, p , 0.05, Cohen’s
d ¼ 0.749; t57 ¼ 4.09, p , 0.05, Cohen’s d ¼ 1.08). Performance

of control participants also exceeded normative accuracy
(t206 ¼ 2.77, p ¼ 0.006, Cohen’s d ¼ 0.385), but performance
did not differ significantly between controls and students
(t44 ¼ 1.51, p . 0.05, Cohen’s d ¼ 0.455). Student performance
did not differ significantly from normative accuracy scores
(t224 ¼ 1.14, p . 0.05, Cohen’s d ¼ 0.1). As far as we are
aware, experts and controls are the only groups reported to
have exceeded normative accuracy on the GFMT.

(b) Person Identification Challenge Test
To analyse scores on the PICT, for each participant we computed the area under the receiver operator characteristic

Proc. R. Soc. B 282: 20151292

50

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

(c) Expertise in facial comparison test

(d) Fusion analysis
In forensic practice, to assure consistency and consensus, it is
common for multiple examiners to repeat a single comparison
judgement. To model the effectiveness of this process, we conducted simulations to ‘fuse’ or combine judgements at the level
of individual image pairs. The simulations followed previous
work showing that aggregating the judgements of multiple
participants improves identification accuracy [26,27]. We
focused on data from the 30 s upright EFCT as this experimental condition most closely resembles working practice of
forensic examiners; however, we also carried out these
simulations with the PICT and found comparable results.
Effects of aggregation were calculated separately for each
group (students, controls, examiners) by resampling participants’ identity ratings (i.e. from 1 ¼ sure same to 5 ¼ sure
different) for 84 image pairs. We randomly sampled n participants from within a group and averaged their responses for
each image pair separately. This sampling procedure was
repeated 100 times for each value of n, and accuracy was
computed at each iteration by calculating the group aROC.
Aggregate accuracy for a given sample size was measured
as the average aROC across all iterations. We report results
for aggregate sample sizes that vary from 1 to 14 participants,
with the upper limit dictated by the smallest group of participants (for controls, by definition, all iterations of sample size
14 include the entire group).
Figure 4 shows the aggregation effect as a function of participant sample size and serves as a practical guide to the
performance benefit that can be expected by combining identity judgements across participants. Closer inspection of
results with smaller n shows substantial improvements
in accuracy as additional judgements were aggregated.
Comparisons highlight differences in the relative value of
participants according to group. That is, one examiner
(0.936) is roughly equal to two controls (0.946) or four students (0.942). With only one subject, examiner performance
surpassed control performance in 62% of the 100 iterations,
and controls surpassed student performance in 79% of the
iterations. As sample size increased, however, aggregated
judgements from examiners were more likely to surpass controls and aggregated judgements by controls were more likely
to surpass student decisions. At the maximum sample size,
examiners surpassed controls in 99% of the iterations and
controls surpassed students in 91% of the iterations. The
bars in the bottom of figure 4 have this analysis for all
participant sample sizes.
So, although examiner performance did not reach ceiling
levels at 30 s exposures when computing performance
measures at the individual level, these limits were largely overcome by response aggregation—which produced near-perfect
accuracy and revealed a highly stable performance advantage
for professional examiner groups. Given the highly challenging
nature of the images used in the EFCT, this suggests that a
fusion approach can help support identification decisions in
forensic practice.

4

Proc. R. Soc. B 282: 20151292

The EFCT was designed to test three key predictions. First,
examiners would be more accurate than both other groups.
Second, this advantage would only be observed in conditions
that enabled careful examination of image pairs (i.e. 30 s
exposure). Third, owing to perceptual expertise comparing
upright facial images, the examiner group would show
larger inversion effects compared with controls and students.
We computed aROC scores individually for each participant (figure 3a). Scores were submitted to a 2  2  3
ANOVA with exposure time (2, 30 s) and orientation (upright,
inverted) as within-subject factors and group (student,
control, examiner) as a between-subjects factor. Significant
main effects were found for exposure time, orientation and
group (F1,70 ¼ 176.33, p , 0.0001, h2p ¼ 0:716; F1,70 ¼ 341.07,
p , 0.0001, h2p ¼ 0:83; F2,70 ¼ 14.54, p , 0.0001, h2p ¼ 0:293).
In line with our second prediction, there was a significant
interaction between group and exposure time (F1,70 ¼ 4.82,
p ¼ 0.0109, h2p ¼ 0:121). In addition, we observed a statistically significant interaction between group and orientation
(F1,70 ¼ 4.02, p ¼ 0.022, h2p ¼ 0:103). The three-way interaction
was non-significant (F , 1). Significant interactions are plotted
in figure 3b.
For the group and exposure time interaction, simple
main effects tests revealed group effects at both shorter and
longer exposure times (F2,70 ¼ 14.55, p , 0.0001, h2p ¼ 0:294;
F2,70 ¼ 46.66, p , 0.00001, h2p ¼ 0:571), but these effects were
more pronounced when participants had more time to examine
each image pair. In the 2 s condition, examiners performed
more accurately than students (F1,70 ¼ 27.02, p ¼ 0.00001,
h2p ¼ 0:279), but did not differ from the controls. In the 30 s condition, the examiners were more accurate than the controls
and students (F1,70 ¼ 9.28, p ¼ 0.003, h2p ¼ 0:117; F1,70 ¼ 91.83,
p , 0.0001, h2p ¼ 0:567), supporting the prediction of greater
differentiation of the examiners from the other groups, when
they had more time to examine the image pairs.
To examine the group and orientation interaction, we collapsed across study duration (figure 3b). Simple main effects
tests revealed significant effects of group for both inverted
and upright faces (F2,70 ¼ 50.38, p , 0.00001, h2p ¼ 0:59;
F2,70 ¼ 18.18, p , 0.0001, h2p ¼ 0:342). This interaction is consistent with differences in the size of the FIE across groups.
Because the FIE is an established index of perceptual expertise,
we predicted a larger inversion effect for the more accurate participant groups (i.e. examiner FIE . control FIE . student
FIE). A cursory examination of the data proved inconsistent
with that prediction. We examined this further by computing
FIE strength (aROC upright – aROC inverted) for each participant, in each exposure duration condition. A 2  3 ANOVA
with study duration (2, 30 s) as a within-subject factor and
group (student, control, examiner) as a between-subjects
factor revealed a significant main effect of group (F2,70 ¼ 4.02,
p ¼ 0.022, h2p ¼ 0:103). Means for these difference scores

appear in figure 3c and show stronger FIE for students than
for examiners (F1,70 ¼ 7.68, p ¼ 0.007, h2p ¼ 0:099) opposite to
predictions based on the perceptual expertise hypothesis. Contrasts between students and controls (F1,70 ¼ 2.53, p ¼ 0.116,
h2p ¼ 0:036), and between examiners and controls (F1,70 ¼
0.425, p ¼ 0.517, h2p ¼ 0:036), were non-significant.

rspb.royalsocietypublishing.org

(aROC). Summary aROC scores are shown in figure 2. Scores
were analysed by ANOVA with group (student, control, examiner) as a between-subjects factor. There was a significant main
effect of group (F2,70 ¼ 4.89, p ¼ 0.01, h2p ¼ 0:122). Contrast
analyses indicated that examiners performed more accurately
than students (F1,70 ¼ 9.66, p ¼ 0.003, h2p ¼ 0:121). Differences
between examiners and controls (F(1,70) ¼ 1.06, p ¼ 0.307,
h2p ¼ 0:015), and between controls and students (F1,70) ¼ 2.18,
p ¼ 0.144, h2p ¼ 0:030), were non-significant.

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

EFCT fusion (30 s)

5

rspb.royalsocietypublishing.org

1.0

0.8

Proc. R. Soc. B 282: 20151292

aROC curve

0.9

0.7
algorithm aROC = 0.687
0.6
examiner
control
student

0.5
1

2

3

4

5

6

8
9
7
no. subjects

10

11

12

13

14

% examiner advantage (versus control)
62 67 70 77 79 88 89 86 89 90 93 92 96 99
% examiner advantage (versus student)
82 82 88 91 97 94 94 98 100 98 100 99 100 100
% control advantage (versus student)
79 70 77 75 77 80 72 88 79 86 89 86 91 91

Figure 4. Accuracy according to number of judgements fused in the EFCT simulations. Accuracy increases as more judgements are fused. As indicated in the bars below the
chart, as the number of subjects fused increases, the reliability of the pattern examiners . controls . students also increases. Clear improvements for aggregating
judgements are seen within all three groups of subjects. For all groups, these improvements plateau well before the maximum sample size: at asymptote, average
aROCs for groups of roughly eight raters were close to perfect for all three groups (examiners ¼ 0.997; controls ¼ 0.987; students ¼ 0.973). (Online version in colour.)

3. Discussion
In this study, we report the first systematic assessment of face
matching performance by a diverse group of international
forensic examiners. The rank order of identification performance in all six experimental tests placed examiners over
controls and students, and examiner performance exceeded
normative levels established in previous studies [9]. Although
examiners’ performance was not statistically superior to controls in all experiments, it was always statistically superior to
the student performance. To the best of our knowledge, this is
the first convincing demonstration of a professional group
showing higher accuracy on face matching tasks.
Closer analysis also revealed two qualitative differences
between examiners and non-expert groups. First, examiners’
superiority was not specific to longer exposures but was
also observed when permitted just 2 s per comparison in
the EFCT, suggesting that expertise improved intuitive as
well as considered judgements. Differences between groups
were most pronounced with 30 s exposure, however, pointing
to a more entailed and effective identity examination process
by examiners than by less experienced participants. This contrasts with accounts of perceptual expertise in radiographers
[28] and fingerprint examiners [29], where expertise appears

primarily driven by a shift in perceptual strategy towards
fast and global image analysis [30]. Consistent with the training forensic examiners receive [14], our results suggest an
opposing trajectory of expertise in forensic facial image comparison characterized by a transition towards controlled and
effortful analysis.
Second, inverting images produced less impairment in
examiners compared with students. This is perhaps surprising,
as it is not entirely consistent with the expertise hypothesis
of face processing, which predicts increasing impairment as
perceptual expertise develops [19,20]. This finding can be
reconciled, however, with existing evidence that unfamiliar
face matching may rely on separate processes to those supporting face memory. For example, individual differences in
unfamiliar face matching accuracy are not predicted by accuracy in face memory tasks [9,24]. Our results extend this
work, suggesting that processes underlying expertise in unfamiliar face matching may be dissociable from those driving
expertise in face memory [20,31]. Because the expertise of forensic examiners extended to images presented upside down,
our results are also consistent with the proposal that mechanisms supporting unfamiliar face matching performance may
not be entirely face-specific, but may instead reflect general
image comparison abilities [24].

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

4. Methods
(a) Participants
Three groups of participants completed each test; we refer to
these groups as examiner, control and student. The examiner
and control groups comprised 41 volunteers (19 females; mean
age ¼ 42.4, s.d. ¼ 9.9) who attended the FISWG meeting in
May 2014. The role of FISWG is to create policy for best practices
in facial image comparison and training. Each participant in
these two groups completed a questionnaire designed to assess
their professional experience and training in forensic and facial
examination. The examiner group consisted of 27 FISWG attendees who stated that they regularly performed facial
examination as a part of their employment (average years
experience ¼ 7.3, s.d. ¼ 5.8; average hours per week ¼ 11.8,
s.d. ¼ 12.1). The remaining 14 attendees were also government
employees. Although this group was relatively small, they provided a valuable control group and so we tested them in the
same experimental session. A detailed comparison of the demographics of the examiner and control groups was not possible
because of requirements to protect the anonymity of participants.
Thus, it was not possible to match groups on age. Although we
have no reason to suspect that these groups differed in average
age characteristics, previous large-scale studies have found that
age of participants is not correlated with accuracy on perceptual
face matching tasks [9].
Students represent the most commonly tested population in
unfamiliar face matching experiments. Thus, we expected that student performance should approximate levels of performance
reported in the literature. We also expected students to be less
cognizant of task demands when compared with controls. The
controls were attending the FISWG meeting to create policy documents outlining ‘best practice’ in facial identification, and so we
anticipated somewhat better performance from the controls than
from the students based on their inherent interest in the task.
Students were undergraduates at the University of New South
Wales (n ¼ 32; 19 females, average age 21.4, s.d. ¼ 5.76).
Participants completed the tests in the following order: EFCT
with 2 s exposures (upright block, then inverted block), PICT,
GFMT and then the EFCT again with 30 s exposures (upright
block, then inverted block). All tests were administered on
laptop computers. Example image pairs from each test are
shown in figure 1.

(b) Glasgow Face Matching Test
The GFMT is a psychometric test designed to evaluate an individual’s ability to match identity across images of unfamiliar

6

Proc. R. Soc. B 282: 20151292

tests, examiners were not permitted unlimited time or access
to digital tools that would support decisions in daily work.
Further, participants made identity judgements on a Likert
scale that may not reflect normal reporting of identity judgements in forensic practice. In a recent study of professional
fingerprint examiners, participants could skip comparison
decisions on the basis that they did not provide sufficient evidence for identification [35]. Although this approach inhibits
the measurement of underlying perceptual skill [35,36], these
types of decisions are critical in minimizing costly workplace
errors, and recent work suggests that forensic examiners are
skilled in these types of judgements [11]. Thus, our results provide an estimate of the perceptual abilities of facial forensic
examiners that can serve as a benchmark for future tests
of identification accuracy in standard forensic practice, and
for computer-based face recognition systems that support
this practice.

rspb.royalsocietypublishing.org

To complicate this issue, deficits in face memory and perception have been associated with smaller FIEs [22,23], and people
with exceptional face processing ability show larger FIEs in both
memorial and perceptual tasks (Cambridge Face Perception
Test [23]). Our data show the opposite pattern. This raises the
additional possibility that visual expertise of forensic facial
examiners differs qualitatively from that which underpins face
processing in the population at large. This type of visual expertise may be dissociable from that shown by high performers
with no specific training in facial image comparison. Given
the emphasis on feature-by-feature approaches to comparison
in professional training [14,32] and the interaction between
image inversion and expertise reported here, it is possible that
increased selective attention to facial features improves performance. Thus, future studies that examine benefits of part-based
comparison strategies and identify visual cues subsisting accuracy in forensic examiners promise to elucidate the
foundations of expertise in forensic comparison.
It is also important to note that control participants performed very well despite not performing facial image
comparison in their daily work. As a group, control participants’ scores exceeded normative levels on the GFMT.
Moreover, in two tests, their performance did not differ statistically from examiners (EFCT 2 s exposure, PICT). This raises the
possibility that controls were more motivated than students,
and that examiners were more motivated than controls. Thus,
it is possible that differing levels of motivation might account
for performance differences across groups. Indeed, in any test
where groups of observers from different backgrounds are
compared, motivation may differ across groups and affect
performance. Here, we think it possible, and even likely, that
students, controls and examiners differed in their ‘selfinvestment’ in the results, and thereby in their motivation to
perform well. Previous research suggests, however, that
benefits of motivation alone are limited. Two different tests
of police and passport officers reported equivalent levels of
face matching accuracy to untrained students [10,11], despite
a clear motivation for these groups to perform well. Moreover,
in the present study, controls did not surpass professionals in
all tests. For example, examiners performed more accurately
in the EFCT test, but only at the longer exposure duration,
when the test conditions supported the employment of the
special skills and experience of the examiners.
Nevertheless, it will be important in future work to establish the relative contribution of natural ability, motivation,
experience and training to expertise in forensic examination.
As is typical in studies of expert populations [33], separating
their contributions to the emergence of expertise is problematic, and so longitudinal studies of forensic professionals
may be necessary to address this important question. Decoupling the influence of these factors will aid development of
recruitment and training methods for forensic examination.
It is also important when evaluating forensic evidence provided in the courtroom, where accurate assessments of
expertise are critical in establishing the appropriate weight
to be given to identity judgements [34]. For now, qualitative
differences in examiner performance suggest that differences
in cognitive processing contribute to their superior accuracy.
Finally, although forensic facial examiners performed
more accurately than the control and student groups, perfect
performance was not attained on any test, with average misclassification rates of around 7% on both the EFCT (30 s
upright) and GFMT. Because these were strictly perceptual

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

The stimuli for the PICT were sampled from those used in a
recent study that compared human and computer algorithm performance on a special set of image pairs for which machine
performance in the face recognition vendor test [15] was 100%
incorrect [17]. Specifically, similarity scores computed between
same-identity faces were uniformly lower than those computed
for the different-identity image pairs, suggesting that they contain no computationally useful identity information in the face.
Interestingly, in a recent study, untrained observers achieved
above-chance identification accuracy for these image pairs
owing to non-face identity cues from the body [17]. We included
this as a test of person identification ability for a set of image
pairs for which face recognition software fails.
We sampled 40 pairs of images (20 same-identity pairs) from
this dataset for the PICT. Participants were presented pairs in a
random order and the image pairs remained on the screen
until the participant’s response was registered. Response options
were as follows: (i) sure they are the same person; (ii) think
they are the same person; (iii) do not know; (iv) think they are
different people; and (v) sure they are different people. After
participants made a response, the next image pair was presented.

(d) Expertise in Facial Comparison Test
In designing the EFCT, our goal was to measure performance of
examiners with image pairs that challenge both computer face
recognition systems and untrained observers. We selected images
from The Good, the Bad and the Ugly Challenge [15], an image
dataset containing images from diverse and unconstrained ambient
conditions. This dataset was specifically designed to test state-ofthe-face recognition algorithms under challenging environmental
conditions, and contains frontal views of faces, taken with minimal
control of illumination, expression and appearance.

Ethics. This study was approved by the Institutional Review Board at
the University of Texas at Dallas. All participants provided written
informed consent and appropriate photographic release.
Data accessibility. Participant performance data: Dryad (http://dx.doi.
org/10.5061/dryad.ng720). Fusion analysis data: Dryad (http://dx.
doi.org/10.5061/dryad.ng720).

Authors’ contributions. D.W., P.J.P., C.A.H., M.H. and A.J.O. designed the
study; D.W., C.A.H., M.H. and A.J.O. collected the data; D.W., P.J.P.,
C.A.H., M.H. and A.J.O. analysed the data. All authors discussed the
results and commented on the manuscript.

Competing interests. The authors have no competing interests.
Funding. This research was supported by US Department of Defense
funding to A.J.O., and funding from the Australian Research Council
and the Australian Passport Office to D.W. and Richard Kemp
(LP110100448, LP130100702).

References
1.

2.

3.

4.

Jain AK, Klare B, Park U. 2012 Face matching
and retrieval in forensics applications. IEEE
MultiMedia 19, 20– 28. (doi:10.1109/MMUL.2012.4)
Dessimoz D, Champod C. 2008 Linkages between
biometrics and forensic science. In Handbook of
biometrics (eds AKJP Flynn, AA Ross), pp. 425–459.
New York, NY: Springer.
Philips PJ, O’Toole AJ. 2014 Comparison of human
and computer performance across face recognition
experiments. Image Vision Comput. 32, 74– 85.
(doi:10.1016/j.imavis.2013.12.002)
Jain AK, Ross A. 2015 Bridging the gap: from
biometrics to forensics. Phil. Trans. R. Soc. B 370,
20140254. (doi:10.1098/rstb.2014.0254)

5.

6.

7.

8.

Sacks MJ, Koehler JJ. 2005 The coming paradigm
shift in forensic identification science. Science 309,
892 –895. (doi:10.1126/science.1111565)
National Research Council. 2009 Strengthening
forensic science in the United States: A path
forward. Washington, DC: The National
Academies Press.
Kemp R, Towell N, Pike G. 1997 When seeing
should not be believing: photographs, credit cards
and fraud. Appl. Cognit. Psychol. 14, 211– 222.
(doi:10.1002/(SICI)1099-0720(199706)11:3,211::
AID-ACP430.3.0.CO;2-O)
Bruce V, Henderson Z, Greenwood K, Hancock PJB,
Burton AM, Miller P. 1999 Verification of face

identities from images captured on video. J. Exp.
Psychol. Appl. 5, 339–360. (doi:10.1037/1076898X.5.4.339)
9. Burton AM, White D, McNeill A. 2010 The Glasgow
face matching test. Behav. Res. Methods 42,
286–291. (doi:10.3758/BRM.42.1.286)
10. Burton AM, Wilson S, Cowan M, Bruce V. 1999 Face
recognition in poor-quality video: evidence from
security surveillance. Psychol. Sci. 10, 243 –248.
(doi:10.1111/1467-9280.00144)
11. White D, Kemp RI, Jenkins R, Matheson M, Burton
AM. 2014 Passport officers’ errors in face matching.
PLoS ONE 9, e103510. (doi:10.1371/journal.pone.
0103510)

7

Proc. R. Soc. B 282: 20151292

(c) Person Identification Challenge Test

First, image pairs (n ¼ 1 177 225) were ranked according to
machine performance using a fusion of top-performing algorithms in the Face Recognition Vendor Test, an international
benchmarking test for leading face recognition algorithms [15].
Images were then stratified according to match score data into
three subsets based on item accuracy: easy (the Good), moderate
(the Bad) or poor (the Ugly). For the ECFT, we choose only image
pairs from the Bad and Ugly portions of this dataset. Second,
performance of untrained human observers on a sample (n ¼
480) of the challenging items [16] was used as the basis for a
second selection. We combined these data with unpublished student performance on the same test, and recalculated item
performance for only the highest-performing 25% of participants.
Items on which 8% or more of high-performing participants
made errors were selected for the EFCT.
Participants completed the EFCT with upright and inverted
image pairs in two conditions that varied by exposure time. The
same image pairs were tested in 2 s and 30 s exposure time conditions. Because PICT and GFMT were administered between
the 2 s and 30 s conditions, there was a gap of roughly 1 h between the 2 s and 30 s EFCT tests. For each trial, images remained
visible for the prescribed exposure duration (2 or 30 s), and then disappeared. Response options were as follows: (i) sure they are the
same person; (ii) think they are the same person; (iii) do not
know; (iv) think they are different people; and (v) sure they are
different people. Participants could enter a response at any time
during the image display or after the image pair disappeared.
The next trial followed immediately. In total, the EFCT consisted
of 168 image pairs (half same-identity, half different-identity),
with half of these allocated to the upright test and half to the
inverted test.

rspb.royalsocietypublishing.org

faces [9]. Stimuli for the short version GFMT consisted of 20
same- and 20 different-identity image pairs. Same-identity
pairs show two images of the same person taken under similar
lighting conditions, on the same day, but using different digital
cameras. For different-identity pairs, one of these images was
paired with a similar-looking person from the database, so that
each identity appears once in a same-identity pair and once in
a different-identity pair. For each image pair, participants
responded ‘same’ or ‘different’ identity. The task was selfpaced, and image pairs remained on the computer monitor
until participants made their response, at which point the next
image pair was presented.

Downloaded from http://rspb.royalsocietypublishing.org/ on September 4, 2015

28. Kundel HL, Nodine CF, Conant EF, Weinstein SP. 2007
Holistic component of image perception in mammogram
interpretation: gaze-tracking study 1. Radiology 242,
396–402. (doi:10.1148/radiol.2422051997)
29. Thompson MB, Tangen JM. 2014 The nature of
expertise in fingerprint matching: experts can do a
lot with a little. PLoS ONE 9, e114759. (doi:10.1371/
journal.pone.0114759)
30. Wolfe JM, Vo˜ MLH, Evans KK, Greene MR. 2011
Visual search in scenes involves selective and
nonselective pathways. Trends Cogn. Sci. 15,
77– 84. (doi:10.1016/j.tics.2010.12.001)
31. Maurer D, Le Grand R, Mondloch CJ. 2002 The many
faces of configural processing. Trends Cogn. Sci. 6,
255–260. (doi:10.1016/S1364-6613(02)01903-4)
32. Woodhead MM, Baddeley AD, Simmonds DCV. 1979
On training people to recognize faces. Ergonomics
22, 333 –343. (doi:10.1080/00140137908924617)
33. Sternberg RJ. 1996 Costs of expertise. In The road to
excellence: the acquisition of expert performance in
the arts and sciences, sports, and games (ed. KA
Ericsson), pp. 347–354. Hillsdale, NJ: Erlbaum.
34. Edmond G et al. 2014 How to cross-examine
forensic scientists: a guide for lawyers. Aust. Bar
Rev. 39, 174 –196.
35. Ulery BT, Hicklin RA, Buscaglia J, Roberts MA. 2011
Accuracy and reliability of forensic latent fingerprint
decisions. Proc. Natl Acad. Sci. USA 108, 7733–
7738. (doi:10.1073/pnas.1018707108)
36. Tangen JM, Thompson MB, McCarthy DJ. 2011
Identifying fingerprint expertise. Psychol. Sci. 22,
995–997. (doi:10.1177/0956797611414729)

8

Proc. R. Soc. B 282: 20151292

19. Carey S, Diamond R. 1977 From piecemeal to
configurational representation of faces. Science 195,
312 –314. (doi:10.1126/science.831281)
20. Carey S. 1992 Becoming a face expert. Phil. Trans. R.
Soc. B. 335, 95 –103. (doi:10.1098/rstb.1992.0012)
21. Diamond R, Carey S. 1986 Why faces are and are not
special: an effect of expertise. J. Exp. Psychol. Gen.
115, 107–117. (doi:10.1037/0096-3445.115.2.107)
22. Avidan G, Tanzer M, Behrmann M. 2011 Impaired
holistic processing in congenital prosopagnosia.
Neuropsychologia 49, 2541 –2552. (doi:10.1016/j.
neuropsychologia.2011.05.002)
23. Russell R, Duchaine B, Nakayama K. 2009 Superrecognizers: people with extraordinary face
recognition ability. Psychon. B, Rev. 16, 252 –257.
(doi:10.3758/PBR.16.2.252)
24. Megreya AM, Burton AM. 2006 Unfamiliar faces are
not faces: evidence from a matching task. Mem.
Cogn. 34, 865– 876. (doi:10.3758/BF03193433)
25. O¨zbek M, Bindemann M. 2011 Exploring the time
course of face matching: temporal constraints
impair unfamiliar face identification under
temporally unconstrained viewing. Vision Res. 51,
2145 –2155. (doi:10.1016/j.visres.2011.08.009)
26. O’Toole AJ, Abdi H, Jiang F, Phillips PJ. 2007 Fusing
face-verification algorithms and humans. IEEE.
Trans. Syst. Man. Cybern. B 37, 1149 –1155. (doi:10.
1109/TSMCB.2007.907034).
27. White D, Burton AM, Kemp RI, Jenkins R. 2013
Crowd effects in unfamiliar face matching. Appl.
Cognitive Psychol. 27, 769–777. (doi:10.1002/
acp.2971)

rspb.royalsocietypublishing.org

12. Norell K, La¨the´n KB, Bergstro¨m P, Rice A, Natu V,
O’Toole A. 2015 The effect of image quality and
forensic expertise in facial image comparisons.
J. Forensic Sci. 60, 331 –340. (doi:10.1111/15564029.12660)
13. Wilkinson C, Evans R. 2011 Are facial image analysis
experts any better than the general public at
identifying individuals from CCTV images? Sci.
Justice 49, 191 –196. (doi:10.1016/j.scijus.2008.
10.011)
14. Facial Identification Scientific Working Group. 2011
Guidelines and recommendations for facial
comparison training to competency. See www.
fiswg.org/document/.
15. Philips PJ et al. 2011 An introduction to The
Good, the Bad, and the Ugly face recognition
challenge problem. In IEEE Int. Conf. on
Automatic Face and Gesture Recognition, 21–25
March 2011, Santa Barbara, CA, pp. 346–353.
Piscataway, NJ: IEEE.
16. O’Toole AJ, An X, Dunlop J, Natu V. 2012 Comparing
face recognition algorithms to humans on
challenging tasks. ACM. Appl. Percept. 9, 16.
17. Rice A, Philips PJ, Natu V, An X, O’Toole AJ.
2013 Unaware person recognition from the
body when face identification fails. Psychol.
Sci. 24, 2235–2243. (doi:10.1177/09567976
13492986)
18. Valentine T. 1988 Upside-down faces: a review of
the effect of inversion upon face recognition.
Brit. J. Psychol. 79, 471 –491. (doi:10.1111/j.20448295.1988.tb02747.x)