An evaluation of two commonly used tests of unfamiliar face recognition

https://doi.org/10.1016/S0028-3932(02)00222-1Get rights and content

Abstract

The Warrington Recognition Memory for Faces (RMF) and the Benton Facial Recognition Test (BFRT) are commercially available tests that are commonly used by clinicians and cognitive neuropsychologists to evaluate unfamiliar face recognition. Yet, it is not clear that a normal score on either instrument demonstrates normal unfamiliar face recognition. Because the RMFs stimuli contain abundant non-internal facial feature information, subjects may be able to score in the normal range without using internal facial features. On the BFRT, subjects commonly rely on feature matching strategies using the hairline and eyebrows rather than recognizing the facial configuration. To test whether these routes to recognition can support normal performance, normal subjects were tested with versions of the RMF and the BFRT in which the faces had been painted over in a way that prevented the operation of some of the procedures normally involved with face recognition. Even though these modifications removed all of the internal feature information in the RMF, many subjects scored in the normal range, and despite precluding the use of configural processing in the BFRT, many of the scores were in the normal range. As a result, it is apparent that normal scores on these tests do not demonstrate normal unfamiliar face recognition and so clinicians should be cautious in interpreting scores in the normal range. Finally, these results place in question models supported by dissociations involving normal performance on these tests.

Introduction

The Warrington Recognition Memory for Faces (RMF) [51] and the Benton Facial Recognition Test (BFRT) [8] are commercially available tests that are widely used by both clinicians and cognitive neuropsychologists to assess face recognition abilities. However, there is reason to question whether normal scores on these tests actually demonstrate normal unfamiliar face recognition. The photos in the RMF include many non-internal facial features, such as clothing and hair, by which they can be recognized. This raises the possibility that participants—in particular individuals with face recognition impairments—might be able to do well on the RMF by using a strategy that does not require the integrity of normal face recognition processes. The BFRT has the same problem, but for different reasons. Because it asks participants to match faces that are presented simultaneously, participants could answer correctly by matching individual features in a piecemeal fashion [41], [42]. This would not, however, show that face recognition processes are intact, because it appears that face recognition is performed by both parts-based procedures that represent facial features and configural processing procedures that represent the spatial relations of the parts of the face [14], [25], [26], [37], [40], [44]. Moreover and possibly more importantly, target faces and test items in the BFRT are presented simultaneously so participants are not required to rely on a memory trace.

As a result of these concerns about the RMF and BFRT, we conducted tests exploring these possibilities with modified versions. In our version of the RMF, parts-based and configural procedures could not be used on internal feature information, and our version of the BFRT precluded the use of configural procedures. If subjects can achieve a normal score nevertheless, then it will be clear that normal scores do not necessarily demonstrate normal face recognition abilities.

The RMF was designed as a test of non-verbal memory, but it is often used as a test a particular type of non-verbal memory—face memory. Subjects in the RMF are presented with black and white photos of 50 unfamiliar men at a rate of one item every 3 s, and they are asked to respond ‘yes’ or ‘no’ according to whether they find the face pleasant or not pleasant. Immediately following presentation of the target photos, memory for the photos is tested with a forced choice between a target photo identical to that presented earlier and a distracter photo. This means that stimulus recognition, not true face recognition, is possible [29]. The number of correct choices determines the score of the subject, so the maximum score is 50. Means range between 42.4 and 44.3 depending on the age group with standard deviations of approximately 3.5.

Even though the RMF is used to test face recognition, the photos contain information that is not internal facial feature information and could be used to discriminate between target and distracter photos. When attempting to test face recognition, the distinction between information handled by face recognition procedures and other information is clearly crucial. Although a number of investigations of face recognition have proposed that both the internal features (eyes, nose, mouth regions) and the external features (hair, ears) are recognized by face recognition procedures [21], [22], [54], neuropsychological evidence paints a different picture. Humans appear to use special procedures for face recognition [11], [24], [31], [37], [40], and the input from the face that these procedures typically operate on defines what counts as facial information. CK, an object agnosic with normal face recognition, has shown in a number of experiments that his recognition of famous faces from the internal features is normal, but his recognition of famous faces from external features is severely impaired [39], [40]. Conversely, prosopagnosics often rely on the external features, in particular the hair, for person recognition [1], [5], [13], [20], [23], [32], [36], [41], [46], although their proficiency with these features has not been formally demonstrated. Thus, although the external features can obviously contribute to person recognition and appear to be more important than internal features when learning to recognize a particular person [21], [22], [54], they do not appear to constitute part of the face for these specialized procedures.

As a result of these considerations, it appears that all information in the RMF other than the internal facial features would not be handled by specialized procedures. The photos consist of shots that display each model’s hair, face, and approximately one-third of their upper body (see Fig. 1). The hair, clothing, head postures, and body positions vary greatly between the models. In addition, some photos contain missing corners, low brightness levels, or developmental imperfections.

Recent experiments with a developmental prosopagnosic vividly illustrate the problem caused by the presence of the non-internal feature information [42]. EP’s performance was clearly impaired on tests of familiar and unfamiliar face recognition. However, EP was able to achieve a normal score on the RMF, and he reported using the non-internal feature information. In order to investigate his claim, EP was presented with the RMF without the information around the face. Not surprisingly, the mean for the control subjects was only two points lower in the modified version, because they were relying primarily on the internal feature information. In contrast, EP scored 31/50 on the modified test, and so it appears that the non-internal feature information does provide a route to normal performance.

Two versions of the BFRT [8] can be used depending upon time considerations: the Short Form has 13 items with 27 possible points while the Long Form has 22 items with 54 possible points. On each item, subjects are presented with a target photo, and they are asked to choose the target individual from six faces presented simultaneously with the target photo (see Fig. 2). There are three parts to the BFRT: (1) matching a frontal view of the target with an identical photograph, (2) matching a frontal view of the target individual with three photos of the target taken from different angles, and (3) matching a frontal view of the target individual with three photos of the target taken under different lighting conditions. No time limits are placed on the BFRT, and scores are classified as normal when 41 or above.

The black and white photos used in the BFRT consist of unfamiliar male and female faces with their hair and clothing shaded out so that subjects must rely on the face. On many of the items, the features that best differentiate the photographed individuals are the eyebrows and hairlines: normal subjects and prosopagnosic subjects to whom we have administered the test, as well as prosopagnosics in [20], [41], [42], have reported that their performance relied heavily on matching the eyebrows and hairlines rather than the internal facial features.

Because of the possibility that normal scores on the RMF and the BFRT do not reflect normal face recognition abilities, we have designed experiments that will determine how well participants can perform on these tasks without using all of the processes involved with normal face recognition. For the RMF, we covered the faces in the test phase so that participants were forced to rely on only the non-internal feature information available in the photos (see Fig. 3). For the BFRT, we covered all of the face except for the eyebrows and the hairline, so that participants could not use the facial configurations for matching (see Fig. 4). We will refer to these modified versions of the task as the mRMF and the mBFRT.

Because we deleted internal feature information, this allows us to test normal subjects under conditions similar to those experienced by individuals with normal parts-based processing but without facial configural processing. However, we deleted internal features that could have been used by parts-based processing so our versions provide an especially demanding test of the hypothesis that these tests can be passed using feature-based procedures. If some participants are able to perform in the normal range on either test, it will indicate that the original test does not require normal face recognition abilities. However, if participants have scores well out of the normal range, it will indicate that the original tests do, in fact, require intact face recognition.

Section snippets

Participants

The participants were 26 undergraduate students, 17 women and 9 men, who participated to satisfy a requirement for their introductory psychology class at UC-Santa Barbara.

Materials

The stimuli were digitized versions of the stimuli from the Warrington Recognition Memory for Faces test. The 50 two-choice test stimuli were scanned into a computer so that the faces could be painted over with a shade of gray. The gray extended from each model’s hairline to his chin vertically and horizontally from ear to ear

Experiment 1b: modified Recognition Memory for Faces (mRMF)

The previous experiment showed that the non-internal feature information alone provides enough information for participants to score in the normal range. However, the procedure differed in a few respects from the standard RMF, and these differences may have inflated the scores on the mRMF. First, participants were not required to make a ‘pleasant/unpleasant’ decision with regard to each face. Although past experiments indicate that such a decision produces better face recognition performance

Participants

The participants were 29 undergraduate students, 10 males and 19 females, from UC-Santa Barbara who participated to satisfy a requirement for their introductory psychology course.

Materials

In order to eliminate the possibility of configural face processing, the stimuli from the BFRT were modified to leave only the eyebrows and the hairline visible. After scanning the original test items, gray masks were painted over the face of each model up to the eyebrows (see Fig. 4). The modified stimuli from each

Discussion

In order to assess the RMF and the BFRT, we tested participants with modified versions of each test. Test faces in the mRMF were painted over so that they provided only non-internal feature information; the stimuli in the mBFRT presented only the eyebrows and the hairline. The mRMF did not allow either parts-based procedures or configural processing to operate on internal feature information in the test phase, and the mBFRT did not permit configural processing. Despite this, many participants

Acknowledgements

Advice and encouragement were provided by Leda Cosmides, John Tooby, and Ken Nakayama, and we also wish to thank Elizabeth Levy, Cathy Galvez, and Crystal McMillan for their assistance in constructing and running these experiments. These studies were supported by the James S. McDonnell Foundation, the NSF (Grant BNS 9157-499), a UCSB Research Across Disciplines Grant, and the NIH (Grant 1 F32 MH64246-01).

References (55)

  • K. Henke et al.

    Specificity of face recognition: recognition of exemplars of non-face objects in prosopagnosia

    Cortex

    (1998)
  • A.J. Mattson et al.

    A case of closed prosopagnosia following moderate closed head injury with left hemisphere focal lesion

    Cortex

    (2000)
  • E.C. Shuttleworth et al.

    Further observations on the nature of prosopagnosia

    Brain and Cognition

    (1982)
  • N. Takahashi et al.

    Prosopagnosia: a clinical and anatomical study of four patients

    Cortex

    (1995)
  • P. Vuilleumier

    Faces call for attention: evidence from patients with visual extinction

    Neuropsychologia

    (2000)
  • P. Vuilleumier et al.

    Two eyes make a pair: facial organization and perceptual learning reduce visual extinction

    Neuropsychologia

    (2001)
  • G. Assal

    Regression des troubles de la reconnaissance des physionomes et de la memoire topographique chez un malade opere d’un hematome intracerebral parieto-temporal droit

    Revue Neurologique

    (1969)
  • R. Barbarotto et al.

    Slowly progressive semantic impairment with category specificity

    Neurocase

    (1995)
  • S. Bentin et al.

    Selective visual streaming in face recognition: evidence from developmental prosopagnosia

    NeuroReport

    (1999)
  • A.L. Benton

    The neuropsychology of face recognition

    American Psychologist

    (1980)
  • Benton AL, Sivan AB, Hamsher K, Varney NR, Spreen O. Contribution to neuropsychological assessment. NY: Oxford...
  • V. Bruce et al.

    Understanding face recognition

    British Journal of Psychology

    (1986)
  • L.J. Buxbaum et al.

    Impaired face and word recognition without object agnosia

    Neuropsychologia

    (1999)
  • R.K. Cabeza

    Features are also important: contributions of featural and configural processing to face recognition

    Psychological Science

    (2000)
  • G.A. Carlesimo et al.

    Components in the visual processing of known and unknown faces

    Journal of Clinical and Experimental Neuropsychology

    (1995)
  • S. Della Sala et al.

    Cognitive modeling of face processing: evidence from Alzheimer patients

    Neuropsychologia

    (1995)
  • B.C. Duchaine

    Developmental prosopagnosia with normal configural processing

    NeuroReport

    (2000)
  • Cited by (118)

    • Relationship between social cognition, general cognition, and risk for suicide in individuals with a psychotic disorder

      2021, Schizophrenia Research
      Citation Excerpt :

      Facial recognition was measured with the Degraded Facial Affect Recognition Task (DFAR (Van't Wout et al., 2004)), by instructing individuals to indicate emotional expressions (neutral, happy, fearful and angry) of 64 facial pictures at 70% intensity, i.e. slightly degraded (Maat et al., 2015; Van't Wout et al., 2004). In order to rule out the effect of a general facial recognition inability on the DFAR task, the Benton Facial Recognition Task (Benton et al., 1983; Duchaine and Weidenfeld, 2003) was administered, and used as a covariate. The Hinting Task (Corcoran et al., 1995) was used to measure ToM and assesses an individual's ability to indicate real intentions from implicit messages in ten short stories containing interaction between two protagonists.

    View all citing articles on Scopus
    View full text