British Journal of Pycbofogy (1993), 84, 433441 Printed in Great Britain 433

0 1993 The British Psychological Society

Effect of disorientation on visual analysis,

familiarity decision and semantic decision on
faces

R. Bruyer*, C. Galvez and C. Prairial

Louvain Universig ( U C L ) , Psychology Department, Cognitive Neuropsychology Unit ( N E C O ) , Voie du
Roman Pays 20, B-1348 Louvain la Neuve, Belgium

The disproportionate effect of vertical inversion on the processing of faces has often
been interpreted as indicating differences in perceptive strategies between upright
and inverted faces. If this is true, then the mathematical relation between the
performance and the angular deviation would not be a linear function. Thirty
subjects processed faces shown under 10 angular deviations from upright (0 deg) to
upside-down (180 deg) in steps of 20 degrees, in a random design. Three cognitive
levels (a between-subject factor) were tested, namely, gender classification of
unfamiliar faces, familiarity decision on famous vs. unknown faces, and semantic
decision (occupation) on famous faces, with 480 items per subject and task. We
found no sign of a change of processing strategies for the upright vs. inverted
presentations: a clear linear function emerged. However, the nature of the task
could influence the effect of the angular deviation on the performance. Results
suggest that, up to now, there is no definite proof of qualitatively different
mechanisms underlying the processing of upright and inverted faces, respectively.

The processing of vertically inverted visual complex stimuli is difficult, especially

when faces are tested (see Valentine, 1988, for a review). This disproportionate
inversion effect (DIE) has been suspected to be face specific (e.g. Yin, 1969), but
recent experimental data by Diamond & Carey (1986) suggest it applies to other
complex visual material (photographs of dogs) for subjects who have developed a
particular skill or expertise in the identification of individual exemplars (dog experts).
Faces are probably a visual material for which such an expertise has been acquired
universally by human beings (see, however, Bruyer & Crispeels, 1992). Moreover,
this DIE is not observed in children (e.g. Carey & Diamond, 1977; Carey, Diamond
& Woods, 1980), nor after damage to the posterior region of the right cerebral
hemisphere (Yin, 1970), nor when the stimuli are displayed in the right visual
hemifield (left hemisphere) of normal adults (e.g. Leehey, Carey, Diamond & Cahn,
1978). In addition, the DIE seems to depend on the level of processing that is tapped
by the task, since real face recognition is required (Valentine & Bruce, 1988; see also
Cochran, Pick & Pick, 1983) and memory must be involved to reveal the inversion
effect (see Valentine, 1988).
Several hypotheses have been suggested to account for this phenomenon.
* Requests for reprints.
16 P S Y 84
434 R. Bruyer, C. Galvex and C. Prairial
Goldstein & Chance (1980) proposed that it results from the development of an
increasingly rigid face schema that disturbs the processing of unusual faces. Yin
(1970) suggested that faces are recognized mainly by extraction of their general
expression, a process that would be prevented by inversion. Carey & Diamond
(1977; see also Rock, 1973) proposed that faces are normally processed by both
component (or piecemeal) and configural items of information (a proposition that
later received some support from Sergent, 1984) and that inversion specifically
prevents configural processing. Diamond & Carey (1986) suggested that the DIE
applies to those stimuli from which second-order relational properties must be
extracted to process them (but see Tanaka & Farah, 1991).
Even though Valentine (1988) concluded his review in stating that there is no
compelling evidence of qualitatively different processing of upright and inverted
faces, authors generally suspect that upright and inverted faces are processed by
qualitatively different strategies or mechanisms. If this interpretation is correct, then
the function relating subjects’ level of efficiency to the degree of departure from the
usual upright orientation should be non-linear. (It could be suggested that this
argument is a bit strong, and that a linear relationship is not sufficient to exclude
different cognitive processes, as illustrated recently by Sergent & Corballis, 1989. We
would like to stress the reverse argument, i.e. a non-linear function is necessarily
incompatible with a single cognitive process.) Nevertheless, such an analysis requires
faces to be displayed at several angles of rotation, while the large majority of
published studies have used only upright vs. inverted faces, i.e. zero vs. 180 degrees.
There are, however, a few papers in which the angle of rotation was manipulated (see
below). In general, the data revealed a linear relationship, which would not firmly
support the two-process interpretation. Nevertheless, beyond the difficulties specific
to each study, few angles were used, and this could have artificially prevented the
discovery of non-linear functions.
Thus, Cochran e t al. (1983) used five angles of rotation (0,45,90,135 and 180 deg)
and found a linear relationship when the task was discrimination between a stimulus
and its mirror image (successive presentation) and a non-linear relationship when the
subjects had to distinguish between different identities. However, the very artificial
nature of the stimuli-highly schematic profile faces-precludes any generalization
to face processing. The subjects in Rock’s (1974) study had to recognize famous faces
under four kinds of rotation (0,45,90 and 180 deg). A linear relation was found, but
the manipulation of the rotation was accomplished by moving the subjects’ head.
Sergent & Corballis (1989) submitted normal and split-brain subjects (i.e. patients
whose cerebral hemispheres are surgically disconnected) to a gender categorization
of faces whose photographs were shown at 0, 60, 120, 180,240 or 300 degrees from
the upright orientation in the left or right visual hemifield. The right hemisphere (left
hemifield) was less disturbed than the left by rotation, and a monotonic relation (with
respect to the departure from 0 degrees, clockwise or counterclockwise) emerged in
the two conditions for both samples. Finally, Valentine & Bruce (1988) showed
photographs of faces at five possible angles (0, 45, 90, 135 and 180 deg) to normal
subjects. In their first experiment, where two successive faces had to be matched for
identity, a linear function emerged and it was not affected by the same vs. different
kind of facial expression. In their second experiment, the subjects had to process one
 Effect of disorientation on visual anabsis 435
stimulus at a time and a linear relationship appeared for the categorization of stimuli
in normal vs. scrambled faces (a facial decision task), as well as for expression and
familiarity decisions, but no specific relation emerged for the discrimination of faces
from common objects (another kind of face decision).
It thus seems that the processing of expression cannot account for the DIE (i.e.
the specific difficulty in processing inverted faces cannot be accounted for by inability
to recognize the inverted facial expression, as suggested by Yin, 1969; this
observation supports the view-for instance, Bruce & Young, 1986-that face
recognition and facial expression recognition are independent parallel processes) and
the data also suggest that a common mechanism underlies the processing of both
upright and inverted faces. However, the small number of angles precluded the
emergence of non-linear relationships (and constitutes a difficulty for regression
analyses as performed by Valentine & Bruce, 1988). In the present study, 10 different
angles are used (in steps of 20 degrees, from 0 to 180). A survey of the literature also
reveals that the DIE appears when memory components are involved and when the
processing of individual realistic representations of faces is required. However, even
within this limited framework, several levels of processing can be distinguished
according to current cognitive models of face processing (see Bruce & Young, 1986).
Therefore, the second purpose of the present study is to compare three tasks using
photographs of faces, namely, a gender categorization task, which requires the
detection of visually derivable semantic information (not needed for the recognition
of faces per se), a familiarity decision task, which taps the level of face-recognition
units (a face-recognition unit is the structural representation, stored in long term
memory, of a face already known by the observer), and a semantic decision task
(occupations), which taps deeper semantic information not visually derivable and
available for well-known faces only.
With regard to this second objective, it should be noted that the previous studies
on mental rotation never went beyond the stage of face-recognition units. One can
therefore speculate that deep, semantic processes could reveal non-linear functions
and thus provide evidence for separate mechanisms for processing upright and
inverted faces.
A pilot experiment had been performed, which motivated the design of the study
reported in this paper. While the results of this pilot study were promising by
showing a linear relationship between performance and the angular deviation, the
experiment was contaminated by enrolling the 30 subjects in a design where the task
was a within-subject factor, and where clockwise deviations only were used with
three displays of each face under each of the 10 angular deviations. Therefore, a new
experiment was undertaken involving 30 new subjects, where (a) the task became a
between-subject factor, (b) each face was shown six times in each orientation, and (c)
these six presentations were three clockwise and three counterclockwise displays of
each face.
Method
Each of the 30 subjects was shown 480 items, in a single task per subject.

16-2
436 R. Bryer, C . Galvex and C. Prairial

Sulyects
Thirty young right-handed adult subjects (15 females and 15 males), were recruited for an entirely
between-subject design. They all had normal or corrected visual acuity.

Material and experimental design

The complete material consisted of full-facephotographs of 16 people. There were eight unknown faces
(four females, four males) and eight well-known faces (celebrities) of which four were of politicians (two
females, two males) and four of people with other activities (two females, two males). All the faces, both
unknown and well known, were taken from magazines, and the various sets were roughly matched for
apparent age. In addition, the non-politicians did not differ from politicians by surface cues (e.g.
athletes were not used) since famous persons were mainly actors and television commentators.
Familiarity and professional activity were unequivocal, according to a pretest by 10 judges. Black-and-
white slides were made of each face at 10 clockwise and 10 counterclockwise angles of departure from
the vertical upright orientation (0 to 180 deg, in steps of 20 degrees, in both directions). Each face
subtended an angle of four degrees of the visual field.
The tasks differed with respect to the subsets of faces used. For the gender categorixation task, only
unknown faces were used. The eight faces were displayed six times (3 clockwise and 3 counterclockwise)
in 10 orientations, so this condition consisted of 480 trials. For the famiharig decision task, the 16 faces
(eight famous, eight unknown) were each shown three times in the 10 angular deviations resulting in
a total of 480 trials. For the semantic decision task, only the eight famous faces were used (480 trials, again).
Within each task the 480 trials were randomly distributed. In each case, the subject had to make a
binary decision, namely, male/female, familiar/unfamiliar, or politician/non-politician. The response
was made by means of two push buttons controlled by the left and right index fingers. The assignation
of buttons was balanced across subjects.

Procedure
Each subject underwent three series of 160 trials, with a short break between each. The sequence of each
trial was as follows : auditory warning signal, delay of 1 s, display of the stimulus until the response was
given, intertrial interval of 8 s. The entire experimental session lasted about 90 min.
The subject was asked to respond as quickly as possible without error. Response times were recorded.

Results
The proportions of errors were generally low (grand mean = 2.7 per cent; see Fig.
1). A mixed, 3 (task) x 10 (angular deviations) ANOVA revealed a marginally
significant main task effect (F(2,27) = 3.29, p = .052), a significant main angular
deviation effect (F(9,243) = 2 . 2 0 , ~= .023), and a significant task x angular deviation
interaction (F(18,243) = 2.88, p < .0002). Orthogonal polynomial analyses were
performed in search of linear trends: these analyses did not reveal any significant
trend (all Fs < 1).
Correct RTs (Fig. 2) were studied by the same procedure as errors. The 3 x 10
ANOVA revealed, first of all, a significant main task effect (F(2,27) = 17.25, p <
.OOOl) : the post hoc comparisons, using the Newman-Keuls test, showed a significant
advantage of gender categorization (572 ms) over the two remaining tasks which did
not differ from each other (semantic decision: 730; familiarity decision: 734). The
main angular-deviation factor was also significant (F(9,243) = 23.94, p < .OOOl), but
the significant task x angular deviation interaction modified this effect (F(18,243) =
2.91, p < .0002). The post hoc analysis of this interaction was computed:
 Effect of disorientation on visual anabsis 437

3.5

I----
3 familiarity
decision
2.5
2
semantic
1.5 decision
1 =--. - * ,
gender
0.5
0
- ;/ categoriz.

0 20 40 60 80 100 : 120 : 140 : 160 : 180 I

Figure 1. Mean number of errors out of 48 (means of 10 subjects per task) according to the task and
the angular deviation.

(a) The main effect of task applied exactly for each angular deviation, except for the
120 deg deviation where the three tasks differed significantly from each other;
(b) For the gender categorization task, the subjects were faster at deviations of 0 and
40 degrees than at all the remaining deviations, and faster at 60 degrees than at
20 degrees ;
(c) For the familiarity decision task, there was an advantage of 0 and 40 degrees over
the others except 160 and 180 degrees, and an advantage of 20 and 60 degrees
over 80, 120, 140 and 180 degrees of angular deviation;
(d) In the semantic decision, we noted (a) an advantage of 0,40,60 and 100 degrees
of deviation over 20,140,160 and 180 degrees, and (b) an advantage of 0 and 60
over 120 degrees, which in turn was executed better than in the 160 and 180 deg
conditions.

semantic
decision

650

500 1 *
0 20 40 60 80 100 120 140 160 180
Figure 2. Mean latency of the correct responses (means of 10 subjects per task) according to the task
and the angular deviation.

Orthogonal polynomial analyses revealed a clear and significant linear component

in each task: for gender categorization (F(1,243) = 1 4 . 6 2 , ~< .OOl), accounting for
53 per cent of the variance; for familiarity decisions (F(1,243) = 58.62, p < .OOl),
 438 R. Brzyer, C. Galvex and C. Prairial
accounting for 76 per cent of the variance; and for semantic decisions (F(1,243) =
68.54, p < .OOl), responsible for 71 per cent of the variance’.
Linear regressions and correlations between performance and the angular deviation
(AD) were computed (Fig. 3). For gender categorization, no significant relationship
emerged for errors (r = .382) or for correct RTs (r = .511). For familiarity decision,
there was no significant relationship between errors and angular deviations
(r = .318), and RT = 702.5 +0.353AD (r = .799, p < .Ol). For semantic decision,
ERR = 0.00012+0.0164AD (r = .706, p < .05) and R T = 695.9+0.382AD
(r = .756, p < .05). The intercepts express global task differences, while the slopes
indicate task differences in the influence of angular deviation on performances.

8 -

6 - 8

4 .)

semantic
600 gender categoriz.
2 ‘I

550
I

Figure 3. Linear regressions computed for the number of errors (left) and the correct latencies (right)
according to the task and the angular deviation.

Discussion
Concerning the task effect, as expected from current cognitive models (for instance,
Bruce & Young, 1986), gender categorization was faster and easier than the other
tasks; in addition, the familiarity decision was not easier nor faster than the semantic
decision task, which was unexpected. Models predict that a semantic decision should
be slower than a familiarity decision because the activation of a face-recognition unit
precedes the semantic judgement and is necessary for an efficient semantic decision.
Nevertheless, only famous faces were used for the semantic decision task, so it can
be argued that the familiarity decision was neither necessary nor made (the subject
knew in advance that faces were well-known), which could have cancelled the
predicted difference. Furthermore, in the familiarity decision task, which was
designed to tap the face-recognition units (Young & Ellis, 1989), the subjects were
asked to make a ‘celebrity decision’. Therefore, it remains possible that such a
Some more complex components reached a significant threshold. For gender classification, the quintic trend
(F(1,243) = 11.6, p < ,001) accounted for 42 per cent of the variance. For familiarity decisions, the quartic
(F(1,243) = 6.12, p < ,025; 8 per cent) and the quintic (F(1,243) = 9.31, p < .005; 12 per cent) trends were
significant. Finally, for semantic decisions, the quadratic (F(1,243) = 6.49, p < .025; 7 per cent), quartic
(F(1,243) = 1 3 . 4 2 , ~< .001; 14 per cent) and quintic (F(1,243)= 6.18,p < .025; 6 per cent) trends were significant.
These effects probably resulted from a ‘peak’ observed for the 20 deg deviation. However, this peak was clearly
weak and the linear component accounted for the major part of the variances.
 Effect of disorientation on vistlal ana4sis 439
decision concerns in fact semantic representations, i.e. a code deeper than recognition
units. Moreover, faces were repeatedly shown so that an experimental familiarity was
probably generated even for unknown faces. It is also possible that this reflects the
necessary inclusion of unfamiliar faces in the familiarity decision task and that the
unexpected result was due to negative (unfamiliar) decisions.
The inversion effect, i.e. a significant difference between upright (0 deg) and inverted
faces (180 deg), was not obvious. Indeed, firstly, it was never observed for errors.
However, for errors there was no effect of angular deviation in the gender
categorization task, no significant trend emerged in the polynomial analysis whatever
the task, the correlation between the error rate and the angular deviation did not
reach a statistical threshold for gender categorization and familiarity decision, and,
more generally, errors were not numerous. Secondly, for RTs, the inversion effect
was significant for gender categorization and semantic decision but not for familiarity
decision. One could speculate that this reduction of the inversion effect was due to
the number of trials per task, that is to say, an effect of practice. More widely, the
weakness of the effect could have resulted from the use of a set of intermediate
angular deviations, as compared with more classical experiments (see Introduction)
where only upright (0 deg) and inverted faces (180 deg) are shown.
The main purpose of this study was the analysis of the ftlnction relatingperfrmance
to the angtllar deviation. The study of errors was not illuminating, probably due to the
low error rate. For RTs, however, the picture seems to be clearer. In each task, the
linear component reached a significant threshold. Furthermore, even if some other
components were also significant, the linear component accounted for the major part
of the variance, even for the weakest (gender categorization: 53 per cent). Finally, the
linear correlation coefficient crossed the critical threshold for the semantic decision
and the familiarity decision tasks. Thus, we could cautiously conclude that a linear
relationship exists. Nevertheless, we do not mean to say that the subjects actually
made a mental rotation in order to obtain an upright orientation of the stimulus seen
(probably due to the kind of task used here: see Sergent & Corballis, 1991). Indeed,
the regression analyses would suggest a rotation speed of about 2770 deg/s, which
is clearly too high when compared with known data involving other types of three-
dimensional material (see, for instance, Metzler & Shepard, 1974).

Conclusions
On the one hand, we did observe the inversion effect for familiarity decisions and
semantic decisions, but not in the gender categorization task. This is consistent with
the literature (see Valentine, 1988, for a review) in showing that memory components
must be involved to reveal the inversion effect for faces. On the other hand,
performance was linearly related to the angular deviation, with this linearity being a
bit weaker for gender categorization than for the two other tasks. This observation
of a linear relationship also concurs with the literature (see the Introduction, and
below). In addition, it reinforces the phenomenon since we tested 10 levels of angular
deviation, used realistic stimuli and tapped three levels of the cognitive architecture
in a single study.
If one considers the available data and organizes them within the cognitive
440 R. Bryer, C. Galvex and C . Prairial
architecture summarized by Bruce & Young (1986) and Ellis (1986), the following
appears. Concerning the very early stages of the processing, no systematic
relationship was observed in the discrimination of faces from common objects
(Valentine & Bruce, 1988), but a linear relation emerged in the discrimination of
faces from scrambled faces (Valentine & Bruce, 1988). With regard to the structural
encoding of faces, a linear function appeared for the discrimination of faces from
their mirror-image (Cochran et al., 1983) and a non-linear function for the matching
of identities (Cochran etal., 1983). However, Valentine & Bruce (1988) noted a linear
relation in the matching of identities by using more realistic stimuli than those used
by Cochran e t al. Concerning the level of face-recognition units, a linear relationship
has been observed for episodic memory recognition of famous faces (Rock, 1974) as
well as for familiarity decisions (Valentine & Bruce, 1988; the present study). With
respect to the identity-specific semantic codes, a linear relation was noted for a
semantic categorization task (the present study). Finally, linear relationships were
noted for visual operations which are not mandatory for the recognition of faces, i.e.
facial expression analysis (Valentine & Bruce, 1988) and gender categorization
(Sergent & Corballis, 1989; the present study).
In short, while a linear relationship can be interpreted as indicating either a single
common mechanism or two (or more), a non-linear (e.g. quadratic) relation is
inconsistent with the single-mechanism explanation and requires considering at least
two different kinds of cognitive operations. So, we can conclude that, up to now, no
firm proof has been provided that upright and inverted faces are processed by
different mechanisms, whatever the cognitive level studied.

References
Bruce, V. & Young, A. (1986). Understanding face recognition. BritishJournalof Psychology, 77,305-327.
Bruyer, R. & Crispeels, G. (1992). Expertise in person recognition. Bulletin of the Psychonomic Socieg (in
press).
Carey, S. & Diamond, R. (1977). From piecemeal to configurational representation of faces. Science, 195,
312-314.
Carey, S., Diamond, R. & Woods, B. (1980). The development of face recognition: A maturational
component? Developmental PgJcbology, 16, 257-269.
Cochran, E. L., Pick, A. D. & Pick, H. L. (1983). Task-specific strategies of mental ‘rotation’ of facial
representations. Memory and Cognition, 11, 41-48.
Diamond, R. & Carey, S. (1986). W h y faces are and are not special: An effect of expertise. Journal of
Experimental Psychology :General, 115, 107-1 17.
Ellis, H. D. (1986). Processes underlying face recognition. In R. Bruyer (Ed.), The Neuropsychology of Face
Perception and Facial Expression, pp. 1-27. Hillsdale, NJ : Erlbaum.
Goldstein, A. G. & Chance, J. E. (1980). Memory for faces and schema theory. Journalof Pycbology, 105,
47-59.
Leehey, S., Carey, S., Diamond, R. & Cahn, A. (1978). Upright and inverted faces: The right
hemisphere knows the difference. Cortex, 14, 41 1-419.
Metzler, J. & Shepard, R. N. (1974). Transformational studies of the internal representation of three-
dimensional objects. In R. L. Solso (Ed.), Theories of Cognitive Psychology: The h y o l a Symposium.
Potomac, MD: Erlbaum.
Roberts, T. & Bruce, V. (1988). Feature saliency in judging the sex and familiarity of faces. Perception,
17, 475-481.
Rock, I. (1973). Orientation and Form. New York: Academic Press.
Rock, I. (1974). The perception of disoriented figures. ScientijiC American, 230, 78-85.
 Effect of disorientation on visaal anaCysis 441
Sergent, J. (1984). An investigation into component and configural processes underlying face
perception. British Journal of Psychology, 75, 221-242.
Sergent, J. & Corballis, M. C. (1989). Categorization of disoriented faces in the cerebral hemispheres
of normal and commissurotomized subjects. Journal of Experimental Psychology :Human Perception and
Performance, 15, 701-710.
Sergent, J. & Corballis, M. C. (1991). Ups and downs in cerebral lateralization. In F. L. Kitterle (Ed.),
Cerebral Laterality : Theory and Rerearch, pp. 175-200. Hillsdale, NJ : Erlbaum.
Tanaka, J. W. & Farah, M. J. (1991). Second-order relational properties and the inversion effect: Testing
a theory of face perception. Perception and Psychopbysics, 50, 367-372.
Valentine, T. (1988). Upside-down faces: A review of the effect of inversion upon face recognition.
British Journal of Psychology, 79, 471491.
Valentine, T. & Bruce, V. (1988). Mental rotation of faces. Memory and Cognition, 16, 556-566.
Yin, R. K. (1969). Looking at upside-down faces. Journal of Experimental Psychology, 81, 141-145.
Yin, R. K. (1970). Face recognition by brain-injured patients : a dissociable ability? Neuropsychologia, 8,
395402.
Young, A. W. & Ellis, H. (1989). Semantic processing. In A. W. Young & H. D. Ellis (Eds), Handbook
of Research on Face Processing, pp. 235-262. Amsterdam: Elsevier.

Received 4 August 1989; revised version received 15 May 1992