Evaluating the effect of motion and body shape on the perceived sex of virtual characters.

Full text

Page 1

20
Evaluating the Effect of Motion and Body Shape
on the Perceived Sex of Virtual Characters
RACHEL MCDONNELL, SOPHIE J¨ORG, FIONA NEWELL, and CAROL O’SULLIVAN
Trinity College Dublin
JESSICA K. HODGINS—Carnegie Mellon University
In this paper, our aim is to determine factors that influence the perceived sex of virtual characters. In Experiment 1, four
different model types were used: highly realistic male and female models, an androgynous character, and a point light walker.
Three different types of motion were applied to all models: motion captured male and female walks, and neutral synthetic
walks. We found that both form and motion influence sex perception for these characters: for neutral synthetic motions, form
determinesperceivedsex,whereasnaturalmotionaffectstheperceivedsexofbothandrogynousandrealisticforms.Theseresults
indicate that the use of neutral walks is better than creating ambiguity by assigning an incongruent motion. In Experiment 2
we investigated further the influence of body shape and motion on realistic male and female models and found that adding
stereotypical indicators of sex to the body shapes influenced sex perception. Also, that exaggerated female body shapes influences
sex judgements more than exaggerated male shapes. These results have implications for variety and realism when simulating
large crowds of virtual characters.
Categories and Subject Descriptors: I.3.7 [Computer Graphics]: Three Dimensional Graphics and Realism—Animation
General Terms: Experimentation, Human Factors
Additional Key Words and Phrases: Perception, graphics, motion capture
ACM Reference Format:
McDonnell, R., J¨ org, S., Hodgins, J. K., Newell, F., and O’Sullivan C. 2009. Evaluating the effect of motion and body shape on
the perceived sex of virtual characters. ACM Trans. Appl. Percpt. 5, 4, Article 20 (January 2009), 14 pages. DOI = 10.1145/
1462048.1462051 http://doi.acm.org/10.1145/1462048.1462051
1.
Animated virtual humans are needed for many applications in entertainment, education, and science.
Their movements can be created by artists, by programs (procedural animation), or recorded from
real people using motion capture technology. Crowd simulation systems, in which thousands of virtual
humansnavigaterealisticenvironmentssuchascitiespresentaparticularchallenge.Crowdssimulated
withsyntheticwalkingmotionscanlackpersonality,somotioncaptureddatacanbeusedtoaddrealism
(Figure 1).
In this paper, we investigate some factors that affect the perceived sex of walking virtual humans,
with a view to increasing the realism of pedestrians in real-time crowd simulations. We cannot simulate
INTRODUCTION
Authors’ address: R. McDonnell, S. J¨ org, F. Newell and C. O’Sullivan, Graphics Research Group & Institute of Neuroscience,
Trinity College Dublin; J. K. Hodgins, School of Computer Science, Carnegie Mellon University.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided
that copies are not made or distributed for profit or direct commercial advantage and that copies show this notice on the first
page or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute
to lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permissions may be
requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481,
or permissions@acm.org.
c ? 2009 ACM 1544-3558/2009/01-ART20 $5.00 DOI 10.1145/1462048.1462051 http://doi.acm.org/10.1145/1462048.1462051
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 2

20:2
•
R. McDonnell et al.
Fig. 1.Crowd simulation requires variety in colour, form, and motion [Dobbyn et al. 2006].
Fig. 2.
Woman model, Man model, Androgynous figure, and Point light walker.
everyone in a crowd with their own personal motion captured walk, as the more motions we use, the
greater the demands on potentially limited computational and memory resources (e.g., a games console
or hand-held device). Therefore, the challenge is to optimise quality and variety with the resources
available. Specifically, we ask the question whether, if there is a clear visual indicator of sex (i.e., a
highly realistic, unambiguously female or male model, as shown in Figure 2), will motion or form
information dominate our perception of the sex of the character? If motion information alone always
determines perceived sex, then we would always need to create templates of every different motion for
both males and females. However, if we find that form dominates or that simulated neutral motions are
as good as captured natural motions under some circumstances, then such duplication may not always
be necessary. Perhaps some actors’ walks can be equally effectively applied to both male and female
models. Any of these results would allow us to create “canonical” motions to which variety could later
be added, irrespective of the sex of the model.
Four model representations were animated with real female, real male, or synthetic neutral motions. From left to right:
2.
People expect a virtual character to behave in a manner befitting its appearance and they find discrep-
ancies in behaviour disturbing [Vinayagamoorthy et al. 2006]. Furthermore, people’s perception of the
sex of a human representation has been studied in the experimental psychology literature, with respect
to human motion with minimal shape information and the effect of body shape. Only recently have the
effects of both shape and motion been considered together. A note on terminology: sometimes gender is
used in the literature to refer to the sex of a person, i.e., male or female. However, gender perception
refers to the classification of the femininity or masculinity of a human, which is a different metric, e.g.,
a man can have a feminine walk but should still be classified as a male.
BACKGROUND
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 3

Evaluating the Effect of Motion and Body Shape on the Perceived Sex of Virtual Characters
•
20:3
Johansson [1973; 1976] reported that natural motion, in the absence of any spatial information, is
a sufficient cue to determine the sex of a walker. His ‘point-light’ displays were designed to separate
biological motion information from other sources of information that are normally intermingled with
the motion of a human, such as form or outline. He showed that 12 moving light points suffice to create
a rich perception of a moving human figure within a very short space of time (200msec, or five frames
of a movie). They also reported that the sex of the walker could be identified even when the number of
lights was reduced to two on the ankles. However, it was later shown that this could have been mainly
due to stride length [Kozlowski and Cutting 1978].
Many studies since have used Johansson’s point-light displays to show that biological motion per-
ception, an extreme example of sophisticated pattern analysis in the brain, extends even further, e.g.,
to recognising a particular walker [Cutting and Kozlowski 1977] or even one’s own walking pattern
[Beardsworth and Buckner 1981]. Kozlowski and Cutting [1977] explored some of the parameters of
sex recognition without familiarity cues and showed that, while point-lights are sufficient for recog-
nising the sex of a walker for dynamic sequences, static images were found to be insufficient, i.e., the
human form cannot be identified when using static point-light stimuli alone. Troje [2002] implemented
a classifier that could discriminate between male and female walkers. When comparing the behaviour
of the classifier with human observers they found that humans rely on dynamic information more than
structural when classifying a walk as male or female.
Mather and Murdoch [1994] found that using views which display sex-specific lateral body sway,
rather than the sagittal (side) views of earlier studies, improved recognition accuracy. Pitting structural
and dynamic cues against each other, their point light walkers were synthetically created with either a
male or female torso or an androgynous structure. They applied male and female synthetic walks to the
male and female body structures and an androgynous walk to the androgynous structure. They found
that, even when looking at incongruent pairings of torso and walks, the motion was a more salient
source of sex information than the body structure. However, altering point light walkers to display
different body types may still leave uncertainty with respect to their sex, so perhaps it is not surprising
thatmotionoverridesshapeinthiscase.Also,usingsyntheticmotionsthathaveexaggeratedproperties
of male and female walks may not produce the same effects as natural motions.
Johnson and Tassinary [2005] studied the effects of both shape and motion on the perception of
sex, gender, and attractiveness. Instead of using point light walkers which give little body shape in-
formation, they used silhouettes of human body shapes with varying waist to hip ratios, from ex-
aggerated hourglass to tubular figures. Even though synthetic motion was restricted to two highly
informative parameters (i.e., swagger for men and hip sway for women), they found that the shape
of the character was more informative of the sex of the individual than its motion. They recently fol-
lowed this up with a study in which they showed that both form and motion information, using these
exaggerated feminine and masculine cues, contributed to participants’ judgments of attractiveness
[Johnson and Tassinary 2007].
In summary, previous work has shown that, when using exaggerated synthetic motion on point light
walkers of differing shapes, motion was more dominant than shape information in sex perception.
Conversely, when using synthetic motion on exaggerated male and female silhouettes, morphology was
moreinfluentialthanmotion.Inthispaper,wedescribetwoexperimentswherenaturalfemaleandmale
motions, as well as synthetic neutral motions, were applied to very realistic male and female models
(e.g., as may be found in modern computer graphics applications). In Experiment 1, we directly compare
results with an androgynous shape and point light walker and also examine the effects of incongruent
shape and motion combinations. In Experiment 2, we further investigate the effect of different male
and female shapes in order to determine whether a scale of effect is evident or whether information
about body shape interacts with motion information in the perception of the sex of the model.
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 4

20:4
•
R. McDonnell et al.
Fig. 3. Group participating in experiment.
3.
Forty-one participants (22M, 19F, aged 20 to 45) took part in this experiment to determine whether
the motion or the shape/form of a virtual character is more salient for judgements of perceived sex of
the model. All participants were na¨ ıve to the purpose of the experiment and had normal or corrected to
normal vision. Participants (consisting of staff, students, and professionals from different educational
backgrounds) viewed the stimuli on a large projected screen in a lecture theatre (see Figure 3). We used
a two-way, repeated measures design and the conditions were motion type (3) and model type (4).
EXPERIMENT 1
3.1
Six undergraduate students (3M, 3F) volunteered to be motion captured, each in a separate session per
actor. Since the walks were motion captured and, therefore, real we chose our volunteers randomly and
did not try to ensure that their walks appeared particularly male or female. They were not informed as
to the purpose of the motion capture session which was conducted using our 10 camera Vicon optical
system, using 41 markers placed according to the Vicon human template. A curved path was drawn on
the floor in the capture area and they were asked to walk naturally up and down this path, which they
didabout20times.Wecapturedsomeofthewalkswithouttheirknowledgetoensuretheywerewalking
naturally, then applied the motion capture data to characters in 3D Studio Max and kept one natural
walk per actor (see video at http://isg.cs.tcd.ie/mcdonner/TAP/). Since it would not be possible to
capture a truly neutral walk, we synthetically generated three different neutral walk motions, using 3D
Studio Max’s footstep modifier. This modifier allows the automatic generation of synthetic walk motions
using a kinematic solver. These walks are neutral as they have neither male nor female characteristics
(such as hip sway or shoulder movement). The distance between footplants was altered in order to
create three unique neutral walks.
Motion Type
3.2
Four different models were used to display the different motions (Figure 2): highly detailed woman
and man models of approximately 35000 polygons each, an androgynous character, and a point light
walker. The woman and man were chosen as typical characters that would be used in a computer
simulation of natural crowds. The androgynous figure was chosen as it did not appear particularly
male or female in shape and so could serve as a control. The point light walker was generated from a
generic neutral skeleton and so contained minimal shape information. We included this representation
both as a baseline and to compare with the results of the earlier studies described above.
Model Type
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 5

Evaluating the Effect of Motion and Body Shape on the Perceived Sex of Virtual Characters
•
20:5
Fig. 4. (L) Walker’s path, (R) Camera view.
A grey ground plane was added to the scene and the camera was placed so that the majority of the
walk was in a straight line coming towards the camera (Figure 4), since Halevina and Troje found the
frontal view facilitates sex classification [2007]. Each of the different motion types (3) from each of the
actors (3) were applied to each of the model types (4), with two repetitions for each condition resulting
in a total of 72 movies, each presented for a duration of approximately 3.5 seconds long (the length of
time it took the actor to walk from one end of the motion capture area to the other).
The movies were sorted randomly in two different playlists and two groups of participants each
viewed one of the sets of 72 movies on a large projected display1(Figure 3). They were told to take
both motion and form/shape into account when making judgements about the sex of the character and
they marked their selections on an answer sheet. Participants sat near to one another, but were not
allowed to discuss their decisions. Each movie was followed by a four-second blank screen, during which
participants categorised the character they just saw on a five-point scale of 1: very male, 2: male, 3:
ambiguous, 4: female, or 5: very female. The number of the next stimulus was then displayed for two
seconds with an accompanying alerting sound.
3.3
Before beginning the experiment, participants were shown a static image of the androgynous figure and
asked to rate it on the same five-point scale. As it was considered to be ambiguous by almost all partici-
pantsweconcludedthat,basedonshape/forminformationalone,thefiguretrulyappearedandrogynous.
We averaged participants’ ratings over the two repetitions and three actors for each combination of
walk and model type. We first found that the sex of the participants had no effect, as in [Kozlowski and
Cutting 1977; Johnson and Tassinary 2005]. A two factor ANalysis Of VAriance (ANOVA) [Howell 1999]
with repeated measures showed a main effect of walk type (F2,80= 130, p < 0.001). Post-hoc analysis
was then performed using a standard Newman-Keuls test for pairwise comparisons among means. In
Figure 5 (left) it can be seen that, as expected, female walks were rated overall as female, neutral walks
as neutral, and male walks as male (p < 0.001 in all cases). There was also a main effect of model type
(F3,120= 24, p < 0.001). Figure 5 (R) shows that the man was considered significantly more male than
the woman or the androgynous figure, while the point walker was considered significantly less male
than the man and less female than the woman or the androgynous figure (p < 0.001 in all cases). There
was also an interaction between model type and walk type (F6,240= 25, p < 0.001).
Figure 6 shows that male walks on the woman affect the sex judgement of the female form such that
it is now rated as ambiguous. A similar effect is observed when a female walk is applied to a male shape.
Results
1Similar group experiments were conducted with only a monitor at the front of the lecture theatre [Kozlowski and Cutting 1977].
Even though viewing angle differed significantly between participants, recognition performance was still well above chance. Our
larger display ensured more consistent viewing angles.
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 6

20:6
•
R. McDonnell et al.
Fig. 5. Main effects of (L) walk type (averaged over all models), (R) model type (averaged over all motions)–with standard errors.
Fig. 6. The interacting effects of model and walk type.
This implies that applying motion captured from actors of the opposite sex to the character will produce
confusing or unsatisfactory results in general. Interestingly, neutral walks were considered male when
viewed on the man and female when viewed on the woman. This implies that for neutral walks, the
shape of the character takes precedence over the motion in determining the sex of the character. This
result has implications for computer graphics applications where resources are limited, as reusing the
same neutral walks on male and female characters would appear to produce the desired effect.
As illustrated in Figure 6, there were three significantly different ratings for the androgynous figure
for male, neutral, and female walks (p < 0.001). This implies that for a character with androgynous
appearance, the motion information is most important when determining the sex (as without motion,
the androgynous figure was consistently rated to be ambiguous). Replicating the results of previous
studies, we showed that participants were able to determine the sex of the walker using point light
walkers with good accuracy.
We then looked more closely at the ratings for each of the motion captured walks. Female walks 1 and
2 (F1, F2) were consistently rated as female for all model types (see Figures 7 and 8). However, F3 was
consistently rated as either ambiguous (for the woman and the androgynous figure) or slightly male (for
the man and the point walker). We noticed that the female actor that performed F3 did not move her
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 7

Evaluating the Effect of Motion and Body Shape on the Perceived Sex of Virtual Characters
•
20:7
Fig. 7.Classifications of individual walkers for each model type (Female walks F1-F3, Males M1-M3, Neutral N1-N3).
Fig. 8.Classifications of individual walkers for each model type (Female walks F1-F3, Males M1-M3, Neutral N1-N3).
hips much, but swung her arms quite a lot and her natural walk was slightly faster than the other two
females. This result implies that care should be taken when choosing an actor, as misinterpretations
can occur. This was also the case in [Kozlowski and Cutting 1977] where they found that one of their
female walks was more difficult to rate than the others.
The three neutral walks were ranked similarly to each other for each of the model types. Also, a
similar trend across the three different male walks was found for each model type, with M1 considered
significantly more male than the other two male walks (p < 0.001 in all cases). For the man, the motion
of M3 was considered more masculine than M2. The actor that performed M2 did appear to have more
hip sway than the other actors when we were capturing the session, therefore it is unsurprising that
his walk was considered less masculine. However, it is interesting that this was most evident on the
male model.
In order to judge overall sex recognition accuracy for the motions and to see if our results are compat-
ible with earlier point-light walker experiments, such as Kozlowski and Cutting’s [1977], we converted
our scaled judgements of sex into percentage correct values (with 4 or 5 being female, 1 or 2 male,
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 8

20:8
•
R. McDonnell et al.
Fig. 9.
three female walks.
Identification accuracy for recognition of walker sex (top) overall; (middle) for the three male walks; (bottom) for the
and 3 neutral). The results are shown in Figure 9. For the point walker, 69% of male walks were
correctly identified, which is consistent with the earlier study where it was found that recognition ac-
curacy hovered around 70%. However, a significantly lower percentage (58%) of females were correctly
identified than males (F82,1= 4.77, p < 0.05). This can be explained by the fact that F3 was consis-
tently misclassified as male and had significantly different results from the other female walks for all
models.
Noteworthy too is the fact that the accuracy for female walks on the male character is equally poor to
that for male walks on the female character (both below chance), while the congruent motion on both
is rated equally high at about 65%. Particularly interesting too are the results for the neutral walks,
where well under 50% of them were actually classified as neutral for any of the models. In particular,
only about 25% of the neutral walks on the realistic characters were perceived to be ambiguous. Closer
examination reveals that 62% of neutral walks were classified as female on the woman, but only 11%
as male (F82,1= 14.22, p < 0.0005), with an equal and opposite effect on the man, with 65% classified
as male and 13% as female (F82,1= 8.19, p < 0.005).
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 9

Evaluating the Effect of Motion and Body Shape on the Perceived Sex of Virtual Characters
•
20:9
Fig. 10.Six body shapes used in Experiment 2: Man 3, Man 2, Man 1, Woman 1, Woman 2, Woman 3
4.
In Experiment 1, we chose only two models to represent male and female appearance. In this experi-
ment,weinvestigatedtheeffectofmotionondifferentmaleandfemaleshapestodetermineifadditional
indicators of sex in the shape will change the overall perception of the sex of the character.
EXPERIMENT 2
4.1
We used the same man and woman models as in the previous experiment. However, this time they both
wore neutral clothing (jeans and a grey t-shirt). Using 3D Studio Max, we manipulated the polygons of
the woman model in order to create a very exaggerated female body shape. We increased the size of her
breasts and hips, in order to give her an hour glass figure with a large waist-to-hip ratio (Woman 3).
We applied equivalent alterations to the man model by increasing his shoulder, waist, and neck width
to create a more exaggerated male body shape (Man 3).
Next, we created a more androgynous version of the female by decreasing her waist-to-hip ratio and
breast size (Woman 1), and of the man by decreasing his chest, neck, and waist size and the distance
between his shoulders (Man 1). Finally, we created a morphed female body shape (Woman 2), which
was a half way morph between Woman 1 and Woman 3. Similarly, we created a halfway morph between
Man 1 and Man 3 (Man 2)—see Figure 10. All bodies were created based on our own estimates of
physiological plausibility which were consistent with standard physical proportions. In participant
feedback, no comments regarding physical abnormalities were made.
Body Shapes
4.2
Each of the nine different motion types from the previous experiment (three male, three neutral, and
three female) were applied to each of the body shapes (6), with two repetitions for each condition,
resulting in a total of 108 movies. Three random playlists, each including the 108 movies, were created
and three groups of participants each viewed one of the playlists on a large projected display. Thirty-
eight participants took part in this experiment (33M, 5F). All participants were na¨ive to the purpose of
theexperimentandhadnormalorcorrectedtonormalvision.Participantsconsistedmainlyofcomputer
science staff and students. As before, they were told to take both motion and form into account, and
they categorised the character on a five-point scale which ranged from very male to very female.
Procedure
4.3
As before, we averaged participants’ ratings over the two repetitions and three actors for each combi-
nation of walk and body shapes. A two factor ANOVA with repeated measures showed a main effect of
walk type (F2,74= 94.65, p < 0.001)—see Figure 11(L). Standard Newman-Keuls tests for comparisons
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.
Results

Page 10

20:10
•
R. McDonnell et al.
Fig. 11.Main effects of (L) walk type (averaged over all models), (R) body shape (averaged over all motions).
Fig. 12.The interacting effects of shape and walk type.
among means showed us that, as before, female walks were rated overall as female, neutral walks as
neutral, and male walks as male (p < 0.001 in all cases). There was also a main effect of body shape
(F5,185 = 39.73, p < 0.001). Figure 11 (R) shows that the three man shapes were considered signifi-
cantly more male than the three woman shapes (p < 0.001 in all cases). Surprisingly, there was no
statistical significance between participants’ ratings of Man 1, Man 2, or Man 3. However, there was
a difference between ratings of Woman 1 and Woman 2, and Woman 1 and Woman 3, with Woman 1
being rated as less female in both cases (p < 0.04 and p < 0.004 resp.). There was no difference in
perception between Woman 2 and Woman 3.
There was also a significant interaction (F10,370 = 10.28, p < 0.001) between walk type and body
shape (see Figure 12). Therefore, a post-hoc analysis was conducted using Newman-Keuls tests for
comparison among means.
Perhaps the most obvious result that may have contributed to this interaction effect is that the male
walks did not affect the perception of sex in the female shapes to the same extent that the female
walk affected the perception of the sex of the male models. For all the male models, applying a female
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 11

Evaluating the Effect of Motion and Body Shape on the Perceived Sex of Virtual Characters
•
20:11
walk biased the perception of the sex of the shape of the male. But this did not occur for two-thirds
of the female shapes. This suggests that female form is slightly more salient than male form in sex
judgements (although a male walk clearly reduced the number of ’female’ judgements of all the female
shapes, so motion was still a powerful cue).
We found that for all walk types, body shapes Man 3 and Man 2 were not significantly different from
each other. This implies that the difference between these two shape types had no effect on judgements
of sex, irrespective of motion type. Perhaps it was the case that Man 2 had already sufficient indicators
of sex that the extra information for Man 3 did not change our perception. Sex judgements for Man 1
were significantly different from both Man 2 and Man 3 for male and neutral walks but not for female
walks.
Woman 1 was considered less female than both Woman 2 and Woman 3 for all walk types. Woman 2
was considered less female than Woman 3 for only female walks. This implies that it was only for the
female walks that the extra female shape characteristics had an effect. Perhaps this was due to the hip
sway in the female walks being exaggerated by the increased waist-to-hip ratio.
As in Experiment 1, male walks on the Woman body shapes rendered sex judgements as more am-
biguous, as did the female walks on the Man body shapes. Neutral walks were still considered female on
the Woman characters, but in this experiment we can see that they are rated significantly less female
than the female walks. Similarly, the neutral walks on the Man characters were considered male, but
again less male than the male walks. This is an interesting result, as it shows that neutral walks could
be used for congruency, as the shape of the character is dominant, but there is a point at which extra
indicators of sex in the shape have no effect.
In Experiment 1, we found no difference between the ratings of motion captured male and neutral
walks for the man model. Similarly for the woman model, there was no difference in ratings between
the female and neutral walks. Perhaps this was due to the clothing worn by the models in Experiment
1 having extra indicators of sex than the neutral clothing in Experiment 2. More likely is the case that,
in Experiment 1 participants used the same rating scale as in Experiment 2 to cover a wider range of
models. There were no androgynous characters in Experiment 2 as this experiment focused more on the
properties of the man and woman models. The presence of these characters in Experiment 1 could have
influenced the participants to rate the neutral walks on the man and woman as more male or female
than they might have otherwise. We feel that the results for the neutral motions on man and woman
models in Experiment 2 could be considered more accurate, since they could not have been influenced
by other model types.
Finally, we looked more closely at the ratings for each of the motion captured walks (see Figure 13).
Female walks 1 and 2 (F1, F2) were consistently rated as female for all body shapes. In Experiment 1,
F3 was consistently rated as either ambiguous or slightly male. The same was true in this experiment,
except that for the three Woman characters, increasingly more ’female’ judgements were made, the
more female the body shape. This suggests that perhaps the female characteristics of the walk were
brought out by the exaggerated female shape.
As before, the three neutral walks were ranked similarly to each other for each of the model types.
Similar effects were also found for two of the different male walks (M2 and M3) for each body shape.
Interesting to note is how M1 was considered significantly less male on Man 1 than on Man 2 or Man
3 (p < 0.001). M1 was the walk judged as the most male in the previous experiment so it is interesting
how the least male shape influenced the overall perception of the character to appear less male.
In this experiment, we found that both shape and motion had an effect on our perception of sex. For
the three Woman body shapes, the female walks were judged as increasingly female the more female
the shapes became. The neutral walks were affected by shape also, but Woman 2 and 3 were judged as
equally female, implying that more exaggerated female shapes had no effect on the perceived sex when
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 12

20:12
•
R. McDonnell et al.
Fig. 13.Average classifications of individual walkers on body shape.
coupled with neutral walks. For the Man models, an increase in male shape between Man 1 and 2 (but
not between 2 and 3) increased the number of male judgements when coupled with neutral and male
walks, implying that extra male information on the Man model did not have an effect past Man 2.
This plateau for Man 2 for male and neutral walks may have been due to the fact that a large
percentage of participants were male in this experiment and perhaps had more sensitivity towards
female body shapes. To test this theory, we ran the experiment on a further 12 females and combined
these results with the original five females, resulting in a total of 17 female participants. We randomly
chose the results of 17 of the male participants and compared them to the female results using an
ANOVA, where the repeated measure factors were model type and motion type, and between subjects
factor was sex of participant. We found no main effect of the sex of the participant. We also found
no interaction between participant sex and walk type. However, we did find a significant interaction
(p < 0.001) between sex of participant and model type. On closer inspection, this interaction was due
to the man models being rated as slightly more male overall and the woman models as slightly more
female by the female participants. However, there were no differences between the ratings of the female
and male participants for the three man or the three woman models. This indicated that the plateau
at Man 2 for male and neutral walks may have been due to some other factor such as the model that
we chose. What this indicates is that, despite [Kozlowski and Cutting 1977; Johnson and Tassinary
2005] and the results of Experiment 1, there is some evidence that there may be slight differences in
perception of body types between males and females.
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 13

Evaluating the Effect of Motion and Body Shape on the Perceived Sex of Virtual Characters
•
20:13
5.
In this paper we have identified some of the factors that influence perception of the sex of virtual
characters. Unlike previous work, we found that both the appearance and the motion of the character
influence perceived sex. We also found that for walk motions with no male or female characteristics, it
is the appearance of the character that will dominate our perception of the character’s sex, whereas it
is the motion that dominates our perception for characters with an androgynous appearance.
This information will be useful in computer graphics applications, as it is now clear that in order
to have a convincing male/female character, you must apply a female walk to a female character and
a male walk to a male character. This implies that consistency is highly important when designing a
virtual character. As discussed in [Vinayagamoorthy et al. 2006], people expect virtual characters to
behave in a manner befitting its appearance, and the results of our study confirm this. However, care
must be taken when choosing an actor, as we found some misinterpretations. From our results, it is
certain that you cannot apply a female walk to a male character, or vice versa, as viewers will judge
this conflicting information to be ambiguous.
A simple solution would be to reduce memory costs by using neutral walks, as overall they will look
female on a female character, and male on a male character, but less so than the motion capture. In
Experiment 2, we found that adding additional indicators of sex in the shape of the character increases
howmaleorfemaleaneutralwalkisconsidered,butonlyuptoapoint.Also,wedonotknowhownatural
the participants found these animations and would like to investigate in future studies whether or not
they found these motions to be natural (we have some implicit evidence that they did, however, as they
had the option to categorise these walks as ambiguous, which they did for the other models). We may
need to add stereotypical characteristics such as hip sway and shoulder movement to increase motion
believability [Cutting 1978]. Therefore, a comparison of synthetic and natural walk motions could also
be beneficial. Something akin to a “Turing test” could be appropriate here, where people would indicate
whether a walking motion was real or synthetic.
It would be interesting to see if our results are replicated with other motions besides walking. Per-
haps it is also the case for gestures; that male actors performing certain stereotypical gestures (e.g.,
woodcutting or boxing) will always be considered male. However, this may not be the case for unfamiliar
motions such as those typically used in games. Perhaps it would be possible to reuse unfamiliar motions
(such as shooting and kicking) from single actors on multiple game characters (both male and female).
Furthermore, highly dynamic motions have less freedom for style than walking (because they are closer
to the energy limit of the human body). Therefore, such motions may naturally be more neutral. How-
ever, the dynamics of the actor will play more of a role in such actions, so the properties of male/female
style that are linked to the kinematics and dynamics of the body (e.g., wider hip spacing, smaller size,
lower moment of inertia in females) could actually cause the differences to be more obvious in dynamic
motions. Such issues merit further investigation.
CONCLUSIONS
REFERENCES
BEARDSWORTH, T. AND BUCKNER, T.
seeing ones body. Bulletin of the Psychonomic Society 18, 1, 19–22.
CUTTING, J.1978. Generation of synthetic male and female walkers through manipulation of a biomechanical invariant.
Perception 7, 4, 393–403.
CUTTING, J. AND KOZLOWSKI, L.1977.Recognizing friends by their walk: Gait perception without familiarity cues. Bulletin of
the Psychonomic Society 9, 5, 353–356.
DOBBYN, S., MCDONNELL, R., KAVAN, L., COLLINS, S., AND O’SULLIVAN, C.
with variation. In Eurographics Short Papers. 103–106.
HALEVINA,A. ANDTROJE,N.2007.Sex-classificationofpoint-lightwalkers:Viewpoint,structure,kinematics. InPosterPresented
at Vision Science Society Meeting, Sarasota, FL.
1981.The ability to recognize oneself from a video recording of ones movements without
2006.Clothing the masses: Real-time clothed crowds
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.

Page 14

20:14
•
R. McDonnell et al.
HOWELL, D.
JOHANSSON, G.
201–211.
JOHANSSON, G.
379–393.
JOHNSON, K. L. AND TASSINARY, L. G.
logical Science 16, 11, 890–897.
JOHNSON, K. L. AND TASSINARY, L. G.
Proceedings of the National Academy of Sciences 104, 12, 5246–5251.
KOZLOWSKI, L. AND CUTTING, J.1977.
Psychophysics 21, 578–580.
KOZLOWSKI, L. AND CUTTING, J. 1978.
thoughts. Perception and Psychophysics 23, 459.
MATHER, G. AND MURDOCH, L.1994.
of the Royal Society of London, Series B 258, 273–279.
TROJE, N. F.2002.Decomposing biological motion: A framework for analysis and synthesis of human gait patterns. Journal
of Vision 2, 5, 371–387.
VINAYAGAMOORTHY, V., GILLIES, M., STEED, A., TANGUY, E., PAN, X., LOSCOS, C., AND SLATER, M.
virtual characters. Eurographics State of the Art Reports.
1999.
1973.
Fundamental Statistics for the Behavioral Sciences. Duxbury Press.
Visual perception of biological motion and a model for its analysis. Perception and Psychophysics 14, 2,
1976. Spatio-temporal differentiation and integration in visual motion perception. Psychological Research 38,
2005.Perceiving sex directly and indirectly: Meaning in motion and morphology. Psycho-
2007.Compatibility of basic social perceptions determines perceived attractiveness.
Recognizing the sex of a walker from a dynamic point-light display. Perception and
Recognizing the gender of walkers from point-lights mounted on ankles: Some second
Gender discrimination in biological motion displays based on dynamic cues. Proceedings
2006. Building expression into
Received December 2007; revised April 2008; accepted June 2008
ACM Transactions on Applied Perception, Vol. 5, No. 4, Article 20, Publication date: January 2009.