Conference PaperPDF Available

Shape Up! Perception based body shape variation for data-driven crowds

Authors:

Abstract and Figures

Representative distribution of body shapes is needed when simulating crowds in real-world situations, e.g., for city or event planning. Visual realism and plausibility are often also required for visualization purposes, while these are the top criteria for crowds in entertainment applications such as games and movie production. Therefore, achieving representative and visually plausible body-shape variation while optimizing available resources is an important goal. We present a data driven approach to generating and selecting models with varied body shapes, based on body measurement and demographic data from the CAESAR anthropometric database. We conducted an online perceptual study to explore the relationship between body shape, distinctiveness and attractiveness for bodies close to the median height and girth. We found that the most salient body differences are in size and upper-lower body ratios, in particular with respect to shoulders, waist and hips. Based on these results, we propose strategies for body shape selection and distribution that we have validated with a lab-based perceptual study. Finally, we demonstrate our results in a data-driven crowd system with perceptually plausible and varied body shape distribution
Content may be subject to copyright.
Shape Up! Perception based body shape variation
for data-driven crowds
Yinxuan Shi, Jan Ondˇ
rej, He Wangand Carol O’Sullivan
Ohio State University, Email: shi.217@osu.edu
Trinity College Dublin, Email: Jan.Ondrej@scss.tcd.ie, Carol.OSullivan@scss.tcd.ie
University of Leeds, Email: H.E.Wang@leeds.ac.uk
Abstract—Representative distribution of body shapes is needed
when simulating crowds in real-world situations, e.g., for city
or event planning. Visual realism and plausibility are often
also required for visualization purposes, while these are the
top criteria for crowds in entertainment applications such as
games and movie production. Therefore, achieving representative
and visually plausible body-shape variation while optimizing
available resources is an important goal. We present a data-
driven approach to generating and selecting models with varied
body shapes, based on body measurement and demographic data
from the CAESAR anthropometric database. We conducted an
online perceptual study to explore the relationship between body
shape, distinctiveness and attractiveness for bodies close to the
median height and girth. We found that the most salient body
differences are in size and upper-lower body ratios, in particular
with respect to shoulders, waist and hips. Based on these results,
we propose strategies for body shape selection and distribution
that we have validated with a lab-based perceptual study. Finally,
we demonstrate our results in a data-driven crowd system with
perceptually plausible and varied body shape distribution.
I. INTRODUCTION
Simulation of crowds is needed for many applications, which
tend to fall into two main categories: those that need as accurate
a representation of a population as possible, and those for
which visual realism and plausibility are the most important
criteria. For example, for games and movies, the most important
criterion is usually that the crowd is perceived to be realistic and
to behave plausibly, whereas when designing a city, stadium
or theme park, the planners will need to be confident that
the movements, behaviors and body sizes of the agents are
representative of the real humans who will actually populate the
future environment. However, for many planning applications a
realistic visual representation of the crowd is also important as it
helps to better understand the real-world data being visualized;
and often for games and movies, the more representative the
virtual crowd is of a real population, the more realistic and
plausible the results will be. Therefore, a data-driven approach
that takes real-world demographics into account will be useful
in both cases.
One of the most important factors for perceived realism of
crowds is the level of variety in the 3D models used to visualize
the virtual humans present. The perception of varied human
appearance and motion have previously been studied in this
Note: This work was done while all authors were at Disney Research
context [1], [2], but to date, body shape variation has received
less attention. However, two of the most functionally relevant
and visually salient features of any crowd are the shapes and
sizes of the people in it, and their relative distributions: i.e.,
under normal circumstances most people will have shapes close
to the median of the population, and there will be decreasing
numbers of more atypical bodies present. For the majority of
practical applications, it is not possible to individually model
each member of a crowd, so a fixed budget of 3D models
is used and replicated to generate a large crowd. Therefore,
strategies are needed for choosing the optimal set of characters
and distributing them realistically, while minimizing the risk
of perceiving visual clones.
Approaches to varying body shapes for crowd simulation
have been presented, and our work is very complementary
to such approaches. For example, Thalmann and Musse
[3] generate models with different body types that may be
distributed according to any available statistics. However, the
resulting models in such systems have tended to be more
stylistic representations rather than physically accurate, and
the perception of body shape differences or motion variation
has not been considered. In the field of psychology, body
perception studies have mainly been focused on gender and
attractiveness (see [4] for a recent overview) and not on the
distinctiveness of different body shapes. To our knowledge,
this is the first study to perceptually evaluate the effects of
body shape on distinctiveness and the first to use body shapes
derived from real data to explore attractiveness.
In this paper, we present a data-driven approach to creating
a crowd with representative and varied body shapes. We
base our model generation and distribution strategies on
body measurement and demographic data from the Civilian
American and European Surface Anthropometry Resource
(CAESAR) anthropometric database. From preliminary studies
and observation, it was obvious that the most salient body
shape differences are due to the height and girth (i.e., waist
circumference) of the captured actors. However, closer to the
median, other types of body shape differences are increasingly
salient. As there are more individuals in this group than any
other, more replications of each template model will be needed.
We were interested in the following question and conducted
two perceptual studies to explore it: What are the most salient
factors that affect the perception of body shape distinctiveness
Fig. 1. A set of varied models created using our data-driven body shape selection and creation approach.
Fig. 2. Example pairs showing 3D models (right) generated from the CAESAR sample measurements (left).
for bodies close to the median height and girth?. We found that
the most salient body differences are in size and upper-lower
body ratios. Based on our results we strategically select of a
number of models which, with clothing variation, can be used
to optimize crowd variety with a limited budget of 3D models.
II. RE LATE D WOR K
McDonnell et al. [1], [5] have studied the effects of different
types of appearance and motion variation on the perceptibility
of visual clones in crowds, while Pra
ˇ
z
´
ak et al. [2] demonstrated
how only three different human motions replicated evenly
through a crowd. Hoyet et al. [6] evaluated the distinctiveness
and attractiveness of different types of human motion and
found that average motions were the least distinctive and
most attractive, consistent with findings in face perception
[7]. Johnson and Tassinary [8], [9] studied the effects of
both shape and motion on the perception of sex, gender, and
attractiveness. Using mainly stylized silhouettes of human body
shapes with varying waist to hip ratios, from exaggerated
hourglass to tubular, they found that both shape (especially
waist to hip ratio for women, and and motion information
contributed to participants judgments of attractiveness. In
particular, the waist-to-hip ratio (WHR) and hip sway are
important for sex categorization and female attractiveness, while
shoulders and their sway are most salient in men. Wellerdiek
et.al. [10] recently explored how body shapes and postures
affect perceived strength and power of male characters. See [4]
for a thorough overview of recent research in body perception.
apart from sex categorization, we are not aware of any studies
where the distinctiveness of different body shapes has been
explored.
Realistic body shape generation usually starts from real data,
e.g. body scans [11] and measurements [12]. To generate a
variety of body shapes, a body shape manifold or statistic model
is typically learned from shape data [13]–[19]. Many assume
a low dimensional manifold and dimensionality reduction tech-
niques such as Principle Component Analysis (PCA) have been
used to learn such representations. Once the manifold is learned,
controlled body generation can be achieved by specifying part
of the parameters and inferring the remainder. The parameter
specification can be either done through sampling or user input
[17], [20]. We use a commercial system (Daz3D) to generate
our crowd characters, but our results would be applicable
to models generated by any of these methods. With respect
to crowd generation, our work is complementary to that of
Thalmann and Musse [3], who generate models with different
somatotypes (endomorph, mesomorph and ectomorph) which
vary based mainly on the distribution of body fat, muscle and
WHR. Then, these body types may be distributed according
to any available statistics. Our model selection strategies and
perceptual results could be used to improve the realism and
representativeness of such a system.
III. MODEL SELECTION AND CREATION
To generate a set of 3D virtual human models for our data-
driven crowd system, we need to select a representative sample
of real body shape measurements and then create a realistic
visual representation for each. Using a set of templates to be
repeated throughout the crowd is efficient in terms of both
Fig. 3. The CAESAR database distribution of body heights and girths, with red dots showing the samples we picked to generate our template models; the red
cluster in the median group shows the samples used to generate our experimental stimuli. The total number of samples within each quadrant is shown at its top
left corner.
artist time and computational resources, as hardware instancing
can be used to replicate selected models.
We use body measurement data sourced from the CAESAR
project. The CAESAR data is based on stratified sampling of
age, ethnicity and gender [21]. It contains demographic data,
3D range scan and measurements for 2391 US residents (52.9%
female, 47.1% male, aged 18-65, with varied ethnicity). Height
and girth (waist circumference) are used to pick reference
samples, as we observe that these are by far the most salient
features that differentiate people from each other, especially
the more they deviate from the median. Then, we use 29
measurements for each sample (see Table I) to direct the
creation of each template 3D human model. To represent
the variation of the height and girth distribution as shown
in Figure 3, we divide the height into 5 groups and girth into 7
groups respectively. We combine some of the outlier quadrants
with small numbers of samples. Then 3 samples are drawn
randomly from each quadrant at least
r
distance apart to avoid
resemblance. After careful examination, one sample is kept to
represent each bucket and is used as templates for generating
the crowd. Due to the large sample count around the median
(row 2, column 2), more distinctive samples are needed to
maximize variety.
We use Daz3D to create realistic 3D models as it provides
useful tools and morphs for creating body shapes, although any
model generation tool with similar properties could be used. A
set of mesh deformers
P
is used to morph the template mesh
T
to the desired measurement
Md
. Thus we can generate models
with measurements that closely match the CAESAR samples
(see Figure 2). Note that some measurements require the avatar
to be seated and in these cases we take the measurement from
the template mesh by setting the joint rotation of the avatars.
We wrote a script to apply the Daz3D mesh deformers
P=
{p1,p2,··· ,pn}
to morph the template mesh
T
using weights
wi
. The post morph measurement
Mt
of the model is shown
in 1. Our script adjusts the morphs iteratively and different
measurements are recorded and compared against the desired
measurements.
Mt=Measure(T+
n
i=0
wipi(T)) (1)
To conform the avatar to the CAESAR subject measurements,
we use conjugate gradient descent to minimize the error
between the template mesh measurements and the desired
ones:
ε=kMdMtk
. A set of clothes for the template female
and male meshes were created by an artist, and morphed
in Daz3D to fit each of our selected models. The models
are exported to Morpheme where retargeting and animation
occurs, and finally they are imported into our crowd system.
There they are rendered with color variation to create our final
varied crowd simulation.
IV. EXP ER IM EN TS
We conducted an online perceptual experiment to investigate
the influence of body measurements on the perception of body
shapes that are close to the median, as these are the templates
that will be most often replicated in a crowd. We are particularly
interested in the distinctiveness of the different body shapes,
and as previous research shows how attractiveness is closely
related, we also explore this metric. We also ran a laboratory-
based experiment with a different method and smaller number
of participants to gain further insights.
Stimuli:
We select 32 male and female CAESAR samples
with girth and height closest to the median (see Figure 3) and
generate the 3D models that match their measurements, as
described in Section III. We also generate average male and
female models, as previous research in face and body motion
Fig. 4. Example models used in Experiment 1, sampled near the median:
models picked to represent the median group (left); most distinctive and
most/least attractive models (right).
perception has shown that the average is usually amongst the
least distinctive and most attractive. Some of the models used
can be seen in Figure 4.
Method:
In order to recruit as many participants as possible,
we created several online surveys and posted the links on
Mechanical Turk (MT). The HITs (Human Intelligence Tasks)
were available only to ‘master MTurkers’, i.e., those who have
a good performance track record. The study was approved
by an institutional ethics review board and all participants
provided informed consent. After first removing the responses
of participants whose accuracy was too low (97 out of 479),
we analyzed the results of 382 participants (206F/176M).
Participants are shown a set of 99 pages, showing either all
male or female models depending on the hit. A target model
is shown on the left, and three smaller sample models are
shown on the right. One of the sample models always matched
the target, while the other two distractor models are chosen
at random from the remaining 32. The task was to first rate
the attractiveness of the target on a 7-point Likert scale from
1 (very unattractive) to 7 (very attractive); then to select the
sample on the right that was the same as the target. Three
repetitions of each target were shown, and the order of all
pages was randomized. Each page was viewed for between
5-10 seconds so the experiment took approximately 15-20
minutes on average.
As it would be infeasible to generate all possible combina-
tions of the 33 models as the time and/or number of participants
needed would be prohibitive, we created two surveys each for
male and female models. This means that each target was
compared with a total of 12 distractor models. To ensure that
Fig. 5. Distinctiveness and Attractiveness results from Experiment 1: the red
dots show the models we chose for the median group. The average Male and
Female models were also chosen.
we were not introducing bias with this limitation, we also ran a
laboratory-based experiment with 17 participants (8M/9F). We
followed a similar procedure to Hoyet et al. [6], where first a
group of two side-by-side models was shown, then one target
model, and the task was to indicate using one of two keys
whether the target was Present or Absent from the previous
group. In 50% of cases the model was present. One group of
9 participants viewed the Females and one group of 8 viewed
the Males. There were 4 repetitions of each target model (2
present, 2 absent) and the distractor models were selected at
random.
V. RE SU LTS AND DISCUSSION
Analysis of Variance (ANOVA):
To ensure that differ-
ences between models were indeed being noticed, we first
performed repeated measures ANOVA on the average accuracy
values and attractiveness ratings for the Mechanical Turk
experiment. For accuracy (i.e., mean percentage of correct
identifications), we found a main effect of Model for the
33 female models (
F(32,7680) = 50.06,p< .00005)
) and the
33 male models (
F(32,7520) = 30.65,p< .00005
). For the
attractiveness ratings, we also found a main effect of Model
for both Females (
F(32,7680) = 134.01,p< .00005)
) and
Males
F(32,7520) = 136.83,p< .00005)
. We also analyzed
the factors of participant sex, age group and the self-reported
display device used and found no significant effects.
We also performed repeated measures Analysis of Variance
(ANOVA) on the accuracy results from the Laboratory exper-
iment for Males and Females separately (a between-groups
analysis showed no main effect of Model Sex). We found
a main effect of both Female (
F(32,256) = 1.88,p< .005)
)
and Male (
F(32,256) = 1.97,p< .00005)
) Models. We also
evaluated whether presence or absence detection accuracy were
different, and found that the former was significantly higher
than the latter (
F(1,8) = 25.70,p< .0005)
). However, a two-
way interaction between presence/absence and Model for the
Female models (
F(32,256) = 1.49,p< .05)
) indicated that
for some actors, absence detection was as good as presence
detection, in one case (90%) whereas for others it was as low
as 20%. There was much less variability in the scores for
presence detection, with participants scoring above 80% on
average.
Correlation Analysis:
One of the main goals of our
experiments was to determine which body shape features have
the strongest effect on body shape perception. Hence we wished
to assess the correlations between our results and the body
measurements that we used to create the 3D models. Based on
the observation that absence detection appeared to yield greater
variability between models in the Laboratory experiment, we
decided to use five metrics based on our results: the Mechanical
Turk Hit rate (MT-H), which was calculated for each model
as the percentage of times it was accurately matched; the
Mechanical Turk False alarm rate (MT-F), i.e., for each model,
we calculated the number of times it was wrongly selected
divided by the number of times it was seen as a distractor
(note that we could not do this on a per-participant basis, as
the distribution of distractors was random); the Laboratory
Hit (LAB-H) and False alarm (LAB-F) rates; and the average
Attractiveness (ATT) ratings.
From viewing the images of the ranked models, it was clear
that the women fell into one or other of two very visually
distinct groups: those with symmetric upper and lower bodies
(i.e., hour-glass) and those whose lower body was wider than
the top (i.e., pear-shaped). Further analysis of the male images
showed that the two main body types were V-shaped (i.e.,
shoulders larger than the lower body) and Block (upper and
lower body widths more or less similar). Furthermore, from
the cited literature we know that Waist to Hip Ratios (WHR)
in women and Waist to shoulder or Chest Ratios (WCR) in
men have been found to be strong predictors of Gender and
Attractiveness. In order to determine whether the Females were
more Hourglass or Pear-shaped, we also calculated WCR-WHR
for the females (i.e., indicates symmetry between the upper
and lower body). We also found high correlations with these
measurements and factors as shown in Table II. We focus on
the most salient features here, and include all other significant
correlations in the supplemental material.
We selected Hourglass and Pear female models from those
above and below the median of WCR-WHR; and VShape and
Block males similarly, based on the median WCR. In Figure
5 we plot the average attractiveness (ATT) ratings against the
average distinctiveness (MT-H) for these models. We can see
that for the Hourglass group, attractiveness and distinctiveness
are positively correlated (0.9), whereas for the Pear group they
are not (the only reason there is not a significant negative
correlation is due to an outlier pear model who was much
thinner and lighter than all the others). Similarly, for the
Fig. 6. Close-up from the crowd simulation.
VShape males, there is a significant positive correlation between
distinctiveness and attractiveness (0.8) and a negative one for
the Blocks (-0.8). The most distinctive hourglass female was
also the most attractive female overall, with the most distinctive
pear the least attractive, with the same being true for the male
VShape and Block models (see Figure 4:right).
Selection Strategy:
These results confirm our intuition that
these different body types are perceived very differently by
participants, and hence we made our choice of the three
representatives of the Median groups (Figure 4:left) by choosing
the average Male and Female (both of whom had the median
WCR or WCR-WHR and were amongst the least distinctive
of all models), along with one of each of the Hourglass, Pear,
VShape and Block models (chosen to be not overly distinctive
but reasonably different from each other). For the non-median
groups, as described in Section III, one sample is selected
among three candidates to represent its body size group. As
shown in Figure 1, we generated a set of crowd template for
various size and shapes to simulate the real-world crowd.
VI. CONCLUSION
We have found that the most visually salient properties of
body shape for models near the median are hips, chest and their
ratios with waist size. Furthermore, the types of bodies that
are defined by these ratios, i.e., V-shape or Block for males;
Hourglass or Pear for females, are perceived differently. The
template models that are selected using our perception based
strategy are most representative of the median population in a
crowd, which combined with the samples we selected from the
other body sizes, allow the generation of a crowd that is both
varied and representative of a real population (see Figure 6).
There are limitations to our work, in that we have only
explored similarity for a very small group of body shapes, and
we did not perform a full confusion analysis due to the nature
of the experimental data. We also did not assess the overall
variation perception of the full crowd, which is an interesting
direction for future work.
REFERENCES
[1]
R. McDonnell, M. Larkin, S. Dobbyn, S. Collins, and C. O’Sullivan,
“Clone attack! perception of crowd variety,” ACM Trans. Graph.,
vol. 27, no. 3, pp. 26:1–26:8, Aug. 2008. [Online]. Available:
http://doi.acm.org/10.1145/1360612.1360625
[2]
M. Pra
ˇ
z
´
ak and C. O’Sullivan, “Perceiving human motion variety,
in Proceedings of the ACM SIGGRAPH Symposium on Applied
Perception in Graphics and Visualization, ser. APGV ’11. New
York, NY, USA: ACM, 2011, pp. 87–92. [Online]. Available:
http://doi.acm.org/10.1145/2077451.2077468
[3]
D. Thalmann and S. R. Musse, “Modeling of populations,” in Crowd
Simulation, 2012, ch. 3, pp. 31–290.
[4]
K. L. Johnson and M. Shiffrar, People watching : social, perceptual, and
neurophysiological studies of body perception. Oxford University Press,
2013.
[5]
R. McDonnell, M. Larkin, B. Hern
´
andez, I. Rudomin, and
C. O’Sullivan, “Eye-catching crowds: Saliency based selective variation,
in ACM SIGGRAPH 2009 Papers, ser. SIGGRAPH ’09. New
York, NY, USA: ACM, 2009, pp. 55:1–55:10. [Online]. Available:
http://doi.acm.org/10.1145/1576246.1531361
[6]
L. Hoyet, K. Ryall, K. Zibrek, H. Park, J. Lee, J. Hodgins, and
C. O’Sullivan, “Evaluating the distinctiveness and attractiveness of
human motions on realistic virtual bodies,” ACM Trans. Graph.,
vol. 32, no. 6, pp. 204:1–204:11, Nov. 2013. [Online]. Available:
http://doi.acm.org/10.1145/2508363.2508367
[7]
G. Rhodes, “The evolutionary psychology of facial beauty,” Annual
Review of Psychology, vol. 57, pp. 199–226, 2006.
[8]
K. Johnson and L. Tassinary, “Perceiving sex directly and indirectly:
Meaning in motion and morphology,Psychological Science, vol. 16,
no. 11, pp. 890–897, 2005.
[9]
——, “Compatibility of basic social perceptions determines perceived
attractiveness,PNAS, vol. 104, no. 12, pp. 5246–5251, 2007.
[10]
A. C. Wellerdiek, M. Breidt, M. N. Geuss, S. Streuber, U. Kloos, M. J.
Black, and B. J. Mohler, “Perception of Strength and Power of Realistic
Male Characters,” in Proceedings of the ACM SIGGRAPH Symposium on
Applied Perception, ser. SAP ’15. New York, NY, USA: ACM, 2015, pp.
7–14. [Online]. Available: http://doi.acm.org/10.1145/2804408.2804413
[11]
B. Allen, B. Curless, and Z. Popovi, “The Space of Human Body
Shapes: Reconstruction and Parameterization from Range Scans,”
in ACM SIGGRAPH 2003 Papers, ser. SIGGRAPH ’03. New
York, NY, USA: ACM, 2003, pp. 587–594. [Online]. Available:
http://doi.acm.org/10.1145/1201775.882311
[12]
S. Wuhrer and C. Shu, “Estimating 3d human shapes from measurements,
Machine Vision and Applications, vol. 24, no. 6, pp. 1133–1147, Dec.
2012. [Online]. Available: http://link.springer.com/article/10.1007/s00138-
012-0472-y
[13]
D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and
J. Davis, “SCAPE: Shape Completion and Animation of People,”
in ACM SIGGRAPH 2005 Papers, ser. SIGGRAPH ’05. New
York, NY, USA: ACM, 2005, pp. 408–416. [Online]. Available:
http://doi.acm.org/10.1145/1186822.1073207
[14]
S.-Y. Baek and K. Lee, “Parametric human body shape modeling
framework for human-centered product design,Computer-Aided
Design, vol. 44, no. 1, pp. 56–67, Jan. 2012. [Online]. Available:
http://linkinghub.elsevier.com/retrieve/pii/S0010448510002289
[15]
Y. Zhang, J. Zheng, and N. Magnenat-Thalmann, “Example-guided
anthropometric human body modeling,” The Visual Computer,
vol. 31, no. 12, pp. 1615–1631, Dec. 2015. [Online]. Available:
http://link.springer.com/10.1007/s00371-014-1043-1
[16]
N. Magnenat-Thalmann and H. Seo, “Data-driven approaches to digital
human modeling,” in 2nd International Symposium on 3D Data Process-
ing, Visualization and Transmission, 2004. 3DPVT 2004. Proceedings,
Sep. 2004, pp. 380–387.
[17]
N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, and H.-P. Seidel, “A
Statistical Model of Human Pose and Body Shape,” Computer Graphics
Forum, vol. 28, no. 2, pp. 337–346, Apr. 2009. [Online]. Available:
http://doi.wiley.com/10.1111/j.1467-8659.2009.01373.x
[18]
H. Seo and N. Magnenat-Thalmann, “An Automatic Modeling of
Human Bodies from Sizing Parameters,” in Proceedings of the
2003 Symposium on Interactive 3D Graphics, ser. I3D ’03. New
York, NY, USA: ACM, 2003, pp. 19–26. [Online]. Available:
http://doi.acm.org/10.1145/641480.641487
[19]
O. Freifeld and M. J. Black, “Lie Bodies: A Manifold Representation of
3d Human Shape,” in Computer Vision ECCV 2012, ser. Lecture Notes
in Computer Science, A. Fitzgibbon, S. Lazebnik, P. Perona, Y. Sato,
and C. Schmid, Eds. Springer Berlin Heidelberg, Oct. 2012, no.
7572, pp. 1–14, dOI: 10.1007/978-3-642-33718-5 1. [Online]. Available:
http://link.springer.com/chapter/10.1007/978-3-642-33718-5 1
[20]
H. Seo and N. Magnenat-Thalmann, “An example-based
approach to human body manipulation,” Graphical Models,
vol. 66, no. 1, pp. 1–23, Jan. 2004. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S1524070303000778
[21]
K. M. Robinette, S. Blackwell, H. Daanen, M. Boehmer, and S. Fleming,
“Civilian american and european surface anthropometry resource (caesar),
final report. volume 1. summary,” DTIC Document, Tech. Rep., 2002.
Measurement Explanation
Acromial Height, Sitting Vertical difference from the sitting surface to the left acromion (shoulder)
Ankle Circumference
Arm Length (Shoulder to Wrist)
Hand Length Distance from tip of middle finger to the end of the palm
Head Circumference Circumference of the hand across knuckles of the index and little fingers
Hip Circumference, Maximum Maximum hip circumference measured parallel to the standing surface
Hip Circumference, Maximum, Height Vertical distance from standing surface to level of maximum hip circumference
Knee Height, Sitting Vertical distance from foot support surface to the top of knee
Neck Base Circumference Circumference of neck measured at the juncture of neck and shoulder
Shoulder Breadth Horizontal distance between maximum protrusion of left and right shoulder
Sitting Height Vertical distance sitting surface to the highest point of the head
Height Vertical distance between the standing surface and the highest point of the head
Thigh Circumference
Total Crotch Length Distance from front to back of the waist passing through the crotch
Waist Circumference
Waist Front Length Distance from front neck base to front of waistline in the median plane
Waist Height Vertical distance from waist to the standing surface
Weight
Armscye Circumference Circumference passing acromion (shoulder) and armpit
Chest Circumference Circumference of the torso at nipple level
Bust/Chest Circumference Under Bust
Buttock-Knee Length, Sitting Horizontal distance from foremost point of kneecap to rearmost point of buttock
Chest Girth at Scye Max circumference of torso passing under the arms and across upper chest
Crotch Height Vertical distance between crotch and standing surface
Eye Height, Sitting Vertical distance between sitting surface to outer corner of the eyes
Foot Length Maximum distance from the rear of the heel to the tip of the longest toe
Hand Circumference Circumference hand passing knuckles of index and little fingers
TABLE I
THE 27 CAESAR B ODY ME AS URE ME NTS U SE D TO GE NER ATE OU R 3D MODELS
Female Correlations All Females Hourglass Females Pear Females
MT-H MT-F LAB-H LAB-F ATT MT-H MT-F LAB-H LAB-F ATT MT-H MT-F LAB-H LAB-F ATT
MTurk Hit Rate (MT-H) - -0.5 - -0.6 - - -0.8 - -0.7 0.9 - - - -0.6 -
MTurk False Alarm Rate (MT-F) -0.5 - - 0.6 -0.5 -0.8 - - 0.8 -0.8 - - - - -0.5
Lab Test Hit Rate (LAB-H) - - - - - - - - - - - - - - -
Lab Test False Alarm Rate (LAB-F) -0.6 0.6 - - - -0.7 0.8 - - -0.7 -0.6 - - - -
Attractiveness Ratings (ATT) - -0.5 - - - 0.9 -0.8 - -0.7 - - -0.5 - - -
Chest Circumference (mm) - 0.4 - 0.4 -0.4 - - - - -0.6 - 0.8 - - -0.7
Hip Circumference, Maximum (mm) - 0.5 - - -0.9 -0.7 0.9 0.1 0.6 -0.9 - - - - -0.9
Waist Circumference, Pref (mm) - - - - -0.6 - - - - - - -0.6 - - -0.7
Waist to Chest Ratio (WCR) - -0.4 - -0.5 - - - - - 0.6 - -0.8 - - 0.6
Waist to Hip Ratio (WHR) - -0.6 - - 0.8 0.7 -0.9 -0.2 -0.7 0.8 - - - - 0.8
WCR-WHR 0.4 - - - -0.4 - - - - - 0.6 - - - -
Male Correlations All Males VShape Males Block Males
MT-H MT-F LAB-H LAB-F ATT MT-H MT-F LAB-H LAB-F ATT MT-H MT-F LAB-H LAB-F ATT
MTurk Hit Rate (MT-H) - -0.8 - -0.5 - - -0.9 - -0.5 0.8 - -0.7 - - -0.8
MTurk False Alarm Rate (MT-F) -0.8 - -0.4 0.6 0.4 -0.9 - - 0.6 -0.7 -0.7 - -0.6 - 0.6
Lab Test Hit Rate (LAB-H) - -0.4 - - - - - - - 0.5 - -0.6 - - -
Lab Test False Alarm Rate (LAB-F) -0.5 0.6 - - - -0.5 0.6 - - -0.6 - - - - -
Attractiveness Ratings (ATT) - 0.4 - - - 0.8 -0.7 0.5 0.6 - -0.8 0.6 - - -
Chest Circumference (mm) -0.1 0.4 -0.1 0.4 0.8 0.8 -0.7 0.4 -0.6 0.9 -0.4 0.7 -0.4 0.4 0.6
Waist Circumference (mm) - - - - - - - - - - - - - - -
Hip Circumference, Maximum (mm) - - - 0.5 - - - - - - - - - - -
Waist to Chest Ratio (WCR) 0.2 -0.3 0.1 -0.4 -0.8 -0.8 0.8 -0.2 0.6 -0.8 0.6 -0.6 0.3 -0.3 -0.7
TABLE II
ALL SIGNIFICANT CORRELATIONS BETWEEN THE BODY MEASUREMENTS AND OUR FIVE METRICS:HIT (H) AND FA LSE A LA RM RAT ES (F) FO R OU R TWO
EXPERIMENTS (MT AN D LAB), AND ATTRACTIVENESS RATINGS (ATT). ALL OT HER S IG NIFI CA NT CO RR ELATI ON S ARE P ROVI DE D IN TH E SU PPL EM ENTA RY
MATE RIA L.

Supplementary resource (1)

... In addition, Shi et al. [SOWO17] implement a data-driven approach to generate a variety of body shapes using the body measurements and demographic data available in the CAESAR database. They found, by conducting perceptual studies, that the height and girth (waist circumference) measurements were the most salient features that differentiate people from each other. ...
Article
Recent advancements in crowd simulation unravel a wide range of functionalities for virtual agents, delivering highly‐realistic, natural virtual crowds. Such systems are of particular importance to a variety of applications in fields such as: entertainment (e.g., movies, computer games); architectural and urban planning; and simulations for sports and training. However, providing their capabilities to untrained users necessitates the development of authoring frameworks. Authoring virtual crowds is a complex and multi‐level task, varying from assuming control and assisting users to realise their creative intents, to delivering intuitive and easy to use interfaces, facilitating such control. In this paper, we present a categorisation of the authorable crowd simulation components, ranging from high‐level behaviours and path‐planning to local movements, as well as animation and visualisation. We provide a review of the most relevant methods in each area, emphasising the amount and nature of influence that the users have over the final result. Moreover, we discuss the currently available authoring tools (e.g., graphical user interfaces, drag‐and‐drop), identifying the trends of early and recent work. Finally, we suggest promising directions for future research that mainly stem from the rise of learning‐based methods, and the need for a unified authoring framework.
... Perceptual models have been further applied to improving virtual simulations [13], character animations [8], [14], human body modeling [15], fluid simulations [16], [17], and crowd simulations [18], [19]. High dynamic range imaging and tone mapping benefits from models of human light adaptation [20], [21], color to grey conversions simulate human color sensitivity [22], [23]. ...
Article
Terrains are visually prominent and commonly needed objects in many computer graphics applications. While there are many algorithms for synthetic terrain generation, it is rather difficult to assess the realism of a generated output. This paper presents a first step towards the direction of perceptual evaluation for terrain models. We gathered and categorized several classes of real terrains, and we generated synthetic terrain models using computer graphics methods. The terrain geometries were rendered by using the same texturing, lighting, and camera position. Two studies on these image sets were conducted, ranking the terrains perceptually, and showing that the synthetic terrains are perceived as lacking realism compared to the real ones. We provide insight into the features that affect the perceived realism by a quantitative evaluation based on localized geomorphology-based landform features (geomorphons) that categorize terrain structures such as valleys, ridges, hollows, etc. We show that the presence or absence of certain features has a significant perceptual effect. The importance and presence of the terrain features were confirmed by using a generative deep neural network that transferred the features between the geometric models of the real terrains and the synthetic ones. The feature transfer was followed by another perceptual experiment that further showed their importance and effect on perceived realism. We then introduce Perceived Terrain Realism Metrics (PTRM) that estimates human perceived realism of a terrain represented as a digital elevation map by relating distribution of terrain features with their perceived realism. This metric can be used on a synthetic terrain, and it will output an estimated level of perceived realism. We validated the proposed metrics on real and synthetic data and compared them to the perceptual studies.
... While these work focus on how motion or appearance clones appear to the viewer, Shi et al. [2017] studied how variety in body shape is perceived and can be used to create a cost-effective realistic crowd. An interesting point is raised about how distinctive the selected representative samples should be such that the overall perceived variety is maximised. ...
... Perceptual models have been further applied to improving virtual simulations [13], character animations [8], [14], human body modeling [15], fluid simulations [16], [17], and crowd simulations [18], [19]. High dynamic range imaging and tone mapping benefits from models of human light adaptation [20], [21], color to grey conversions simulate human color sensitivity [22], [23]. ...
Preprint
Terrains are visually important and commonly used in computer graphics. While many algorithms for their generation exist, it is difficult to assess the realism of a generated terrain. This paper presents a first step in the direction of perceptual evaluation of terrain models. We gathered and categorized several classes of real terrains and we generated synthetic terrains by using methods from computer graphics. We then conducted two large studies ranking the terrains perceptually and showing that the synthetic terrains are perceived as lacking realism as compared to the real ones. Then we provide insight into the features that affect the perceived realism by a quantitative evaluation based on localized geomorphology-based landform features (geomorphons) that categorize terrain structures such as valleys, ridges, hollows, etc. We show that the presence or absence of certain features have a significant perceptual effect. We then introduce Perceived Terrain Realism Metrics (PTRM); a perceptual metrics that estimates perceived realism of a terrain represented as a digital elevation map by relating distribution of terrain features with their perceived realism. We validated PTRM on real and synthetic data and compared it to the perceptual studies. To confirm the importance of the presence of these features, we used a generative deep neural network to transfer them between real terrains and synthetic ones and we performed another perceptual experiment that further confirmed their importance for perceived realism.
Article
Data-driven methods for modeling the realistic shapeof 3D human bodies need to access datasets that contain a large amount of 3D human models. We develop a method based on sparse representation in this paper to represent 3D human models as signals of patches. Unlike the general mesh compression approaches, all mesh models used in a data-driven human modeling framework have the same mesh connectivity. By using this property, we segment a human model into patches containing the same number of vertices. L0-learning algorithm is selected to train an overcomplete dictionary matrix, which in turn introduces sparse representation of the dataset. Patch signals of individual human models can then be extracted by using the dictionary matrix. With the ease of balance control between sparsity and accuracy that is featured by the chosen learning algorithm, a representation with high compression ratio and low shape-approximation error can be determined. The results have been compared with the widely used statistic representation based on principal component analysis (PCA) to verify the effectiveness of our approach. Moreover, the method for using sparse representation in the regression-based statistical modeling of 3D human models has been presented at the end of the paper.
Conference Paper
Full-text available
We investigated the influence of body shape and pose on the perception of physical strength and social power for male virtual characters. In the first experiment, participants judged the physical strength of varying body shapes, derived from a statistical 3D body model. Based on these ratings, we determined three body shapes (weak, average, and strong) and animated them with a set of power poses for the second experiment. Participants rated how strong or powerful they perceived virtual characters of varying body shapes that were displayed in different poses. Our results show that perception of physical strength was mainly driven by the shape of the body. However, the social attribute of power was influenced by an interaction between pose and shape. Specifically, the effect of pose on power ratings was greater for weak body shapes. These results demonstrate that a character with a weak shape can be perceived as more powerful when in a high-power pose.
Article
Full-text available
Recent advances in rendering and data-driven animation have enabled the creation of compelling characters with impressive levels of realism. While data-driven techniques can produce animations that are extremely faithful to the original motion, many challenging problems remain because of the high complexity of human motion. A better understanding of the factors that make human motion recognizable and appealing would be of great value in industries where creating a variety of appealing virtual characters with realistic motion is required. To investigate these issues, we captured thirty actors walking, jogging and dancing, and applied their motions to the same virtual character (one each for the males and females). We then conducted a series of perceptual experiments to explore the distinctiveness and attractiveness of these human motions, and whether characteristic motion features transfer across an individual's different gaits. Average faces are perceived to be less distinctive but more attractive, so we explored whether this was also true for body motion. We found that dancing motions were most easily recognized and that distinctiveness in one gait does not predict how recognizable the same actor is when performing a different motion. As hypothesized, average motions were always amongst the least distinctive and most attractive. Furthermore, as 50% of participants in the experiment were Caucasian European and 50% were Asian Korean, we found that the latter were as good as or better at recognizing the motions of the Caucasian actors than their European counterparts, in particular for dancing males, whom they also rated more highly for attractiveness.
Article
Full-text available
The Civilian Americana and European Surface Anthropometry Resource (CAESAR) project was a survey of the civilian populations of three countries representing the North Atlantic Treaty Organization (NATO) countries: The United States of America (USA), The Netherlands, and Italy (Robinette et al. 1999, Robinette 2000). One site in Ottawa, Canada was added to the USA sample and it is henceforth referred to as the North American sample. The survey was carried out by the U.S. Air Force, with the of 1) the contractor, Sytronics Inc., 2) The Netherlands Organization for Applied Scientific Research (TNO), 3) the subcontractor D'Appolonia in Italy, and 4) a consortium of companies under the umbrella of the Society of automotive Engineers (SAB).
Article
Full-text available
We introduce the SCAPE method (Shape Completion and Animation for PEople)---a data-driven method for building a human shape model that spans variation in both subject shape and pose. The method is based on a representation that incorporates both articulated and non-rigid deformations. We learn a pose deformation model that derives the non-rigid surface deformation as a function of the pose of the articulated skeleton. We also learn a separate model of variation based on body shape. Our two models can be combined to produce 3D surface models with realistic muscle deformation for different people in different poses, when neither appear in the training set. We show how the model can be used for shape completion --- generating a complete surface mesh given a limited set of markers specifying the target shape. We present applications of shape completion to partial view completion and motion capture animation. In particular, our method is capable of constructing a high-quality animated surface model of a moving person, with realistic muscle deformation, using just a single static scan and a marker motion capture sequence of the person.
Article
Full-text available
When simulating large crowds, it is inevitable that the models and motions of many virtual characters will be cloned. However, the perceptual impact of this trade-off has never been studied. In this paper, we consider the ways in which an impression of variety can be created and the perceptual consequences of certain design choices. In a series of experiments designed to test people's perception of variety in crowds, we found that clones of appearance are far easier to detect than motion clones. Furthermore, we established that cloned models can be masked by color variation, random orientation, and motion. Conversely, the perception of cloned motions remains unaffected by the model on which they are displayed. Other factors that influence the ability to detect clones were examined, such as proximity, model type and characteristic motion. Our results provide novel insights and useful thresholds that will assist in creating more realistic, heterogeneous crowds.
Chapter
Virtual human models are becoming widespread in computer graphics applications such as virtual reality applications, simulations, and games. Usually, due to the complexity of human bodies, realistic shapes are built by means of an exhaustive and long process of graphic design. In addition, if an application requires the presence of groups and crowds of characters, a diversity of shapes and types of animations is necessary to realistically populate a virtual environment. As a consequence, the artists involved should manually create such multiplicity, increasing both the complexity of their task and the time needed to accomplish it.
Book
The human body has long been a rich source of inspiration for the arts, and artists have long recognized the body's special status. While the scientific study of body perception also has an important history, recent technological advances have triggered an explosion of research on the visual perception of the human body in motion, or as it is traditionally called, biological motion perception. Now reaching a point of burgeoning inter-disciplinary focus, biological motion perception research is poised to transform our understanding of person construal. Indeed, several factors highlight a privileged role for the human body as one of the most critical classes of stimuli affecting social perception. Human bodies in motion, for example, are among the most frequent moving stimulus in our environment. They can be readily perceived at a physical distance or visual vantage that precludes face perception. Moreover, body motion conveys meaningful psychological information such as social categories, emotion state, intentions, and underlying dispositions. Thus, body perception appears to serve as a first-pass filter for a vast array of social judgments from the routine (e.g., perceived friendliness in interactions) to the grave (e.g., perceived threat by law enforcement). This book provides an exciting integration of theory and findings that clarify how the human body is perceived by observers.
Article
This paper presents an example-guided, anthro-pometry-based modeling method for creating 3D human body models from users’ input of partial anthropometric measurements with a given example dataset. Rather than directly forming a mapping between the partial measurements and the body model, we first estimate a set of chosen 30 measurements from the input based on the example-oriented measurement analysis. We then create an initial 3D model using the example-oriented radial basis function model that maps the set of 30 measurements to the body shape space and is established based on the given examples. We finally refine the 3D model by constrained optimization to create the target body model. Our method has several advantages: (1) the created model is guaranteed to match the input measurements and reflects the shape characteristics of examples; (2) the input requirement is modest, which makes it useful in practice; and (3) the information of both the measurements and examples is fully utilized. We demonstrate the effectiveness, accuracy, flexibility and extensibility of the method by various experimental evaluations and a Kinect-based body customization application.