RJTA Vol. 9 No.

1 2005

A Virtual Clothing System for Retail and Design

Bernhard Spanlang1 , Tzvetomir Vassilev2 , Jonathan Walters3 and Bernard F. Buxton1
1
Dept. of Computer Science, University College London, UK, {b.spanlang, b.buxton}@cs.ucl.ac.uk
2
Department of Informatics, University of Rousse, Bulgaria, t.vassilev@ecs.ru.acad.bg
3
Contact through Bodymetrics Ltd. UK.

ABSTRACT

In this paper we describe a virtual clothing system for retail and design, created for
Bodymetrics Ltd., at their request. In the retail setup we installed the system at Selfridges, a
well known department store in London. To our knowledge this is a world first installation
of a fully automatic virtual try-on system. Sizing and body landmark information is
extracted from 3D scanner data in a fully automatic process and customers can try-on
garments from a database on their virtual selves within seconds. The system uses very fast
numerical methods and collision detection mechanisms that harness the capabilities of
graphics hardware for cloth body collision detection and response. Fabric property
measurements from a Kawabata evaluation system are mapped onto our cloth model to
ensure appropriate virtual drape behaviour. The information required in order to simulate
particular garments is provided in a format tailored to the computational requirements for
defining the garment pieces, seaming, partitioning on the body, cloth reflectance and
patterning, and location of accessories such as buttons, etc. Based on the realistic visual
feedback customers are able to decide whether to buy or not. The whole process in which a
customer is scanned, registered to the system, and virtually tries on ten different garments in
different sizes takes less than ten minutes. The acceptance of the system is shown by the
high interest and demand, the willingness of customers to pay for the service, and an
increased sale of items available in the virtual garment database. The system was also
designed to be usable over the Internet and has been made freely available. We show that the
same system in combination with global illumination for near photo-realistic augmented
reality is also of interest for garment designers.

Keywords: virtual try-on, cloth simulation, body scanning, textile properties,

global illumination, electronic garment exchange.

1. Introduction reflectance we are able to produce very realistic

The aim of this work was to build a system for
virtual try-on of garments for retail and design.
The system allows customers to browse and try
garments on their 3D scan and gives sufficient
visual feedback to decide whether to buy or not.
In the process important body landmarks and a
large number of size measurements are extracted
from the customers' body scan in a completely
automatic way. The fully automatic virtual
clothing system uses very fast numerical methods
and collision detection mechanisms that harness
the capabilities of graphics hardware for cloth
body collision detection and response. By these
means, we are able to virtually dress a customer
completely automatically with a chosen garment
within 1 to 2 seconds on a state of the art PC. By
using photographs of the real garment as an
approximation of a garment's inhomogeneous
74
images of a dressed customer. Fabric property
measurements from a Kawabata Evaluation
System (KES) (Kawabata, 1980) are mapped onto
our cloth model to ensure appropriate virtual
drape behaviour. The system has advantages over
a conventional changing room and mirror in that
customers can view themselves from any distance
and angle and try on different garments within
seconds. The system is built on Internet-based
technology and customers can view their body
from their home PC over a secure link and try on
new garments added to the system's central
database.
We have also integrated the system with a global
illumination system which allows a garment
designer to create near photo-realistic images of a
customer in a new garment design, naturally
illuminated at a location of choice. A client may
RJTA Vol. 9 No. 1 2005
approach a garment designer with a specific
occasion in mind where she wants to wear a new
design. For a designer it would then be beneficial
to experiment with the new garment in the target
environment and to be able to present near photo-
realistic images of the customer wearing the new
clothes at the specified location. Such a system
will save time and cost for the process of clothing
design.
We briefly discuss related work in cloth
simulation in the next section. Implementation
details about the garment simulation and
visualisation system are given in sections 3 and 4.
Details of the integration to a natural global
illumination system are given in section 5.
Feedback of customers and a designer is
summarised and example images that show the
quality of our simulation are given in the results
section. Finally, we draw conclusions and give
suggestions for future work in section 7.
2. Related Work
Methods to model cloth for computer graphics
have been investigated for about two decades.
Weil (Weil, 1986) was the first to report a model
for cloth objects using a two step geometric
process. Feynman (Feynman, 1986), around 1986,
developed the first physically-based model which
was based on elastic shell theory. Since then
researchers have found many ways to describe
cloth dynamics that differ mainly in their physical
accuracy, numerical stability and computational
cost. Mass-spring particle systems are mainly
used while some (Terzopoulos, et al., 1987;
Volino, et al., 1995; Carignan, et al., 1992;
Eischen, et al., 1996) employ finite element
methods (FEM). Breen et al. (1994) first
employed particle systems to approximate textile
behaviour from a KES (Kawabata, 1980).
Eberhardt et al (Eberhardt, et al., 1996) improved
Breens model by adding simulation of hysteresis
effects obtained directly from a KES. Provot
(1995) introduced a simple mass-spring topology
(see Figure 1) which is commonly used owing to
its efficiency and simplicity. Provot uses linear
(Hook) springs and applies explicit Euler
integration. To account for super-elongation,
caused by the linear springs, he constrains
particles' positions in a post correction step so that
springs can not extend above a certain threshold
(usually 5-10% of their natural length, depending
on the material properties to be simulated).
Repositioning according to the length constraint
75
of one spring, however, can lead to over-
elongation of other springs and may require
several iterations of the post correction steps to
resolve. Vassilev et al (2001) improved this by
modifying the particles' velocities instead of their
positions.
Bend
Shear Stretch
Fig. 1. Spring types in the cloth model.
The elastic model of cloth is a mesh of ln mass
points, each of them being linked to its
neighbours by massless springs of natural length
greater than zero. There are three different types
of spring:
Springs linking vertices [i, j] with [i+1, j], and
[i, j] with [i, j+1] are called stretch springs;
Springs linking vertices [i, j] with [i+1, j+1],
and [i+1, j] with [i, j+1] are called shear
springs;
Springs linking vertices [i, j] with [i+2, j], and
[i, j] with [i, j+2] are called bend springs.
As the names indicate, the first type of spring
implements resistance to stretching, the second
to shearing and the third to bending.
Let pij(t), vij(t), aij(t), where i=1,,l and j=1,,n,
be respectively the positions, velocities, and
accelerations of the mass points at time t. The
system is governed by the basic Newtons law:
fij = mij aij, (1)
where mij is the mass of point ij and fij is the sum
of all forces applied at point ij. The force fij can be
divided in two categories.
Internal forces arising from the tensions of the
springs. The overall internal force applied at the
point ij is a result of the stiffness of all springs
linking this point to its neighbours:
RJTA Vol. 9 No. 1 2005
p kl p ij
fint (p ij ) = k ijkl (p kl p ij ) lijkl
0 ,
p kl p ij
k ,l
where kijkl is the stiffness of the spring linking ij
0
and kl, and lijkl is the natural length of the same
spring.
The external forces can differ in nature
depending on what type of simulation we wish to
make. The forces most frequently included are:
Gravity: f ijgr = mg, where g is the gravity
acceleration;
Viscous damping: f ijvd = Cvd(vi-vj), where Cvd
is a damping coefficient,
Collision response.
From the above we may compute the force fij(t)
applied to point ij at any time t. The fundamental
equations of Newtonian dynamics can be
integrated over time by a simple Euler method:
1
a ij (t + t ) = f ij (t )
mij
v ij (t + t ) = v ij (t ) + t a ij (t + t ) ,
p ij (t + t ) = p ij (t ) + t v ij (t + t )
where t is a chosen time step. The Euler
Equations 3 are known to be very fast and to give
good results, provided the time step t is less than
the natural period of the system T0 m K .
Where K is the highest stiffness in the system.
Numerous recent work in cloth simulation (see for
example (Baraff and Witkin, 1998; Desbrun, et al.,
1999; Etzmuss et al., 2000) has shown that
improvements in stability are possible by using
implicit integration. However, for complex
garments with mapping of KES measurements to
the spring properties, explicit integration still
proved to be beneficial in terms of efficiency in
our case (Vassilev, et al., 2001). The advantages
of Euler integration became particularly apparent
when computation of the collision detection and
response, which require small time steps, were
taken into consideration. Similar results were also
indicated by Volino and Magnenat-Thalmann
(2001) in [29].
Collision Detection and Response
Collision detection proves to be a bottleneck of
dynamic simulation algorithms. This is especially
true if surfaces are highly discretised owing to the
number of potential colliding particles. Most of
(2)
the existing algorithms (Teschner, 2004) for
detecting collisions between cloth and other
objects in the scene are based on geometrical
object-space interference tests. They use
geometric calculations to detect penetration
between a cloth particle and a face of the object
together with strategies to minimise these
geometric calculations. Bounding volume
hierarchy approaches divide the space into simple
bounding volumes such as bounding spheres,
axis-aligned bounding boxes (AABB), object
oriented-bounding boxes (OOBB) or k-Dop
volumes (Lin and Gottschalk, 1998). The
hierarchies are represented by tree structures
which reduce the complexity from O(n2) of a
nave implementation to O(n log n), where n is the
number of polygons in cloth and object.
Additional culling of possible interactions
between cloth and cloth can be achieved by taking
the surface curvature of cloth into account. The
basic idea is that particles close to each other on a
low curvature surface are unlikely to collide. Use
of surface curvature is described in Volino et al .
(1995) and Provot (1997).
(3)
By partitioning the space around the cloth into
uniformly sized cells (voxels), the complexity of a
collision query can be reduced further from O(n
log n) to O(n). This is possible because voxel size
and positions are known and reference to them
can be made directly. Bigliani et al. (2000) and
Zhang et al. (2000) report such techniques. In
order to reduce memory requirements, they use
hash tables for the voxel data.
Vassilev et al. (2001) use similar ideas as that of
the voxel based approaches, described above, but
reduce the problem from 3D to 2 1/2D depth
maps. This approach is an image space based
collision detection in which the modern graphics
hardware is employed for rasterisation and depth
map generation for collision tests. Memory
requirements are thus drastically reduced.
Furthermore, interpolation units of the graphics
hardware are harnessed to compute smooth
surface normals for collision response. Vertex
velocities of dynamic objects are interpolated in a
similar way to allow quick calculation of the
response to dynamic collisions. The performance
of image space based approaches is independent
of complexity of the objects' geometry and
therefore particularly well suited for use in
combination with highly accurate densely
sampled 3D body scans.
76
RJTA Vol. 9 No. 1 2005
3. Customer Scanning and Garment
Description
In this section, we describe the scanning process
that we use to acquire shape information of a
customer. The garment exchange format which
includes description of cutting patterns, fabric
drape and reflectance properties, seaming, and
garment accessories will be detailed then.
Customer Shape Acquisition and
Measurement extraction
Several companies provide technology to acquire
human shape information by means of 3D
scanners. Since we already had experience with
the scanning system from the Textile/Clothing
Technology Corporation (TC) [15] from the UK
national sizing survey [16] we employed it again
here. The TC system uses a technique called
white light phase measurement profilometry in
which images of a structured light grid projected
onto the front and the back of the body are taken
and from distortions of the light grid the body
shape is acquired. This approach is well suited for
our requirements owing to the short acquisition
time (less than 10 seconds), safety, accuracy and
relatively low cost.
The direct output of the scanner is a 3D point
cloud. The body surface is reconstructed and
important body landmarks are extracted
automatically by TC software. A rule-based
language allows users to customise the extraction
of body landmarks and measurements.
In the application, the 3D models are stored in the
virtual reality modelling language (VRML)
format (Bell, et al., 1996). For simulation we
remove parts of the geometry that are not essential
for the virtual try-on and may not be scanned well
such as the head, hands and feet. To leave enough
space for dressing we lift the arms of the
customer's scan by rotating them around the
shoulders.
The TC2 surface reconstruction leaves holes in
areas that are partly obstructed during the
acquisition process. Such holes in the armpits,
crotch and hips areas are identified and patched in
the process. Fig. 2 shows an original scan on the
left, in the middle the arms were lifted and holes
patched, and on the right parts not essential for
virtual try-on were removed.
77
Bodymetrics software merges landmarks and
geometry in a compressed binary format. The
resulting file size of the body scan including all
required information for dressing and
visualisation, is approximately 150KB. Small file
size is important for efficient loading into main
and graphics memory for visualisation as well as
for possible transfer over the Internet.
Fig. 2. fltr: original scan; arms were lifted
and most holes patched; head,
hands and feet removed.
Garment Exchange Format
A cloth model such as that described in section 2,
is, on its own, not sufficient for garment
simulation. While standards to describe garments
exist [Error! Reference source not
found.,Error! Reference source not found.]
they are vulnerable for misinterpretations by
different CAD systems, they are not commonly
used by designers and do not provide sufficient
information for a virtual try-on system. To meet
our requirements, we have designed a simple
garment description in text format. The format
describes global garment properties such as the
mesh resolution and where the garment should
arrive on the body. A garment type is used to
classify the garment in order to correctly and
automatically pre-position it around a body scan.
At the moment four garment categories proved to
be sufficient for pre-positioning: bra, shirt, skirt,
and trousers.
The exchange format also defines the set of panels
comprising the garment. Garment panels are
described by their cutting pattern, (see Fig. 3)
fabric properties, Uniform Resource Locators
(URL) to images that approximate the reflectance
of the fabric, and URLs to garment accessories
(rigid geometry that might be attached to the
garment such as buttons, cuff links, etc.). In
compressed form even very complex garments
can be described by less than 10KB of data in the
format.
RJTA Vol. 9 No. 1 2005
Fig. 3. Garment panels of a pair of jeans in the
2D GUI on the left, and meshed
for simulation on the right.
For direct use of CAD files we integrated garment
mark-up functionality into an illustration program
commonly used by garment designers. This
allows a designer to import from CAD files and to
define in a graphical user interface all the
information required for virtual dressing. Export
to a file in the garment exchange format permits
immediate real-time simulation and visualisation
of the garment on a 3D scan.
For a skilled designer, marking up a garment for
virtual try-on takes no more than thirty minutes.
The graphical user interface and 3D visualisation
can be harnessed as an interactive 3D garment
design tool. Cutting pattern, drape properties,
seaming information and garment accessories are
described in the exchange format as follows.
Cutting Pattern Definition
Cutting patterns are described by a sequence of
2D coordinates that when connected by lines form
the outline of the individual garment panels. Files
exported from CAD systems usually contain a
seaming allowance for the manufacturing of real
garments. Since we are directly joining garment
panels at their edges as described below, such
allowances are not used by the seaming
simulation and need to be removed from the
cutting patterns. Also, cutting patterns from CAD
systems are often defined by spline curves which
we approximate to the required precision by
sequences of straight line segments. For pre-
positioning, each panel is classified in
combination with the garment type as to where it
should be placed around the body (front, back,
around arms, etc.)
Fabric Drape and Reflectance Properties
In order to make the virtual garment drape in a
way that a real fabric would, we use
measurements of real fabrics and incorporated
78
them into the system. Measurements from a KES
(Kawabata, 1980) were taken to obtain
information about fabric weight, resistance to
stretching, shearing and bending of the fabrics in
question. Stretch, and bend coefficients were
measured both in the warp and weft directions.
We fitted the measurements to simple linear
functions and mapped them to the corresponding
spring coefficients of our mass-spring model in a
similar manner as described in (Breen, et al., 1994;
Eberhardt, et al., 1996). The grain line, which
defines the orientation of the fabric in the garment
panel, is taken into consideration and is
approximated by linear interpolation of the warp
and weft coefficients according to a rotation angle
defined for each panel. Fabric thickness is
visualised by adding an extra polygon layer
displaced along the garment's surface normals.
The displacement distance is defined per panel.
Fabric reflectance and transparency is described
by photographs which are linked via a URL,
typically as JPEG images. Their position is
defined by UV texture coordinates of the
bounding rectangle of the relevant garment panel.
More information on reflectance estimation can
be found below.
Seaming Information and Garment
Accessories
In general, seaming lines describe pairs of
garment panel edges that need to be joined.
Usually, seaming pairs are on different garment
panels. They are described by a unique panel ID
and start and end distances from a defined origin
along the bounding contour of the cutting pattern.
Seaming information of pockets and garment
panels that are not sewn on the edge of a panel but
within one are defined by means of UV
coordinates of the bounding rectangle of the
panel's cutting pattern contour. Another case
important for seaming, is that of darts which
describe neighbouring pairs of edges on the same
panel that have to be joined. Darts are a form of
suppression and are used by designers mainly to
make a garment fit the body-shape more closely.
Often garments have rigid objects attached the
positions of which are dependent on the fabric's
drape. Examples of such objects are buttons, cuff
links, etc. We define them by using a URL to their
geometry and again use the panel ID and UV
coordinates for the position of the attachment on
the garment panel bounding rectangle.
RJTA Vol. 9 No. 1 2005
4. Garment Simulation System
In this section, we describe the virtual garment
simulation and visualisation system, indicate how
it is implemented and discuss briefly the user
interface. The interface was designed to be very
simple for use by customers in a store and to
protect their privacy.
Simulation and Visualisation
The dynamics and visualisation for real-time
garment simulation was implemented as an
ActiveX control by using the Active Template
Library (ATL) and the Interface Description
Language (IDL) (Rector and Sells, 1999) in
combination with C++ and OpenInventor
(Wernecke, 1994) ActiveX was chosen because it
can easily be embedded into HTML Internet
pages with JavaScript while at the same time it
gives access to the full processing and graphics
acceleration power of the hardware. The control is
divided into four main processes: URL loader,
interface, visualisation and simulation. Simulation
consists of three parts, the meshing, the numerical
simulation (ie. integration of Newton's law of
motion) and collision detection and response. We
employ directional velocity modification (DVM)
to compensate for spring over-elongation with an
explicit Euler integration as numerical methods
for the simulation as described in (Vassilev, et al.,
2001). Cloth-body collision detection and
response is carried out with our image space
based approach also described in (Vassilev, et al.,
2001) and for cloth-cloth collisions we employ
bounding volume hierarchies.
Garment Panel Meshing, Pre-Positioning,
Seaming and Draping
In order to apply the simulation algorithm, we
implemented a meshing method that reads the
cutting patterns and fabric properties and
produces a mass-spring system with the
rectangular topology as shown in Figure 1 and
Figure 3.
We decided to keep the meshing process as part
of our simulation module so that we could adjust
the drape simulation accuracy and speed by
setting the mesh resolution appropriately. Cutting
pattern contours require far less space than the
mass-spring mesh and therefore can be transferred
more efficiently over the Internet.
79
The defined garment type is used to pre-position
the garment accordingly around the scanned body
so that their seaming lines can be joined.
After the garment panels have been meshed and
pre-positioned, elastic forces are applied along the
seaming lines to join them together. This seaming
process is complete once all seaming points have
converged to a close proximity. If convergence
does not take place within a predefined number of
iterations the seaming process stops. This means
that the garment is too small and does not fit the
customer. We use the percentage of sewn points
as a fit feedback measure. After seaming has
finished we apply a number of iterations with
gravitational forces enabled to obtain realistic
drape behaviour.
Since the computations for the simulation mainly
involve vector operations we achieved
performance increases by exploiting some of the
latest processor architectures which implement
parallel data manipulation. Without optimising the
code for parallel computation but simply by
changing compiler settings to exploit single
instruction multiple data (SIMD) and parallel
execution a simulation speed up of 20% could be
achieved.
User Interface and Customer Discretion
Interaction with the dressed 3D body is restricted
to garment and viewpoint changes. The user can
zoom in and out, pan and rotate around the 3D
model by dragging the mouse.
For the try-on of garments we followed a simple
point and click paradigm. Once a garment has
been chosen it is immediately loaded into the
visualisation and the virtual customer is dressed.
Not surprisingly, a research study on body
scanning has revealed issues about customers'
embarrassment when viewing their bodies (Istook
and McKinnon, 1999). We assumed that this
would be particularly critical in a public place and
took measures to address such issues for the
garment retail setup.
Firstly, we designed clothing to make sure the
customer is always dressed properly. We call such
clothing default wear. An example is shown on
the left of Figure 7. The visualisation only shows
the customer when she is dressed. Secondly, when
a different garment is chosen we keep on
displaying the previous garment and wait until the
RJTA Vol. 9 No. 1 2005
simulation has converged to a certain degree
before we swap to visualise the simulation of the
new choice. The customer thus gets important
visual feedback from the seaming simulation
animation but at the same time we make sure she
always appears dressed.
5. Natural Global Illumination
Although not available for customer use in a store,
the system we have developed includes facilities
for rendering and visualising garments under
natural, global illumination. This is provided as a
facility for designers as indicated in the
introduction.
Global Illumination
The goal of global illumination is to approximate
the behaviour of light interactions in a scene.
Global illumination is of particular interest for the
realistic illumination of clothing because of
complex light interreflections between fabrics for
example, between clothing on the torso and on the
arms, or in the ridges, creases and folds of cloth.
Such interreflections and soft shadow effects
cannot be simulated with faster local illumination
systems. Radiosity and Monte Carlo path tracing
are the most commonly used approaches to find a
solution to the rendering equation that takes such
light interactions into account. RADIANCE
(Ward and Shakespeare, 1998) is one of the
systems that approximates the real behaviour of
light very well and is freely available in open
source. It uses a Monte Carlo path tracer with an
octree data structure and an irradiance caching
strategy for improved speed. RADIANCE was
mainly used for the illumination of architectural
designs but is very flexible and can also be
harnessed for the illumination of clothes. In
(Debevec, 1998) Debevec proposed to use
RADIANCE with high dynamic range (HDR)
panoramic photographs as light sources for global
illumination in order seamlessly to insert virtual
objects into real scenes. He coined this technique,
Natural Illumination. In Debevec's approach
the panoramic HDR images are composed of
photographs of a chrome ball taken from the same
view with different exposure settings (eg. varied
shutter speed on the camera). We employ this
technique for the system and exploit the panorama
transformation and HDR composition features of
HDRShop (Tchou and Devevec, 2001) to create a
so called light-probe of the target environment. A
light-probe contains the radiance measurements
80
of the scene at a certain point, therefore, the light-
probe should be taken with the chrome ball
located near where the client is placed in the
target photograph. An example light-probe of our
designer's office is depicted in Figure 4.
Fig. 4
Fig. 4. A HDR Light probe of a target location
for illumination of virtual garments
shown at increasing f-stops from
left to right.
After the real-time garment simulation has
converged to a stable state we convert the 3D
mesh and reflectance data to a format that can be
read by RADIANCE. The appropriate HDR
panoramic image or light-probe is loaded into
RADIANCE for lighting and an image is rendered
with the camera parameters acquired in the
registration process (see next sub-section). In
order to render with indirect illumination, such as
a light-probe, the number of diffuse reflections of
a ray (also called ambient bounces in
RADIANCE) simulated has to be greater than
zero. In fact, we set it to two to account for fabric
interreflections as well. Ambient sampling
parameters are set relatively high to achieve good
rendering quality.
Merging the Real and Virtual
In order to make the virtual garment appear on the
client's photograph, as if the client were dressed in
the virtual garment, the photograph of the client
has to be aligned with an image of the 3D scan of
the client. This could be achieved by extracting
camera position and direction by using a camera
calibration technique such as that described by
Zhang in (Zhang, 2000) and by rendering the 3D
scan and garment with the camera settings of the
photograph. However, for camera calibration,
multiple photographs of a calibration pattern in
the target scene are required which may not be
available. Automatic alignment of the model with
the image can also be found with image
registration techniques. See Lester and Arridge
(Lester and Arridge, 1999) for a survey. If the
pose of the client in the image is significantly
different to that of the client's 3D scan, the scan
RJTA Vol. 9 No. 1 2005
has to be animated with the garment to approach
the pose in the photograph as closely as possible.
An automatic character animation system
combined with an image registration method
would be required to find the right pose
automatically. For this system we did not
concentrate on automatic model alignment.
Instead, we superimposed the virtual model over
the client's photograph and adjusted the virtual
camera to allow us to visually match photograph
and scan. In addition, rather than attempting to
register the limbs, in particular the client's arms,
in the scan to match the photograph, we chose to
replicate appropriate background texture over the
original position of the arms in the photograph.
Part of the subject's left arm shown in Figure 8 to
illustrate this process.
Fabric Reflectance Estimation
Owing to its micro- and meso-structure the
bidirectional reflectance distribution function
(BRDF) of fabric can be very complicated. We
approximated the spatially varying fabric
reflectance, described by a bidirectional texture
function (BTF), by taking photographs of a
sample of the fabric in diffuse light. Therefore the
approximation is described by a Lambertian
texture. Colour and light calibration was achieved
by inserting a MacBeth colour checker in the
image with the fabric. The image is balanced
afterwards so that the known values of the
MacBeth chart best match those of the image in a
least squares sense. Since our fabric reflectance
sample was not always large enough to create the
whole virtual garment we used a texture synthesis
method briefly described below.
Masking the Rendering for Photo Composition
In order to subtract a garment from the
background of a synthetic image it is necessary
to create a mask image. The problem is that
client and garment interact and it would be
difficult manually to generate a good mask to
sufficient accuracy. Instead we use an
automatic depth buffer approach. We create a
depth buffer image for the 3D scan and
background (the light probe) without the
garment and another depth buffer image of
scan with the simulated garment. A simple
filter function is then used to generate a mask
image. The filter function returns 1 if the depth
value of the garment is greater than the depth
value of the client and 0 otherwise. An example
mask is depicted on the left in Figure 5 Fig.
81
5and on the right a rendered jacket extracted from
the background is depicted.
Fig. 5. Left: automatically generated mask to
extract a garment from a background,
right: extracted garment.
Synthesis of Occluded Background
When merging images, usually the background,
where the new image is inserted, is lost.
Sometimes it is also required to recreate parts of
the background texture occluded by a real
garment. For example, parts of our client's real
garments may not be sufficiently covered by the
synthetic garment and therefore need to be
replaced by the background expected in the image
as indicated in Figure 8. Ideally, an image of the
background without the client is available, but if
this were not the case, the background could be
approximated by use of texture synthesis
techniques. Generally texture synthesis takes as
input an area of the image that is similar to what
should be regenerated. The area that should be
recovered is marked and synthesized with a
texture that most likely would have covered that
area according to a probability distribution of the
specified sample region. We employed a
technique described by Harrison (Harrison, 2001)
for this, but the results, though promising, were
not as good as those we could obtain by using
parts of a background image to replace areas in
the foreground of Figure 9.
6. Results
The system was tested in two setups. In the retail
service without use of the global illumination
methods and in the augmented reality garment
design application with the natural global
illumination methods.
Retail Installation
Statements about the retail service are based on
informal interviews with some of the operators
that run the service. Generally, these operators
RJTA Vol. 9 No. 1 2005
were female students from the London College of
Fashion (LCF). Their education provides a very
good background to give recommendations on the
fit of garments. Also, they know how to work
with computers though they are not expert
computer users. A brief training of less than an
hour is sufficient for them to run the service.
System Reliability
About one in every fifty customers is scanned a
second time owing to size and landmark
extraction problems of the TC2 software. No
problems have been reported with the
visualisation and simulation system.
Customer Feedback
At the moment the system is for women only, this
limitation is because of the type of garments
available for simulation. Ten different designer
jeans brands in five different sizes are offered for
virtual try-on at the moment. New customers are
added to the system every day. They read about
the service in the news, have heard about it on TV
or are informed by word of mouth. They are
coming mainly alone or with a male friend.
Customers are normally coming to buy jeans
anyway and they are willing to pay for the service.
In general, they are finding the system useful and
they find the real-time garment visualisation
sufficiently realistic. A few customers completely
trust the simulation and would buy the garment
without trying them on in person. Operators
recommend to try-on the real garment before
buying however. Used in this manner, the system
helps customers find the best fitting size and style
of jeans that they like. Customers are very
interested to try out new virtual garment items
added to the database.
We found that most of the customers are very
confident with their body and embarrassment
about viewing their body as suggested by (Istook
and McKinnon, 1999) only occurs rarely.
Some of the customers would like to enter the
scanner with high heels to appear differently.
Quite a few female customers seem to be
encouraged by their male friends to try-out the
system. It is also mainly the males who are
interested in the data format of the scan and
whether they could access it from the Internet.
Other Visitors
There are many people visiting the installation
just because of interest in the technology. These
82
are mainly people from the film and games
industries but also university students who want
to find out what the system does and how it works.
Internet Access
Once scanned, customers can access the system
over a secure link on the Internet. There are now
several thousand scans stored in the central
database of Bodymetrics.
Visual Quality
Images that illustrate the visual quality of the
simulation are given below. Figure 7 shows one
of our customers in default wear on the left. A
front view of the customer in a pair of jeans is
given in the middle of Figure 7 and a view from
the back of the same customer in the same pair of
jeans is depicted on the right. An illustration of
the visual feedback a customer would get if a pair
of jeans does not fit so well is depicted in Figure 6.
On the left we show a customer in jeans that are
too tight for the seams to join properly and on the
right jeans that are too large for this customer and
therefore appear very baggy.
Augmented Reality Garment Design
We collaborated with a designer to test the system
in a real world design setting. As a simple
garment, our designer was asked to make a jacket.
In order to verify the simulation our designer
created the jacket not only in the CAD system for
simulation but also made a real toile of the
garment. The designer's client was scanned with
the TC2 system and we took photographs of the
client in the real jacket for comparison. A light
probe was taken in our designer's office as
depicted in Figure 4. This was done under real-
life conditions, with the office otherwise in near-
normal use, as can be seen from the fact that the
door is slightly open in Figure 8 (and therefore
also in Figure 9, right) but closed in Figure 9, left).
Figure 9 depicts a client in a real jacket toile on
the left and on the right the client dressed in a
virtual jacket. We note that owing to natural
illumination the simulated garment fits
realistically into the photograph. Since hands
were chopped off the 3D scan they are not present
in the image with the virtual garment either. There
are creases in the real jacket which do not appear
in the simulation. In general the simulation
appears much smoother. This may be for two
reasons. Firstly, the garment simulation does not
account for memory effects of the fabric, while,
secondly, the resolution of the mass-spring mesh
could be increased to generate more detailed
RJTA Vol. 9 No. 1 2005

Fig. 6. Jeans that do not fit well; left a customer with jeans too tight; right a customer with very baggy jeans.

Fig. 7. fltr: Customer in default wear; front view of a customer in jeans;

back view of the same customer in the same pair of jeans.

83
RJTA Vol. 9 No. 1 2005

Fig. 8. 3D scan superimposed on the photograph to approximate camera parameters.

Fig. 9. Left client in the real jacket toile, right the naturally illuminated and
simulated jacket composited with a photograph of a client.

84
RJTA Vol. 9 No. 1 2005
creases. Also, some shadows, for example, under
the arms, appear stronger and sharper in the
simulation than in real-life. Finally, we note that
parts of the simulated garment, such as on the
pockets, right sleeve and left lapel, appear less
smooth on the simulation than in real-life. This
appears to be because of limitations on the
interpolation of normals in the RADIANCE
system when they are used for calculating
interreflection effects. Such artefacts could also
be reduced by using higher resolution meshes. In
spite of these deficiencies, both the designer and
client have found the system very useful and
interesting. Our designer commented on possible
improvements mainly on the garment design and
simulation interfaces.
7. Conclusions and Future Work
A system for fully automatic virtual clothing
simulation was installed successfully in a
department store. The same system was tested by
a garment designer as an augmented reality
garment design tool to create near photo-realistic
images of a client in virtual garments. At the time
of writing the retail installation has run for over
eighteen month. There is high customer demand
for the system and, to meet this, we need to add
more virtual garments to the database. The system
is also freely accessible on the Internet over a
secure link. There are now several thousand
customer scans in the central database. A market
research study will be carried out to identify the
main customer demands to improve the system. A
possible system improvement may be to allow the
virtual customer to change poses or even to carry
out animation sequences based on motion capture
data in order to get a better idea about how the
garment will interact with the body. The
visualisation quality can be improved by adding
more detailed fabric reflectance properties.
Quantitative data, such as the percentage of sewn
points mentioned in section 4, to give fit feedback
to the customer could be extracted from the
garment simulation data. We believe that
automatic techniques to extract a full garment
description from photographs of real garments
will be a fruitful area of future research.
For the augmented reality garment design setup
an automatic technique visually to register a
client's photograph with a 3D scan in arbitrary
pose will be a challenge. In addition, it would be
interesting to extend the system to a collaborative
garment design tool for use over the Internet.
85
Acknowledgement
The system described in this paper is the property
of Bodymetrics Ltd 1 and was commissioned by
them. The system utilises components licensed
from UCL.
Thanks also to Anki Lindblom for designing the
jacket and Kelvin Lawrance for work on the on-
line database.
REFERENCES
[1] 1992, Aama/ansi 292 American national
standard for pattern data interchange.
[2] 2001, D6673-01 standard practice for
sewn products pattern data interchange-
data format.
[3] Baraff, David and Witkin, Andrew (1998),
Large steps in cloth simulation, In
Computer Graphics Proceedings, Annual
Conference Series, pages 43-54.
SIGGRAPH.
[4] Bell, G., Carey, R. and Marrin, C. (1996),
The Virtual Reality Modeling Language
Specification. Technical report, Version
2.0, ISO/IEC CD 14772, August.
[5] Bigliani, R. and Eischen, J.W. (2000),
Collision detection in cloth modeling. In
D.H. House and D.E. Breen, editors, Cloth
Modeling and Animation, pages 197217.
A K Peters.
[6] Breen, D.E., House, D.H. and Wozny, M.J.
(1994), Predicting the drape of woven
cloth using interacting particles. In
Computer Graphics Proceedings, Annual
Conference Series, volume 94, pages 365-
372.
[7] Carignan, M., Yang, Y., Magnenat
Thalmann, N. and Thalmann, D. (1992),
Dressing animated synthetic actors with
complex deformable clothes, In Computer
Graphics Proceedings, Annual Conference
Series, volume 92, pages 99-104.
[8] Debevec, P. (1998), Rendering synthetic
objects into real scenes: Bridging
traditional and image based graphics with
global illumination and high dynamic
range photography, 32(Annual
Conference Series):189-198, August.
1
More information about Bodymetrics Ltd. can be found on
the Internet at www.bodymetrics.com
RJTA Vol. 9 No. 1 2005

[9] Desbrun, M., Schroeder, P. and Barr, A.
(1999), Interactive animation
structured deformable objects.
Proceedings of Graphics Interface
(GI1999), pages 1-8. Canadian Computer-
Human Communications Society.
[10] Eberhardt, B., Weber, A. and Strasser, W.
(1996), A fast, flexible, particle-system
model for cloth draping, j-IEEECGA,
16(5):52-59, September.
[11] Eischen, J.W., Deng, S. and Clapp, T.G.
(1996), Finite element modeling and
control of flexible fabric parts, IEEE
Computer Graphics and Applications,
16(5):7180, ISSN 0272-1716, September.
[12] Etzmuss, Olaf, Eberhardt, B. and Hauth,
Michael, Implicit-explicit schemes for
fast animation with particle systems. In
Computer Animation and Simulation 2000,
pages 138-151, Eurographics, August
2000.ISBN 3-211-83392-7.
[13] Feynman, C.R. (1986), Modeling the
Appearance of Cloth, MS thesis, EECS
dept, MIT, May.
[14] Harrison, P. (2001), Non-hierarchical
procedure for re-synthesis of complex
textures. In V. Skala, editor, WSCG 2001
Conference Proceedings.
[15] Home page of the textile clothing
technology cooperation. www.tc2.com.
[16] Home page of the UK national sizing
survey. www.sizeuk.org.
[17] Istook, C.L. and McKinnon, L. (1999),
Psychological Issues Concerning Body
Scanning, Annual International Textile
and Apparel Association Meeting,
November 10-13, Santa Fe, NM
(November ).
[18] Kawabata, S. (1980), The standardization
and analysis of hand evaluation, The
Textile Machinery Society of Japan.
[19] Lester, H. and Arridge, S. (1999), A
survey of hierarchical non-linear medical
image registration, Pattern Recognition,
32(1):129-149.
[20] Lin, M. and Gottschalk, S. (1998),
Collision detection between geometric
models: A survey.
[21] Provot, X. (1995), Deformation
constraints in a mass-spring model to
describe rigid cloth behaviour,
Proceedings of Graphics Interface, 141-
155.
[22] Provot, X. (1997), Collision and self-
of collision handling in cloth model dedicated
In to design, Computer Animation and
Simulation 97, pages 177-190, ISBN 3-
211-83048-0. Held in Budapest, Hungary
September.
[23] Rector, B.E., Sells, C. (1999), ATL
Internals (The Addison-Wesley Object
Technology Series) by Pearson
Educational; 1st edition ISBN:
0201695898 March 31
[24] Tchou, C. and Debevec, P. (2001),
HDRShop.
[25] Terzopoulos, D., Platt, J., Barr, A. and
Fleischer, K. (1987), Elastically
deformable models, Computer Graphics
(Proc. SIGGRAPH87), 21(4):205214.
[26] Teschner, M., Kimmerle, S., Heidelberger,
B., Teschner, M., Zachmann, G.,
Raghupathi, L., Fuhrmann, A., Cani, M.P.,
Faure, F., Magnenat Thalmann, N.,
Strasser, N. and Volino, P., Collision
detection for deformable objects,
Eurographics, Sep 2004.
[27] Vassilev, T., Spanlang, B. and
Chrysanthou, Y. (2001), Fast cloth
animation on walking avatars, Computer
Graphics Forum, 20(3):260267, ISSN
1067-7055.
[28] Volino, P. and Magnenat-Thalmann, N.
(1995), Collision and self-collision
detection: Effcient and robust solutions for
highly deformable surfaces, Computer
Animation and Simulation 95, pages 55
65, September. ISBN 3-211-82738-2.
[29] Volino, P. and Magnenat-Thalmann, N.
(2001), Comparing efficiency of
integration methods for cloth simulation,
In Computer Graphics International 2001,
pages 265-272, July. ISBN 0-7695-1007-8.
[30] Volino, P., Courchesne, M. and Magnenat-
Thalmann, N. (1995), Versatile and
efficient techniques for simulating cloth
and other deformable objects, In Proc.
SIGGRAPH95, pages 137-144.
[31] Ward, G. and Shakespeare, R. (1998),
Rendering with radiance: the art and
science of lighting visualization, Morgan
Kaufmann Publishers Inc.
[32] Weil, J. (1986), The synthesis of cloth
objects, In David C. Evans and Russell J.
Athay, editors, Computer Graphics
(SIGGRAPH 86 Proceedings), volume 20,
pages 49-54, August.

86
RJTA Vol. 9 No. 1 2005

[33] Wernecke, J. (1994), The Inventor Mentor:
Programming Object-Oriented 3D
Graphics with Open Inventor, Release 2,
Addison Wesley.
[34] Zhang, D. and Yuen, M.M.F. (2000),
Collision detection for clothed human
animation, 8th Pacific Conference on
.
Computer Graphics and Applications,
pages 328-337, October. ISBN 0-7695-
0868-5.
[35] Zhang, Z. (2000), A flexible new
technique for camera calibration, IEEE
Transactions on Pattern Analysis and
Machine Intelligence, 22(11):13301334.

87