See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318127007

Fit evaluation of virtual garment try-on by learning from digital pressure data

Article  in  Knowledge-Based Systems · July 2017

DOI: 10.1016/j.knosys.2017.07.007

CITATIONS READS

10 1,980

6 authors, including:

Kaixuan LIU Xianyi Zeng

Ecole d'ingénierie et d'innovation textile Ecole d'ingénierie et d'innovation textile
31 PUBLICATIONS   213 CITATIONS    270 PUBLICATIONS   1,819 CITATIONS   

SEE PROFILE SEE PROFILE

Pascal Bruniaux Edwin Kamalha

Ecole d'ingénierie et d'innovation textile Politecnico di Torino
111 PUBLICATIONS   501 CITATIONS    19 PUBLICATIONS   135 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

IOTFetMov View project

SUCRé: Human-Human and Human-Robots Cooperation in hostile environment View project

All content following this page was uploaded by Kaixuan LIU on 05 October 2017.

The user has requested enhancement of the downloaded file.

 Knowledge-Based Systems 133 (2017) 174–182

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

Fit evaluation of virtual garment try-on by learning from digital

pressure data
Kaixuan Liu a,b,c,d,∗, Xianyi Zeng c,d, Pascal Bruniaux c,d, Jianping Wang b, Edwin Kamalha c,d,
Xuyuan Tao c,d
a
Apparel and Art Design College, Xi’an Polytechnic University, Xi’an 710048, China
b
College of fashion and design, Donghua University, Shanghai 200051, China
c
University of Lille 1, Nord de France, 59000 Lille, France
d
GEMTEX Laboratory, ENSAIT, 59100 Roubaix, France

a r t i c l e i n f o a b s t r a c t

Article history: Presently, garment fit evaluation mainly focuses on real try-on, and rarely deals with virtual try-on. With
Received 25 March 2016 the rapid development of E-commerce, there is a profound growth of garment purchases through the
Revised 29 June 2017
internet. In this context, fit evaluation of virtual garment try-on is vital in the clothing industry. In this
Accepted 2 July 2017
paper, we propose a Naive Bayes-based model to evaluate garment fit. The inputs of the proposed model
Available online 3 July 2017
are digital clothing pressures of different body parts, generated from a 3D garment CAD software; while
Keywords: the output is the predicted result of garment fit (fit or unfit). To construct and train the proposed model,
Digital clothing pressure data on digital clothing pressures and garment real fit was collected for input and output learning data
Support vector machines respectively. By learning from these data, our proposed model can predict garment fit rapidly and auto-
Naive Bayes matically without any real try-on; therefore, it can be applied to remote garment fit evaluation in the
Active learning context of e-shopping. Finally, the effectiveness of our proposed method was validated using a set of
Ease allowance
test samples. Test results showed that digital clothing pressure is a better index than ease allowance to
Real try-on
evaluate garment fit, and machine learning-based garment fit evaluation methods have higher prediction
accuracies.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction module, 2) a fabric properties module, and 3) a virtual pattern

sewing module. To model the human body rapidly, the 3D para-
Today garment e-shopping has become more prominent world- metric mannequin module is used to construct a personalized 3D
wide [1]. However, an important technical barrier that garments human model from measurement of a 3D body scanner or a mea-
displayed online cannot be physically evaluated for fitting effects suring tape, related to a specific customer. Several key body dimen-
on a specific consumer [2]. Thus, virtual try-on technology was sions, such as height, waist circumference, and hip circumference
developed to evaluate garment fit [2,3], finding wide application [6], control the parametric mannequin’s dimensions. By adjusting
in the clothing industry in the last ten years. A number of virtual key body dimensions, the 3D parametric mannequin module can
try-on programs, such as Clo 3D, Lectra 3D Prototype, OptiTex and create various body shapes and dimensions rapidly and automati-
V-Stitcher 3D, are available on the market for garment fit evalu- cally, meeting customers’ body shapes and dimensions. Then, the
ation [4]. These 3D virtual try-on software systems follow simi- fabric properties module, usually based on a mechanical model,
lar principles, i.e. showing virtual garment static and dynamic per- will permit the simulation of different perceived properties (drap-
formance from identified human morphological and fabric proper- ing, texture, elasticity, bending, etc.) of a virtual fabric through
ties and their interactions. Defining this performance involves the adjustable fabric technical parameters. Finally, the virtual pattern-
use of complex mechanical and geometric modeling and simula- sewing module assembles the predefined garment patterns on the
tion techniques, such as finite elements [5]. The software, normally specific 3D human body and sews the patterns together, taking
include four main modules [4]: 1) a 3D parametric mannequin into account the performance of the simulated fabric. The combi-
nation of these three modules constitutes a virtual try-on system,
permitting the simulation of the real garment making process. In
∗
Corresponding author at: College of fashion and design, Donghua University, a virtual 3D try-on process, consumers and designers can visualize
Shanghai 20 0 051, China. the static and dynamic performance of the selected fabrics and
E-mail address: [email protected] (K. Liu).

http://dx.doi.org/10.1016/j.knosys.2017.07.007
0950-7051/© 2017 Elsevier B.V. All rights reserved.
 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182 175

garment fit effects in terms of comfort, expressed by simulated tal pressure-based method will be more efficient for fit evaluation
pressures between the human body and fabrics, and fashion styles. than ease allowance-based methods, which are more adapted to
Through virtual try-on, consumers can easily decide whether loose garments.
they like a garment style or not. However, these virtual try-on ap- Bayes classifiers have been applied successfully in a wide vari-
plications, are strongly dependent on mathematical models used in ety of domains [42]. Their representations are quite intuitive and
the software [5], and cannot give full accurate garment fit evalua- easy to understand [43]. The advantages of Naive Bayes are: 1) A
tion. Moreover, for real try-on, the wearer can feel whether a gar- Naive Bayes model has a solid mathematical foundation, as well
ment fits; unlike, with virtual try-on [7]. The issue of garment fit as the stability of the classification efficiency. 2) The number of
evaluation is a research hotspot and a great challenge [8]. In prac- estimated parameters for modeling with Naive Bayes is relatively
tice, no matter how beautiful a garment is, and how excellent the fewer, the model is less sensitive to missing data, and the algo-
fabric’s properties are, a customer will not select it if it is unfit [8]. rithm is relatively simpler. 3) A Naive Bayes model has very high
Garment fit is a major factor affecting customers’ purchasing deci- classification accuracy in many practical cases. Due to these advan-
sions [9,10]. For garment e-shopping, consumers cannot physically tages, Naive Bayes is applied to model the relationship between
try on garments; therefore, estimating the garment fit without real digital clothing pressures and garment fit level. The inputs of the
try-on is still an issue for researchers. proposed model are digital clothing pressures on different body
Lately, there are mainly two methods to evaluate garment fit parts, while the output is the fit evaluation result (fit or unfit). By
through virtual try-on. One approach is that the visual evaluation learning from a number of experimental data measured on a num-
is carried out on a 3D garment by expert fashion designers [11–13]. ber of samples, we set up the model, capable of quickly estimating
Obviously, this subjective visual evaluation is neither accurate nor the fit for a new garment without any real try-on. Our proposed
convincing. The other approach is to measure the ease allowances, model can be used to help consumers in realizing online efficient
which is the dimensional difference between human body and gar- garment shopping.
ment in the girth direction [14]. Then, expert fashion designers The following sections are organized as follows. Section 2 in-
analyze these measured ease allowances based on their own em- troduces the general scheme and data formalization.
pirical knowledge to evaluate garment fit. However, the ease al- Sections 3 presents the collection of input and output learn-
lowance can reflect the fitting feeling when neither it is less or ing data respectively. Section 4 expounds the construction of
equal to zero (tight garment style) nor does it take into account garment fit prediction model. In Section 5, we evaluate the accu-
fabric properties. With the same value of ease allowance, the fit- racy of the proposed model and give a practical application of the
ting effects could be different if the fabrics are different. Evidently, proposed model. In Section 6, we discuss the application prospect,
ease allowance only is not enough for characterizing the fitting ef- limitation, etc. Finally, we present some conclusions and possible
fects of a garment try-on. Moreover, these two approaches to gar- further works in Section 7.
ment fit need empirical knowledge. Often, the predicted results
are entirely dependent on the subjectivity of designers, such as
2. General principle and formalization
experience and personal preference, which erodes their accuracy.
Therefore, people without fashion design knowledge cannot easily
2.1. General principle
use the traditional fit evaluation methods of virtual garments. For
shopping online, there are thousands of garments purchased in a
The general scheme of the mentioned garment fit evaluation
short time. If every garment fit is evaluated using traditional meth-
model is described in Fig. 1.
ods, the work is so enormous. Therefore, it is necessary to find a
First, Experiment I is aimed at collecting the output learning
method that can evaluate garment fit automatically, rapidly and ac-
data. Nine subjects try on 72 pairs of pants respectively. They di-
curately. In this context, we proposed a machine learning-based
vide all the pants into fit pants and unfit pants.
model to evaluate garment fit. The input item of the proposed
Second, Experiment II is aimed at collecting the input learn-
model is an indicator reflecting the garment fit condition, whose
ing data. We measured digital clothing pressures of the 72 pairs
output is fit or unfit. Compared to traditional garment fit evalua-
of pants respectively by virtual try-on technology.
tion methods, the greatest advantage of our proposed method is
Next, garment fit evaluation model based on Naive Bayes and
that it can predict garment fit rapidly and automatically, without
Support Vector Machines (SVMs) is trained by the input and out-
any real try-on and designers’ involvement.
put learning data.
Modeling by learning from experimental data has been widely
Finally, the proposed model predicts garment fit without any
used in the clothing industry [15,16], including evaluation of gar-
real try-on after learning from the collected data.
ment and fabric products [12,17–19], evaluation of wear comfort
[20], garment CAD systems [21–24], clothing manufacturing [25–
30]; clothing retailing [31–36], and apparel supply chain manage- 2.2. Formalization of the concepts and data
ment [37,38]. However, few studies have focused on fit evaluation
of virtual try-on using machine learning. In order to increase the As shown in Fig. 2, we built a machine learning-based garment
accuracy of fit evaluation using 3D virtual body shapes and vir- fit evaluation model, the input of which are digital clothing pres-
tual garments, we introduce a data machine learning-based model. sures and the output of which is garment fit level. The data and
This method requires inputting an indicator that can reflect gar- concepts involved in this study are formalized as follows:
ment fit and returning an output as the predicted result of gar- Let FL be the fit level of a garment, i.e. 1-fit, 0-unfit.
ment fit. In the preceding sections, we have discussed and indi- Let G = {g1 , g2 , . . . , gm } be a set of m real garments used in our
cated that the ease allowance between a garment and the human study.
j
body is not a good indicator of reflecting garment fit. We, there- Let Pi = ( p1i , . . . , pi , . . . , pki ) be a vector of digital clothing pres-
fore, opted to find a more suitable indicator. The influence of fab- sures obtained during the virtual try-on of the garment gi where
j
ric properties can be measured using the digital clothing pressures, pi is the pressure on the key position j of the garment gi (we
distributed over the human body of the wearer and provided by a suppose that there exist k key positions on the whole garment
3D garment CAD software like CLO 3D [39–41]. It is for this reason surface). In a general case, the vector of digital pressures Pnew =
j
that, in our study, we selected digital clothing pressures as a key ( p1new , . . . , pnew , . . . , pknew ) of a new garment g, is taken as input
indicator for remote garment fit prediction. In practice, the digi- variables of the model.
176 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182

Experiment I: Output data collection Experiment II: Input data collection

Real Garment
garments patterns

Garment fit Body 3D virtual try- Digital clothing

Real try-on evaluation results dimensions pressures
on software
Real people Fabric
properties

Learning database
A customer

3D virtual
Fabric Digital clothing Garment fit evaluation model
properties
try-on pressures based on Naïve Bayes
software
A new garment
pattern

Unfit fit
Garment fit levels: Purchase or
Fit or unfit Production

Fig. 1. General scheme of MLBGFET.

Input Garment fit prediction model

Clothing pressure p1new
Output
Clothing pressure P2new Machine learning-based garment fit
.. Garment fit level
. evaluation models
Clothing pressure Pknew FLę(0, 1)

Fig. 2. Modeling the relation between digital clothing pressures and garment fit.

3. Learning data acquisition the total female population of China [45]. The corresponding 3D
virtual body models used for virtual try-on are shown in Fig. 3.
3.1. Preparation work for experiments (Note: In China, female body shapes are classified in four cate-
gories (Y, A, B, C) according to the difference of bust-waist. The
We design Experiments I and II to collect data. Experiment I body shape belongs to the type Y if this value is located in the
aims to acquire output learning data on garment fit by using real range of 19–24 cm, the type A for the range of 14–18 cm, the type
try-on; Experiment II aims to acquire input learning data on digital B for the range of 9–13 cm, and the type C for the range of 4–8 cm.
clothing pressures by using virtual try-on. Anthropometric equip- 155/60A means that the body type is A, the stature 155 cm and the
ment, software, subjects, garments, fabrics, etc. involved in Experi- waist 71 cm).
ment I and II are expounded below respectively.

3.1.1. Anthropometric equipment

3.1.4. Garments
The Vitus Smart 3D body scanner is applied to collect human
72 pairs of straight pants, which cover most of pants’ sizes,
body dimensions for virtual try-on. This device captures body mea-
are involved in the real try-on experiments for data collection
surements with a ± 1 mm level of accuracy, in accordance with
(Table 1). We select the pant type to test our proposed method
the international standard DIN EN ISO 20,685.
because that they are the most challenging clothing item for a
good fit [2]. If the proposed model predicts pants’ fit accurately,
3.1.2. Software
this method could be also available for other styles.
The software CLO 3D is applied to measure digital clothing
pressures. This software permits to create virtual, close-to-life gar-
ment visualization with cutting-edge simulation technologies. Vir-
tual fabrics available in CLO 3D are based on actual fabrics com- 3.1.5. Fabric
monly used in the industry and they currently have a 95% accuracy Fabric physical properties influence digital clothing pressures
rate [44]. significantly. Therefore, they should be considered in the virtual
try-on experiment. However, these fabric properties, as well as gar-
3.1.3. Subjects ment styles, have already been taken into account in the digital
Nine female subjects with representative body shapes are se- clothing pressures measured in the 3D garment CAD environment.
lected for performing real try-on and body dimension measure- Therefore, we do not need to specially study the effects of fabric
ment. According to China National Standard (GBT 1335.2–2008), properties on garment fit. In Experiment II, we just selected a fre-
their body dimensions (155/60A, 155/62A, 160/64A, 160/66A, quently used jeans fabric with the mechanical properties shown in
160/68A, 165/70A, 165/72A, 170/74A and 170/76A) can account for Table 2 for making different garments of virtual try-on.
 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182 177

Fig. 3. Nine virtual bodies used for virtual try-on generated based on the nine subjects.

Table 1
Waist girths, hip girths of 72 pairs of pants used in the experiments (unit: cm).

60.0/75.5 62.5/78.0 65.0/80.5 67.5/83.0 70.0/85.5 72.5/88.0 75.0/90.5 77.5/93.0 80.0/95.5

60.0/78.0 62.5/80.5 65.0/83.0 67.5/85.5 70.0/88.0 72.5/90.5 75.0/93.0 77.5/95.5 80.0/98.0

60.0/80.5 62.5/83.0 65.0/85.5 67.5/88.0 70.0/90.5 72.5/93.0 75.0/95.5 77.5/98.0 80.0/100.5
60.0/83.0 62.5/85.5 65.0/88.0 67.5/90.5 70.0/93.0 72.5/95.5 75.0/98.0 77.5/100.5 80.0/103.0
60.0/85.5 62.5/88.0 65.0/90.5 67.5/93.0 70.0/95.5 72.5/98.0 75.0/100.5 77.5/103.0 80.0/105.5
60.0/88.0 62.5/90.5 65.0/93.0 67.5/95.5 70.0/98.0 72.5/100.5 75.0/103.0 77.5/105.5 80.0/108.0
60.0/90.5 62.5/93.0 65.0/95.5 67.5/98.0 70.0/100.5 72.5/103.0 75.0/105.5 77.5/108.0 80.0/110.5
60.0/93.0 62.5/95.5 65.0/98.0 67.5/100.5 70.0/103.0 72.5/105.5 75.0/108.0 77.5/110.5 80.0/113.0

Note: 60/67.5 means that the garment’s waist girth is 60 cm, whose hip girth is 75.5 cm; 60/78.0 means that the garment’s
waist girth is 60 cm, whose hip girth is 78.0 cm; and so on.

Table 2
Values of the fabric mechanical properties.

Abbr. BST BSP BRT BRP ST SW BT BP SH DE ID FC

Value 30 30 50 50 32 32 35 35 23 35 1 3

Note: BST is buckling stiffness-weft; BSP is buckling stiffness-warp; BRT is buckling ratio-
weft; BRP is buckling ratio-warp; ST is stretch-weft; SW is stretch-warp; BT is bending-
weft; BP is bending-warp; SH is shear; DE is density; ID is internal damping; FC is friction
coefficient. The fabric mechanical properties are relative values in the range [1–99], defined
by the software.

3.1.6. Garment fit level pants is selected by only one subject. The selected results
In this research, we classify all garment fit values into two lev- are shown in Table 3.
els fit or unfit. These fit levels are used in both real and virtual Step 2: Each subject realizes her try-on with the selected pants
garment try-on. by performing a number of gestures: sitting down, standing,
squatting, running and walking (See Fig. 4). After that, she
3.1.7. Try on condition gives an overall fit level of the evaluated pants using one of
Before each real try-on evaluation, each subject wears a piece the two scores (fit or unfit).
of underwear that is thin and neither tight nor loose. The try-on
experiment is carried out indoor under a temperature of 18–20 °C. Finally, the nine subjects evaluate the fit levels of all the 72
pairs of pants. According to these evaluation results, the 72 pairs
3.2. Experiment I: acquisition of the data on garment fit of pants are classified into the set of fit pants (23 pairs) and the
set of unfit pants (49 pairs) (see Table 4). The data will be taken
Experiment I is designed to evaluate garment fit levels using as input learning data to build the proposed models.
the real try-on. 72 pairs of pants are made according to the sizes in
Table 1. The experiment procedure is shown in Fig. 4. Nine selected 3.3. Experiment II: acquisition of the data on digital clothing
female subjects with different body shapes (Fig. 3) participate in pressures
the garment fit evaluation procedure. The details are given below.
We design Experiment II to measure the digital clothing pres-
Step 1: Each of the nine subjects selects eight pants from the sures at the key positions of the garment surface using the CLO
72 pairs of real pants according to her personal preference, 3D software (Fig. 5). The concrete scheme of Experiment II is de-
like what she usually does in a garment shop. One pair of scribed as follows.
178 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182

Fit level = 0
garments
Real garments Fit level
(72 pairs of pants) 0-unfit, 1-fit

sit stand squat walk run Fit level = 1

garments

Fig. 4. Garment fit data collection by real try on (output learning data).

Table 3
Size distribution of 72 pairs of pants selected by the nine subjects.

155/60A 155/62A 160/64A 160/66A 160/68A 165/70A 165/72A 170/74A 170/76A

60.0/78.0 62.5/78.0 60.0/75.5 60.0/80.5 60/83.0 67.5/95.5 70.0/85.5 72.5/105.5 72.5/93.0

60.0/88.0 62.5/83.0 60.0/85.5 65.0/80.5 62.5/85.5 70.0/93.0 70.0/100.5 75.0/90.5 77.5/100.5
62.5/80.5 60.0/90.5 62.5/90.5 67.5/85.5 65.0/98.0 72.5/95.5 72.5/88.0 75.0/103.0 77.5/105.5
62.5/88.0 60.0/93.0 62.5/95.5 70.0/95.5 65.0/93.0 72.5/98.0 72.5/90.5 75.0/105.5 77.5/110.5
62.5/93.0 65.0/88.0 65.0/90.5 70.0/98.0 67.5/83.0 75.0/95.5 75.0/108.0 77.5/93.0 80.0/105.5
65.0/83.0 67.5/90.5 65.0/85.5 72.5/100.5 67.5/100.5 75.0/98.0 77.5/103.0 77.5/95.5 80.0/108.0
65.0/95.5 67.5/98.0 67.5/93.0 75.0/93.0 70.0/88.0 77.5/98.0 80.0/95.5 80.0/100.5 80.0/110.5
67.5/88.0 70.0/103.0 72.5/103.0 75.0/100.5 70.0/90.5 77.5/108.0 80.0/98.0 80.0/103.0 80.0/113.0

Note: each column represents eight pairs of pants selected by one subject.

Fig. 5. Digital clothing pressure measurement by virtual try-on (input learning data).

Table 4 Step 4: We measure the digital clothing pressures of each pair

Garment fitness data collected by real try-on (output learning data).
of pants on predefined 20 key positions of each garment
FL Fit pants Unfit pants during its virtual try-on (Fig. 5(c)).
1 1 1 1 0 0 0 0 0
The digital clothing pressure data of the 72 pairs of pants are
Sample No. 1 2 3  23 24 25 26  72 shown in Table 5. The corresponding data (input data) will be com-
bined with the data of garment fit evaluation, collected in Section
5.2.2 (output data), for building the fit prediction models in Fig. 2.

4. Modeling the relation between clothing pressures and

Step 1: We built nine 3D human models whose body dimen- garment fit level
sions are equal to those of the nine subjects (see Fig. 3).
Step 2: We determine key positions F1, F2, …, F15 and B1, B2, As an effective tool for modeling with data learning, the Naive
…, B5 of each pair of pants, which are uniformly distributed Bayes classifier is used in our approach for constructing the gar-
on the front piece pattern and on the back piece pattern re- ment fit evaluation model. We present the specific modeling pro-
spectively (Fig. 5(a)). As the parts below knee have little ef- cedure in Fig. 6. It is composed of the following five steps:
fect on clothing fit, we do not define any key positions on
them. Step 1: Determining the characteristics of attributes.
Step 3: According to Table 3, we make virtual try-on with the In the procedure of modeling, the digital clothing pressures
patterns of the 72 pairs of pants on the 3D human models measures on the k predefined key positions on the garment
corresponding to the body dimensions of nine subjects pre- are taken as the characteristics attributes of the model. Ac-
viously selected (Figs. 5(b) and 3). cording to the general principle of Naive Bayes, we suppose
 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182 179

Table 5
Digital clothing pressures data collected by virtual try-on (input learning data).

Sample No. Digital clothing pressures (unit: KPa).

F1 F2 F3 F4 F5 F6 F7 F8 F9  B5

1 6.92 9.57 12.34 3.22 4.35 5.66 7.82 4.35 3.19  1.67
2 6.58 8.98 14.74 3.12 6.03 6.35 7.39 5.41 2.64  1.15
3 7.58 9.12 8.12 1.78 5.37 7.30 13.33 4.95 1.69  0.47
4 10.62 13.12 12.80 3.02 5.95 5.23 9.76 5.39 3.68  1.15
5 10.01 4.65 13.47 3.08 5.54 8.68 10.87 5.06 3.65  0.33
6 9.27 11.53 12.31 3.29 6.10 7.23 7.92 5.46 2.59  0.21
7 5.05 7.05 14.47 2.83 4.77 5.86 10.32 4.44 2.37  0.94
8 13.31 11.80 30.57 2.43 4.23 8.86 23.15 4.37 1.24  0.50
9 4.83 8.92 10.85 2.38 4.11 6.19 11.18 4.72 2.67  1.02
          
72 8.00 30.81 36.80 36.99 20.15 25.20 18.77 58.13 15.70  15.13

Note: The numbers of the samples are the same as those of Table 4; F1, F2,…, B5 refer to the key positions of the
digital clothing pressures, see Fig. 5(a).

Fig. 6. Modeling with Naive Bayes.

that these k characteristics attributes are independent each Table 6

Prediction accuracies calculated by the k-fold cross validation (k = 10).
other and all respect normal distributions.
Step 2: Acquiring training samples. Algorithms Prediction accuracy
Two experiments I and II are carried out to collect training Digital clothing Pressures Ease allowance
data by real and virtual try-on. The input and output train-
Naive Bayes 93.1% 76.4%
ing data are digital clothing pressures and garment fit levels SVMs 84.7% 77.8%
respectively.
Step 3: Computing the prior probabilities of each category.
We have two levels of garment fit (categories). Thus, the
prior probability of each category i (i ∈ {1, 0} ) can be:
5. Model validation
the number o f f it evaluations corresponding to the i − th level
P ( F Li ) =
the total number o f f it evaluations In this section, we compare the performances of the proposed
Step 4: Computing the conditional probability of the new sam- model. The input and output data for setting up the model are
ple Pnew . found in Tables 4 and 5 respectively. As the number of learning

k   data is very limited in this research, which may cause ‘‘over fitting’’
P (Pnew |F L = i) = j
P pnew /F L = i (i ∈ {1, 0} ) problem, leading to a very unstable performance of the model out-
j=1 put, we apply K-fold cross validation approach to calculate the pre-
P (F L = i )P (Pnew |F L = i) diction accuracies of the proposed model. To compare with other
P (F L = i|Pnew ) = 5 (i ∈ {1, 0}) machine learning algorithms, we also calculate the prediction ac-
i=1 P (F L = i )P (Pnew |F L = i ) curacies of the SVMs-based and ease allowance-based garment fit
Step 5: Predicting with Naive Bayes classifier. evaluation models. According to the defination of ease allowance,
The classification rule of the Naive Bayes classifier is given the waist’s ease allowace equals to garment’s waist girth minus hu-
below. man body’s wiast girth; the hip’s ease allowace equals to garment’s
If P (F L = l |Pnew ) = max {P (F L = i|Pnew )}, (l ∈ {1, 0}), then hip girth minus human body’s hip girth.
i=1 or 0
the new sample Pnew corresponds to the fit level l. As shown in Table 6, the prediction accuracy of Naive Bayes
model (93.1%) is better than that of SVMs model (84.7%) with se-
180 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182

Fig. 7. Process 4: remote garment fit prediction for online shopping.

lected digital clothing pressure as a fit evaluation index. The pre- posed model are both based on digital clothing pressures and do
diction accuracy of Naive Bayes model (76.4%) is slightly worse not deal with any real clothing pressures, we do not need to pre-
than that of SVMs model (77.8%) with selected ease allowance as a cisely identify real clothing pressures. Even if there are some dif-
fit evaluation index. ferences between digital and real clothing pressures, the prediction
accuracies of the proposed models are not affected.

6. Discussion
6.2. Application prospect
6.1. Influence of the difference between real and digital pressures on
the prediction results With increasing online sales, the fit of garments has serious im-
plications for a fashion retailer because ill-fitting garments are di-
In the introduction, we have pointed out that the ease al- rectly related to product return rates [49,50]. The evaluation of gar-
lowance between a garment and the human body is not a good in- ment fit without the physical participation of customers and de-
dicator of reflecting garment fit. Therefore, we opted to find a more signers is very useful for online clothing shoppers. In this context,
suitable indicator. The influence of fabric properties can be mea- we introduce an application based on the proposed method to pre-
sured using the digital clothing pressures distributed on the gar- dict garment fit in an e-shopping environment (Fig. 7).
ment surface covering the human body of the wearer. These digi- For a specific customer, a parametric human model is used to
tal clothing pressures are easily measured in a garment CAD soft- be adapted or adjusted to the real dimensions of the concerned
ware environment like CLO 3D. Our previous research shows that human body (Fig. 8(a) and (b)). Next, we search for the garment
the digital clothing pressures can reflect garment wear comfort ac- patterns from the database of the company according to the pre-
curately [39]. It is for this reason that we select the digital cloth- vious body dimensions (Fig. 8(c)). Next, a number of red points
ing pressures as a key indicator for performing remote garment fit will be marked on the selected patterns in order to measure the
prediction without real try-on. The test result indicates the digital clothing pressures at these key positions (Fig. 8(d)). Then, gar-
pressure-based methods are more efficient in fit evaluation than ment patterns are assembled on the adjusted digital human model
the ease allowance-based methods, which can be adapted to loose (Fig. 8(e)). Next, the assembled patterns are seamed together to
garments only instead of tight ones. form a 3D virtual garment (Fig. 8(f)). Finally, digital clothing pres-
The proposed models enable to set up accurate and quantitative sures are measured on the predefined key positions (Fig. 8(g)).
relations between digital clothing pressure data measured during a Having performed the previous operations, the collected digi-
virtual try-on and garment fit data evaluated during a real try-on. tal clothing pressures are introduced to the garment fit evaluation
For a new garment with an unknown fit level, we can measure model (see the red wireframe in Fig. 7) for predicting the garment
its digital clothing pressures and then apply a previously proposed fit automatically. If the predicted result meets the customer’s re-
model for predicting its fit according to the measured digital cloth- quirement, we recommend the concerned customer to buy the gar-
ing pressures. These models are significant and can accurately re- ment. Otherwise, she/he will be invited to try another one with a
flect comfort feeling of garments with different fabric mechanical different size or style. This procedure repeats until the satisfaction
properties because the digital and real clothing pressures not only of the result.
have the same variation trends (i.e., the digital clothing pressure
at a position is high when a subject feels tight at the same posi- 6.3. Limitation and future research
tion, and vice versa.) [5,39,40,46–48], but also are rather close each
other in a certain range. As the learning and prediction of the pro- The limitations and future research are summarized, as follows:
 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182 181

Fig. 8. 3D human model adjustment and virtual try-on.

1) In order to get reliable data to train the proposed method, gar- scenario. More data should be collected in the future practical ap-
ments with different sizes and styles need to be made firstly. As plications.
garment patterns are the business secret for fashion companies,
we needed to make garment patterns and real garments by our- Acknowledgments
selves. Due to this reason, we only collected a small dataset
with 72 samples to train the proposed model. Thus, further re- This paper was financially supported by China Scholarship
search can be combined with a specific garment company to Council (No. 201406630077) and the Fundamental Research Funds
train the proposed model using their existing clothing. As gar- for the Central Universities (No.CUSF-DH-D-2015082).
ment patterns are the business secret, our proposed approach
might be only suitable for companies do both production and References
sale of garments by themselves.
[1] E. Young Kim, Y.K. Kim, Predicting online purchase intentions for clothing
2) The two fit levels (fit and unfit) are too simplified. For exam- products, Eur. J. Market. 38 (2004) 883–897.
ple, during the real try-on, we only evaluate the overall fit level [2] H.K. Song, S.P. Ashdown, Investigation of the validity of 3-D virtual fitting for
for all the gestures and all the positions. In fact, more accu- pants, Cloth. Textiles Res. J. 33 (2015) 314–330.
[3] J. Kim, S. Forsythe, Adoption of virtual try-on technology for online apparel
rate results can be obtained if we propose to evaluate a series
shopping, J. Interact. Mark. 22 (2008) 45–59.
of local fit levels (hip fit level, waist fit level, etc.) each corre- [4] A.S.M. Sayem, R. Kennon, N. Clarke, 3D CAD systems for the clothing industry,
sponding to one body position of the wearer and then properly Intl. J. Fashion Des. Tech. Educ. 3 (2010) 45–53.
aggregate them for generating an overall fit level. In this situ- [5] X. Zhang, K. Yeung, Y. Li, Numerical simulation of 3D dynamic garment pres-
sure, Text. Res. J. 72 (2002) 245–252.
ation, all local discomfort feeling can also be considered in the [6] K. Liu, J. Wang, X. Zeng, X. Tao, P. Bruniaux, Fuzzy classification of young
fit prediction models in the further research. women’s lower body based on anthropometric measurement, Int. J. Ind. Er-
3) Digital clothing pressure is selected as the index of garment fit gonom. 55 (2016) 60–68.
[7] E. Shin, F. Baytar, Apparel fit and size concerns and intentions to use virtual
evaluation. However, it is possible that some parts of a loose try-on: impacts of body satisfaction and images of models’ bodies, Cloth. Tex-
garment are not in contact with the human body and the cor- tiles Res. J. 32 (2013) 20–33.
responding clothing pressures could be near zero. In further re- [8] J. Fan, W. Yu, L. Hunter, Clothing Appearance and Fit: Science and Technology,
Woodhead publishing Limited, Cambridge, UK, 2004.
search, ease allowance and clothing pressure can be combined [9] H.K. Song, S.P. Ashdown, Development of automated custom-made pants
together for evaluating garment fit in a complementary way. driven by body shape, Cloth. Textiles Res. J. 30 (2012) 315–329.
4) The used digital clothing pressures are static values measured [10] E. Kamalha, Y. Zeng, J.I. Mwasiagi, S. Kyatuheire, The comfort dimension; a re-
view of perception in clothing, J. Sens. Stud. 28 (2013) 423–444.
at different key positions related to a given gesture. The dy- [11] X. Tao, P. Bruniaux, Toward advanced three-dimensional modeling of gar-
namic aspect, i.e. the clothing pressures varying with time dur- ment prototype from draping technique, Int. J. Cloth. Sci. Technol. 25 (2013)
ing a movement is not considered. We need to apply time se- 266–283.
[12] X. Chen, X. Tao, X. Zeng, L. Koehl, J. Boulenguez-Phippen, Control and opti-
ries analysis to study these clothing pressures and form new
mization of human perception on virtual garment products by learning from
input variables of the fit evaluation model in further research. experimental data, Knowl. Based Syst. 87 (2015) 92–101.
[13] S. Thomassey, P. Bruniaux, A template of ease allowance for garments based
on a 3D reverse methodology, Int. J. Ind. Ergonom. 43 (2013) 406–416.
[14] J. Lee, Y. Nam, M.H. Cui, K.M. Choi, Y.L. Choi, Fit evaluation of 3D virtual gar-
7. Conclusion ment, in: N. Aykin (Ed.), Usability and Internationalization. HCI and Culture,
Springer, Berlin Heidelberg, Berlin, Heidelberg, 2007, pp. 550–558.
[15] Z. Guo, W. Wong, S. Leung, M. Li, Applications of artificial intelligence in the
In this research, we proposed a machine-learning-based model apparel industry: a review, Text. Res. J. 81 (2011) 1871–1892.
to predict garment fit. The results indicate that: 1) digital clothing [16] K. Liu, J. Wang, E. Kamalha, V. Li, X. Zeng, Construction of a body dimensions’
prediction model for garment pattern making based on anthropometric data
pressure is a better garment evaluation index than ease allowance;
learning, J. Text. Inst. (2017). https://doi.org/10.1080/0 04050 0 0.2017.1315794.
2) Naive Bayes is a good classifier and not inferior to other classi- [17] X. Zeng, L. Koehl, Representation of the subjective evaluation of the fabric hand
fiers in the field of garment fit evaluation, and even better than using fuzzy techniques, Int. J. Intell. Syst. 18 (2003) 355–366.
[18] X. Zeng, Z. Liu, A learning automata based algorithm for optimization of con-
SVMs in some cases. Compared with the traditional garment fit
tinuous complex functions, Inform. Sciences 174 (2005) 165–175.
evaluation methods, the proposed approach has a number of ad- [19] X. Zeng, D. Ruan, L. Koehl, Intelligent sensory evaluation: concepts, implemen-
vantages: 1) continuous improvement of the model’s performance tations, and applications, Math. Comput. Simulat. 77 (2008) 443–452.
with new learning data, 2) independence of any real try-on, 3) re- [20] A. Wong, Y. Li, P. Yeung, Predicting clothing sensory comfort with artificial in-
telligence hybrid models, Text. Res. J. 74 (2004) 13–19.
moval of human involvement. Due to the dataset is very small in [21] H.-S. Kim, S.-B. Cho, Application of interactive genetic algorithm to fashion de-
this research; the approach might not be easy to use in a real-life sign, Eng. Appl. Artif. Intel. 13 (20 0 0) 635–644.
182 K. Liu et al. / Knowledge-Based Systems 133 (2017) 174–182

[22] A. Vuruskan, T. Ince, E. Bulgun, C. Guzelis, Intelligent fashion styling using [35] M. Xia, Y. Zhang, L. Weng, X. Ye, Fashion retailing forecasting based on extreme
genetic search and neural classification, Int. J. Cloth. Sci. Technol. 27 (2015) learning machine with adaptive metrics of inputs, Knowl. Based Syst. 36 (2012)
283–301. 253–259.
[23] L.C. Wang, X.Y. Zeng, L. Koehl, Y. Chen, Intelligent fashion recommender sys- [36] M. Xia, W.K. Wong, A seasonal discrete grey forecasting model for fashion re-
tem: fuzzy logic in personalized garment design, IEEE Trans. Human Mach. tailing, Knowl. Based Syst. 57 (2014) 119–126.
Syst 45 (2015) 95–109. [37] W.-S. Lo, T.-P. Hong, R. Jeng, A framework of E-SCM multi-agent systems in the
[24] Z.-H. Hu, Y.-S. Ding, W.-B. Zhang, Q. Yan, An interactive co-evolutionary CAD fashion industry, Int. J. Prod. Econ. 114 (2008) 594–614.
system for garment pattern design, Comput. Aid. Des. 40 (2008) 1094–1104. [38] A. Pan, S. Leung, K. Moon, K. Yeung, Optimal reorder decision-making in the
[25] W. Wong, S. Leung, K. Au, Real-time GA-based rescheduling approach for the agent-based apparel supply chain, Expert Syst. Appl. 36 (2009) 8571–8581.
pre-sewing stage of an apparel manufacturing process, Int. J. Adv. Manuf. Tech. [39] K. Liu, E. Kamalha, J. Wang, T.-K. Agrawal, Optimization design of cycling
25 (2005) 180–188. clothes’ patterns based on digital clothing pressures, Fiber. Polym. 17 (2016)
[26] Z.X. Guo, W.K. Wong, S.Y.S. Leung, J.T. Fan, S.F. Chan, Genetic optimization 1522–1529.
of order scheduling with multiple uncertainties, Expert Syst. Appl. 35 (2008) [40] K. Liu, J. Wang, C. Zhu, Y. Hong, Development of upper cycling clothes using
1788–1801. 3D-to-2D flattening technology and evaluation of dynamic wear comfort from
[27] M.-T. Lin, The single-row machine layout problem in apparel manufacturing by the aspect of clothing pressure, Int. J. Cloth. Sci. Technol. 28 (2016) 736–749.
hierarchical order-based genetic algorithm, Int. J. Cloth. Sci. Technol. 21 (2009) [41] K. Liu, J. Wang, Y. Hong, Wearing comfort analysis from aspect of numerical
31–43. garment pressure using 3D virtual-reality and data mining technology, Int. J.
[28] P.C.L. Hui, K.C.C. Chan, K.W. Yeung, F.S.F. Ng, Fuzzy operator allocation for bal- Cloth. Sci. Technol. 29 (2017) 166–179.
ance control of assembly lines in apparel manufacturing, IEEE T. Eng. Manage. [42] P. Langley, W. Iba, K. Thompson, An analysis of Bayesian classifiers, in:
49 (2002) 173–180. C.A.A.f.A. Intelligence (Ed.), Proceedings of the Tenth National Conference on
[29] Z.X. Guo, W.K. Wong, S.Y.S. Leung, J.T. Fan, Intelligent production control de- Artificial Intelligence, AAAI Press, Menlo Park, CA, 1992, pp. 223–228.
cision support system for flexible assembly lines, Expert Syst. Appl. 36 (2009) [43] P. Domingos, M. Pazzani, On the optimality of the simple bayesian classifier
4268–4277. under zero-one loss, Mach. Learn. 29 (1997) 103–130.
[30] C. Çiflikli, E. Kahya-Özyirmidokuz, Implementing a data mining solution for [44] C.D. enterprice, CLO 3D, in, 2016.
enhancing carpet manufacturing productivity, Knowl. Based Syst. 23 (2010) [45] C.N.S.M. Committee, GBT 1335.2-2008, Standard Sizing Systems For Garments,
783–788. Standards Press of China, Beijing, 2008.
[31] Z.-L. Sun, T.-M. Choi, K.-F. Au, Y. Yu, Sales forecasting using extreme learning [46] H. Seo, S.-J. Kim, F. Cordier, K. Hong, Validating a cloth simulator for measuring
machine with applications in fashion retailing, Decis. Support. Syst. 46 (2008) tight-fit clothing pressure, in: Proceedings of the 2007 ACM Symposium on
411–419. Solid and Physical Modeling, ACM, Beijing, China, 2007, pp. 431–437.
[32] W. Wong, Z. Guo, A hybrid intelligent model for medium-term sales forecast- [47] L. Yanmei, Z. Weiwei, J. Fan, H. Qingyun, Study on clothing pressure distribu-
ing in fashion retail supply chains using extreme learning machine and har- tion of calf based on finite element method, J. Text. Inst. 105 (2014) 955–961.
mony search algorithm, Int. J. Prod. Econ. 128 (2010) 614–624. [48] M. Zhang, H. Dong, X. Fan, R. Dan, Finite element simulation on clothing pres-
[33] W.K. Wong, X. Zeng, W. Au, P. Mok, S. Leung, A fashion mix-and-match expert sure and body deformation of the top part of men’s socks using curve fitting
system for fashion retailers using fuzzy screening approach, Expert Syst. Appl. equations, Int. J. Cloth. Sci. Technol. 27 (2015) 207–220.
36 (2009) 1750–1764. [49] K. DesMarteau, CAD: let the fit revolution begin, Bobbin 42 (20 0 0) 42–56.
[34] S. Choi, Y. Yang, B. Yang, H. Cheung, Item-level RFID for enhancement of cus- [50] H. Kim, M.L. Damhorst, The relationship of body-related self-discrepancy to
tomer shopping experience in apparel retail, Comput. Ind. 71 (2015) 10–23. body dissatisfaction, apparel involvement, concerns with fit and size of gar-
ments, and purchase intentions in online apparel shopping, Cloth. Textiles Res.
J. 28 (2010) 239–254.

View publication stats