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

net/publication/265105658

Developing Body Measurement Charts for Garment Manufacture Based on a Linear

Programming Approach

Article in Journal of Textile and Apparel, Technology and Management · January 2006

CITATIONS READS

7 295

4 authors, including:

Deepti Gupta Naveen Garg

Indian Institute of Technology Delhi Indian Institute of Technology Delhi
105 PUBLICATIONS 1,152 CITATIONS 77 PUBLICATIONS 4,210 CITATIONS

SEE PROFILE SEE PROFILE

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

Sericin based finishing of textiles View project

sericin based finishes View project

All content following this page was uploaded by Naveen Garg on 06 November 2014.

The user has requested enhancement of the downloaded file.

 Volume 5, Issue 1, Winter 2006

Developing Body Measurement Charts for Garment Manufacture Based on a Linear

Programming Approach.

Deepti Gupta, Department of Textile Technology, Indian Institute of Technology, Hauz Khas,
New Delhi –110016, India.
Naveen Garg, Department of Computer Science and Engineering, Indian Institute of Technology,
Hauz Khas, New Delhi –110016, India.
Komal Arora and Neha Priyadarshini, Department of Computer Science and Engineering, Indian
Institute of Technology, Hauz Khas, New Delhi –110016, India.

ABSTRACT

The process of developing body size charts for a given population is a highly complex one as too
many variables are involved. The requirements are often contradictory as in trying to provide the
best fit using a minimum number of sizes. With the availability of advanced mathematical tools it
is now possible to address the issue as an optimization problem. In the present study, an
algorithm based on the Linear Programming approach has been developed specifically to cluster
a given population data into homogenous body size groups. The theoretical efficiency of the
approach has been demonstrated on an anthropometric database of 1900 young Indian women.
The mathematical tool developed is flexible enough to be adapted for use for mass production as
well as mass customization of garments. It is extremely versatile in that garment specific size
tables can be developed. The degree of fit desired at each body dimension as well as the body
dimensions used as the basis of clustering can be changed with ease. It is also a great tool for
inventory management as it gives exactly the number of people covered by each cluster thus
giving the manufacturer and retailer the choice of deciding how many pieces to make in each
style and in what sizes.

Keywords: Body Measurements, CAD, Garment Fit, Garment Sizing, Linear Programming,
Optimization.

Introduction measurements. Further work is ongoing for

developing a user friendly software based on
The process of developing body size charts the algorithm, for use by the garment
for garment manufacturers is a very complex manufacturing industry of India.
one. In this paper, the development and
testing of a novel mathematical solution Background
based on the LP approach has been
proposed. The results have been extensively A “size” is an item having specif ied
validated mathematically, using the measurements along certain dimensions,
anthropometric database of a group of young such that it will fit perfectly a person with
Indian men and women. However, wearer measurements equal to that size (Tryfos,
trials need to be conducted for converting 1986). The purpose of an apparel sizing
these body measurements into garment system is to divide a varied population into
Article Designation: Refereed 1 JTATM
Volume 5, Issue 1,Winter 2006
homogeneous subgroups. Members of a is then to choose the sizes in order to
subgroup are similar to each other in body optimize sales of the garment. The problem
size and shape so that a single garment can is formulated as a “p median” or “Facility
adequately fit each of them. Members of Location problem.”
different subgroups are dissimilar and would
therefore require different garments A novel approach for the construction
(Ashdown & Delong, 1995). Fit of a of apparel sizing systems has been proposed
garment depends on the correlation between by McCulloch (McCulloch et al.1998). The
garment measurements and the body concept of garment fit is captured by a
measurements for which it is intended. In distance measure, which is calculated from
general, very little correlation exists among the discrepancies between the body
the human body measurements. Along with measurements of an individual in the sample
the large number of relevant body and the prototype design values of a size.
dimensions, the body proportions can also Using this measure, known as aggregate
vary enormously. This is obvious from the loss, various existing sizing systems were
variety of body shapes that can be seen in a compared (Ashdown, 1998). Non-linear
group of people. Body size as well as body optimization techniques were used to derive
shape can be a characteristic of ethnic a set of possible sizing systems using multi-
groups. Classifying a population into dimensional information from
homogeneous body sizes is hence a highly anthropometric data. Results showed that
complex problem. endogenous size assignment and selection of
disaccommodated individuals together with
The last decade has seen a remarkable relaxation of the requirement of a
surge in the number of studies reported on “Stepwise” size structure resulted in
the subject from all over the world substantial improvements in fit over an
(McCulloch et al, 1998, Rifkin, 1994). existing sizing system.
Various statistical methods ranging from
simple percentiles to complex combinations Several approaches have been tried
of multivariate and regression analyses have and tested as reported but what the industry
been employed for distribution of population is really looking for is a tool which can yield
into subgroups. More recently, powerful accurate, flexible and quick solutions to the
mathematical techniques have been extremely complex problem of what sizes to
employed with good results. make and how many garments to make in
each size. These values will vary from one
Gupta & Gangadhar, (2004) used geographical location to another even for the
Principal Component Analysis to identify same retail chain. Customized size charts are
the key body measurements which can form thus required for each location to maximize
the basis for classifying a population data the fit of garments for the intended clients
set. Body shapes and their distribution and minimize inventories.
within the population were identified.
Validation of size charts was achieved by It is obvious that linearly graded
calculating the aggregate loss of fit. sizing systems that range from very small to
very large cannot cover a diverse population
Tryfos, (1986) has suggested an adequately. Human bodies come in all
integer programming approach to optimize shapes and sizes and it is not possible to
the number of sizes so as to maximize divide them on the basis of simple averages
expected sales or minimize an index of which lead to unrealistic sizes corresponding
aggregate discomfort. He divides the space to the so called “standard” or “ideal” body
of body dimensions artificially into a set of measurements. What is needed is a
discrete possible sizes. The probability of completely random system, which can yield
the sale of a garment from one company to a an optimum number of body sizes reflecting
person falling in another category is the true body measurements and proportions
modeled as a simple function of fit. The goal existing in the target population. The sizes
Article Designation: Refereed 2 JTATM
Volume 5, Issue 1,Winter 2006
thus obtained can be expected to provide the of a functio (x1 ,x2 ,…,xn), which furthermore
best possible fit for garments. satisfies a set of constraints, e.g.,
g(x1 ,x2 ,…,x n)>=b. Linear programming is
Furthermore, it would be the specialization of mathematical
advantageous to provide the users with the programming to the case where both,
ability to choose their target population. function f - to be called the objective
Thus if a certain business manufactures function - and the problem constraints are
clothes for young adults, it would be logical linear.
to do the clustering only for the data
corresponding to the target group in the An important factor for the
population rather than using the complete applicability of the mathematical
data set for adult women. programming methodology in various
application contexts, is the computational
All this necessitates the development of tractability of the resulting analytical
a clustering algorithm, which, since it would models. Under the advent of modern
have to be invoked frequently, has a small computing technology, this tractability
running time. All these requirements make requirement translates to the existence of
the clustering problem all the more effective and efficient algorithmic
challenging. Bearing these considerations in procedures which can provide a systematic
mind, the problem of sizing in the current and fast solution to these models.
study was modeled with the objective of
developing: The general form for a Linear Programming
problem is as follows:
a) A suitable clustering algorithm for
dividing population database into a Objective Function:
predetermined number of
homogenous size groups. max/min f(X 1 ,X2 ,…,Xn):= c1 X 1 + c2 X 2 +... cn
b) A system with an in-built flexible fit Xn
function to minimize the deviation (1)
of proposed measurements from the
actual body dimensions. Technological Constraints:
c) A system such that the largest
possible fraction of the population
can be accommodated in minimum
number of sizes.
d) A system which allows development
of sizes for sub groups of population (2)
with ease and accuracy.
Sign Restrictions:
The fit-function (i.e. the tolerance or
range) also depends to a great extent on the
garment chosen. For instance, if the garment
is a woman’s skirt with an elastic band at the (3)
waist then the fit function for the waist
measurement can be comfortably large. where “urs” implies unrestricted in sign.

Linear Programming The functions involved in the problem

objective and the left-hand-side of the
A Linear Programming problem is a technological constraints are linear. It is the
special case of Mathematical Programming assumptions implied by linearity that to a
problem. From an analytical perspective, a large extent determine the applicability of
mathematic al program tries to identify an the above model in real-world applications.
extreme (i.e., minimum or maximum) point
Article Designation: Refereed 3 JTATM
Volume 5, Issue 1,Winter 2006
 Another approximating element in proportion of the population, we developed
many real-life LP applications results from a greedy algorithm along the lines of the
the so called divisibility assumption. This classical algorithm for set-cover. Our
assumption refers to the fact that for LP algorithm is iterative in nature; in each step
theory and algorithms to work, the problem it picks that potential garment size which
variables must be real. However, in many “covers” the largest number of uncovered
LP formulations, meaningful values for the points. The algorithm stops when it has
levels of the activities involved can be only picked the number of sizes permitted.
integer. Introducing integrality requirements
for some of the variables in an LP We compared the efficiency of our
formulation turns the problem to one algorithm against the highest possible
belonging in the class of (Mixed) Integer number of individuals in which can be
Programming (MIP) which have been theoretically accommodated (theoretical
shown to belong to the notorious class of upper bound) in the chosen number of sizes.
NP-complete problems (problems that have These values were determined by solving an
been ``formally'' shown to be extremely existing linear program (using an LP
“hard” computationally). Given the Solver). We tested the algorithm on a data
increased difficulty of solving IP problems, set of 2000 Indian males and 1900 Indian
sometimes, in practice, near optimal females with 20 body measurements each
solutions are obtained by solving the LP and found that in most instances the
formulation resulting by relaxing the algorithm, which is very fast in practice,
integrality requirements - known as the LP came very close to the upper bound values
relaxation of the corresponding IP - and yielded by the LP solver.
(judiciously) rounding off the fractional
values for the integral varia bles in the Materials and Methods:
optimal solution.
Anthropometric Data
Specifically, we shall define as the
feasible region of the LP of Equations 1 A data set comprising of body
to 3, the entire set of vectors measurements for 1900 Indian women (18 to
35 years) has been used for the study. The
that satisfy the anthropometric survey was conducted on
technological constraints of Eq. 2 and the behalf of a leading garment brand of India
sign restrictions of Eq. 3. An optimal across six cities in India. Twenty
solution to the problem is any feasible measurements were taken for each
vector that further satisfies the optimality individual.
requirement expressed by Eq. 1.
The approach
The points covered by each data entry
taken as a potential garment size was Three approaches were used for
calculated. The linear problem was solved clustering the population. The approach 1
using GLPSOL LP-SOLVER to maximize yielded the theoretical optima possible for a
the number of points covered. However, the given set of variables. Approach 2 was the
solution to the linear program is a fractional algorithm specially written for this
solution that cannot be used directly to application and approach 3 was used to
obtain garment sizes. validate the results. These are described
below:
In this paper, we report the
development of an algorithm for finding the 1. An LP solver- GLPSOL was used
body sizes so as to accommodate the to determine the maximum
maximum number of individuals under the (theoretical) number of points that
constraints imposed by the fitness function. could be covered under a given
To obtain garment sizes which cover a large number (say n) f sizes.
Article Designation: Refereed 4 JTATM
Volume 5, Issue 1,Winter 2006
 2. Algorithm development: Given an charts, a judicious combination of the girth
arbitrary fit function, an algorithm and length measures is thus required. It
was written to cluster the population determined that better fit could be obtained
into a predetermined number of in garments, if the size classification was
sizes (1-20) so as to maximize the done separately for the top body and the
number of individual lower body. For the purpose of analysis,
accommodated. lesser is the number of body measurements
3. Grid Approach: In this system taken for clustering, larger would be the
each point was considered in a grid population covered under each size.
of predetermined granularity as a
potential garment size and the The algorithm allows the use of any
number of points in the data set selected measurements as the basis of
covered by that size was calculated. clustering. In the present work clustering for
The process was repeated to get the the lower body garments was done on the
required number of sizes. basis of hip, waist (girth) and outer leg
length (length) measures. Clustering for top
The three approaches were tested on body garments was done on the basis of bust
women’s’ top and lower body (girth), cervical height and waist from
measurements. Fit functions were changed Centre front (length measures). Size charts
to see how the coverage of data was affected were generated using four different fit
in each case. functions namely (i) +1, -1,” (ii) +1, -0.5,”
(iii) +1.5, - 0.5” and (iv) +2,- 1.” In each
Results and discussion case, output was taken using all the three
approaches mentioned above. Results
Multiple correlation analysis of the data obtained are reported and discussed below.
conducted earlier (Gupta and Gangadhar,
2004) showed that there was poor Case 1. Fit Function = +1, -1” (i.e.
correlation between the length and girth point * is covered by the size if size-1<
measurements. However, good correlation point <size+1 for all dimensions). Table I
existed between the major girth dimensions shows the size chart obtained using fit
namely bust, hip and waist and within the function of + 1.”
length measures. For developing body size

Article Designation: Refereed 5 JTATM

Volume 5, Issue 1,Winter 2006
 Table I. Size chart for 20 sizes using fit function of + 1”

Body measurements
Hip Waist Length
No. of sizes Points covered % Cover
1 30 22 60 26 1.4

2 30 24 62 39 3.16

3 31 23 63 33 5.44

4 31 24 59 50 8.22

5 32 24 61 88 13.11

6 32 25 63 61 16.5

7 32 26 60 48 19.16

8 33 25 60 28 20.72

9 34 28 62 32 22.5

10 34 26 62 87 27.33

11 34 27 60 68 31.11

12 34 25 64 38 33.22

13 34 28 64 37 35.27

14 35 28 59 38 37.38

15 36 26 61 27 38.88

16 36 27 63 45 41.38

17 36 28 62 60 44.72

18 36 28 64 30 46.38

19 37 30 61 41 48.66

20 38 30 63 40 50.88

Total 916 51

Article Designation: Refereed 6 JTATM

Volume 5, Issue 1,Winter 2006
 It can be seen from Table 1 that if 39% are covered and with 20 sizes 51% of
there was to be only one size for the given the population was covered. Comparison of
data set, then only 26 (1.4%) persons (out of the results obtained by three approaches is
1900) would be covered if the fit function shown in Fig. 1. In all cases, the values
was defined as + 1.” However, as the obtained by approach 2 and 3 fall within 5-
number of sizes increases to 10, 27% of 10% of the theoretical upper bound value
population is covered, while with 15 sizes obtained by Approach 1.

Figure 1. Number of points covered by three approaches (FF=

+1")

1000

950

900
No of points covered

850

800

750

700

650 Approach1
Approach2
600
Approach3
550

500
10 11 12 13 14 15 16 17 18 19 20
No of sizes

Note: FF = Fit Function

Sizes in Table I have been arranged in refers to a person in the population) are
the ascending order of hip measurements. being catered. Difference in length
There are two size options each for women measurements is also reflected in the size
having hip measurement equal to 30 and 31” chart. For example, Size 12 is for a taller
and one size each or hip measure equal to woman having a narrower waist as
35, 37 and 38” respectively. For hip size 34” compared to size 11.
there are 4 options while for hip 36” there
are 5 options. The distribution clearly shows The 3D and 2D plots of the proposed
that unlike the empirical systems in practice, sizes superimposed over the actual
this system is completely random and truly population data are shown in Figures 2-4. In
reflects the distribution of measurements in all cases, the cover of population was seen
a given data set. To illustrate, look at the to be quite good. Data sets having extreme
five sizes having hip of 34.” In these sizes, values are not covered by the system.
the difference between the hip and waist Special sizes may have to be developed for
ranges from 6” to 9” meaning that all such people who have odd dimensions and
different body proportions corresponding to proportions.
different body shapes (*each point of data

Article Designation: Refereed 7 JTATM

Volume 5, Issue 1,Winter 2006
 Figure 2. 3D A plot showing the proposed sizes superimposed over the population
measurements for hip, waist and length.

Figure 3. 2D A plot showing the proposed size measurements for hip and waist
superimposed over the data measurements.

Article Designation: Refereed 8 JTATM

Volume 5, Issue 1,Winter 2006
 Figure 4. 2D A plot showing the proposed size measurements for hip and length
superimposed over the data measurements.

Case 2. Fit function = +1, -0.5, i.e. reduction in total cover with only 31% of the
point is covered by the size if size- population being covered under 20 sizes. A
0.5<point<size +1. The proposed size chart completely different set of sizes is obtained
is shown in Table II. As expected, the in this case as compared to Case 1. For
number of people covered by this function is example, the 4 sizes for hip 36” have an
less than that covered by Case 1. There is a identical waist measure of 27” but vary
consistent reduction of about 30-40% in significantly in the length measure.
cover for a given number of sizes and a 40%

Article Designation: Refereed 9 JTATM

Volume 5, Issue 1,Winter 2006
 Table II. Size chart for 20 sizes using fit function of + 1,” -0.5.”

Body measurements
Points % Cover
No. of sizes Hip Waist Length covered
1 30 24 61 23 1.27
2 30 22 60 23 2.55
3 31 24 59 28 4.11
4 31 24 61 45 6.61
5 32 24 62 41 8.88
6 33 26 60 32 10.66
7 33 25 63 37 12.72
8 33 24 60 21 13.88
9 33 25 62 45 16.38
10 34 27 62 25 17.77
11 34 24 63 19 18.83
12 34 27 60 36 20.83
13 35 26 63 29 22.44
14 35 29 60 18 23.44
15 36 27 65 20 24.55
16 36 27 60 22 25.77
17 36 27 62 29 27.38
18 36 27 63 19 28.44
19 37 30 61 24 29.77
20 37 29 62 21 30.94

Total 557 31

Case 3. Range = + 1.5, - 0.5, i.e. point previous cases. Hip measurements from 29”
is covered by the size, if size-0.5<point<size to 39” are covered, even though the total
+1.5 for all dimensions. Results are shown population covered is slightly less than that
in Table III. It is interesting to note that with in case 1. There is only one size for hip 36”
this fit function the overall range of hip in this case.
measurements covered is more than the

Article Designation: Refereed 10 JTATM

Volume 5, Issue 1,Winter 2006
 Table III. Size chart for 20 sizes using fit function of + 1.5,” -0.5.”

Body measurements % Cover

Points
No. of sizes Hip Waist Length covered
1 29 22 60 24 1.33
2 30 23 61 30 3
3 31 23 60 51 5.83
4 31 24 63 40 8.05
5 31 24 61 80 12.5
6 32 25 59 29 14.11
7 32 25 63 54 17.11
8 33 26 60 62 20.55
9 33 25 62 88 25.44
10 33 27 63 36 27.44
11 33 28 60 29 29.05
12 34 25 64 31 30.77
13 35 26 59 29 32.38
14 35 27 61 55 35.44
15 35 26 62 45 37.94
16 35 28 60 40 40.16
17 36 27 65 36 42.16
18 37 30 61 39 44.33
19 37 28 63 35 46.27
20 39 31 63 25 47.66
Total 858 47.7

Case 4 Range = +2, -1, i.e. point is improvement in the total covered under this
covered by the size if size-1<point<size+2 fit function with 81% of the population
for all dimensions. Results are reported in being covered. The range of measurements
Table IV. There is a significant covered is wider as expected.

Article Designation: Refereed 11 JTATM

Volume 5, Issue 1,Winter 2006
 Table IV. Size chart for 20 sizes using fit function of + 2,” - 1.”

Body measurements
Points % Cover
No. of sizes Hip Waist Length covered
1 29 22 59 33 1.8
2 30 23 63 59 4.5
3 30 23 61 151 12.9
4 30 24 57 43 15.3
5 31 26 61 29 16.9
6 32 25 58 92 22
7 32 24 64 99 27
8 32 25 61 224 40
9 33 24 61 37 42
10 34 27 63 48 44.7
11 34 27 60 100 50.2
12 34 25 59 26 51.7
13 35 27 62 161 60.6
14 35 27 65 77 64.9
15 36 29 59 51 67.7
16 37 30 62 102 73.4
17 38 28 60 24 74.7
18 38 29 63 27 76.2
19 39 32 63 33 78.1
20 40 31 61 50 80.8
Total 1466 81

Conclusions know in advance how many pieces to

manufacture and stock. These numbers
It has been possible to demonstrate would vary from one store location to other
through this work the potential of an and can be worked out exactly for each
algorithm developed by us for clustering a country/ state or location.
population into homogenous groups which
can be fitted within a pre-determined This is the first step in the process of
number of garment sizes. The fit function development of a garment sizing system for
for each measurement can be varied. It can the Indian population. The tool can be
be different for each measurement modified to suit the requirements of ready
depending upon the criticality of fit desired made garment mass production as well a
at a particular body area. It is possible to mass customization. Investigators are further
change the measurements and the number of collecting up-to-date anthropometric data an
measurements used for clustering. analyzing body shape and proportions in the
Indian population. Work is ongoing to
One of the strongest points in favor of determine the fit functions required for each
this system is that it tells exactly the number body measurement for each garment type.
of people falling under each size category. Fitness trials will be conducted to establish
This allows a retailer or a manufacturer to the fit preferences of Indian women leading
choose the sizes of interest to them and
Article Designation: Refereed 12 JTATM
Volume 5, Issue 1,Winter 2006
 to development of digitized pattern drafts to 3. Gupta D and Gangadhar B R,
suit the various Indian figure types. IJCST, Vol.16 (5) 2004, 458-469
4. McCulloch CE, Paal B. and
References Ashdown SP, J. Opl. Res. Soc., Vol.
49, p 492 (1998)
1. Ashdown S. P. and Delong M, App. 5. Rifkin G, The New York Times,
Ergo., Vol. 26 (1), p 47 (1995) Dec.8.1994, A1, D4.
2. Ashdown SP, IJCST, Vol. 10(5), p 6. Tryfos P, J.Opl Res. Soc. Vol.37
324 (1998) (10), p1001(1986).

Article Designation: Refereed 13 JTATM

Volume 5, Issue 1,Winter 2006

View publication stats