heuristic misfit reduction: a programmable approach for 3d

13
Computers & Graphics 71 (2018) 1–13 Contents lists available at ScienceDirect Computers & Graphics journal homepage: www.elsevier.com/locate/cag Technical Section Heuristic misfit reduction: A programmable approach for 3D garment fit customization Wonseop Lee, Hyeong-Seok Ko Department of Electrical and Computer Engineering, Seoul National University, Seoul, Republic of Korea a r t i c l e i n f o Article history: Received 2 May 2017 Revised 11 October 2017 Accepted 16 October 2017 Available online 6 November 2017 Keywords: Clothing simulation Fit customization Pattern-making Computer animation a b s t r a c t Based on the physically-based clothing simulation, this paper develops a novel method to customize the fit of the given garment to the reference body. The method defines three misfit measures, namely the landmark point misfit, landmark line misfit, and circumferential misfit, based on the correspondence be- tween the landmark points and landmark lines in the body and garment. In terms of the above mis- fit measures, this paper proposes the heuristic misfit reduction method. The heuristic misfit reduction method works in the following way. Starting from a given preliminary garment, it (1) performs the fit evaluation, (2) modifies the panels based on the evaluation to enhance the fit, then (3) performs the draping simulation with the modified panels. The procedure is repeated until a satisfactory fit is achieved. According to our experiments, the proposed method successfully customizes a given garment to fit to the given body. © 2017 Elsevier Ltd. All rights reserved. 1. Introduction Finding out whether a given garment fits to the body (we will call this part the fit evaluation) and being able to modify the panels to customize the fit (we will call this part the panel modification) are both fundamentally and practically important in the production of clothes. This paper is about customizing the fit of the given 3D garment to the given 3D avatar based on fit evaluation and panel modification. In the clothing field, several drafting schemes such as ESMOD [1], BUNKA [2] have long been established for producing gar- ments. From the given primary body sizes (e.g., waist back length, waist circumference, etc. for the case of the bodice), each drafting scheme specifies how to construct the panels that supposedly fit the target body. (In this work, we use the term panel to mean the sewing pattern.) When a given garment which has been constructed targeting a specific body A (Fig. 2(a)) needs to be customized to fit a different body B (Fig. 2(b)), by employing a drafting scheme, a preliminary panel resizing can be performed. In fact, we have programmed an automatic drafting scheme which takes the panels for A (Fig. 2(c)) and resizes them for B (Fig. 2(d)) by examining the differences in Corresponding author. E-mail addresses: [email protected] (W. Lee), [email protected] (H.-S. Ko). the two bodies. However, note that the drafting scheme provides just a general guideline. That is why we call it “preliminary”. With the same primary body sizes, the body can have different shapes. For example, the bust circumference does not tell the shape or volume of the bust itself. The particular (bent, straight, or cam- ber) shape of the torso is not reflected in the primary body sizes. Therefore a drafting scheme itself can not be regarded as a fit cus- tomization. This paper aims to develop a fit customization method using a physically-based clothing simulator. Fig. 1(a) shows the front and back panels used for the bodice, Fig. 1(b) shows their draping sim- ulation, and Fig. 1(c) shows the result after the fit is customized to the body with the proposed method. As the input, the proposed fit customization method takes (1) the 3D geometry of the avatar in a static pose and (2) the prelimi- nary 3D garment (i.e., the panels comprising the garment with the seams defined between the panels). It assumes that a number of landmark points and landmark lines are pre-marked in both the avatar body and the panels, as summarized in Appendices. The proposed method is basically a loop. It firstly drapes the given garment on the given body by running the clothing simula- tor. It analyzes the fit of the simulated results, makes necessary modifications to the panels (according to the method described subsequently), then the resultant garment is simulated again. The above are repeated until the desired fit is achieved. The purpose of this paper is to develop a method that can be potentially used for the tailoring task. This paper is not the first https://doi.org/10.1016/j.cag.2017.10.004 0097-8493/© 2017 Elsevier Ltd. All rights reserved.

Upload: others

Post on 26-Feb-2022

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Heuristic misfit reduction: A programmable approach for 3D

Computers & Graphics 71 (2018) 1–13

Contents lists available at ScienceDirect

Computers & Graphics

journal homepage: www.elsevier.com/locate/cag

Technical Section

Heuristic misfit reduction: A programmable approach for 3D garment

fit customization

Wonseop Lee, Hyeong-Seok Ko

Department of Electrical and Computer Engineering, Seoul National University, Seoul, Republic of Korea

a r t i c l e i n f o

Article history:

Received 2 May 2017

Revised 11 October 2017

Accepted 16 October 2017

Available online 6 November 2017

Keywords:

Clothing simulation

Fit customization

Pattern-making

Computer animation

a b s t r a c t

Based on the physically-based clothing simulation, this paper develops a novel method to customize the

fit of the given garment to the reference body. The method defines three misfit measures, namely the

landmark point misfit, landmark line misfit, and circumferential misfit, based on the correspondence be-

tween the landmark points and landmark lines in the body and garment. In terms of the above mis-

fit measures, this paper proposes the heuristic misfit reduction method. The heuristic misfit reduction

method works in the following way. Starting from a given preliminary garment, it (1) performs the fit

evaluation, (2) modifies the panels based on the evaluation to enhance the fit, then (3) performs the

draping simulation with the modified panels. The procedure is repeated until a satisfactory fit is achieved.

According to our experiments, the proposed method successfully customizes a given garment to fit to the

given body.

© 2017 Elsevier Ltd. All rights reserved.

1

c

t

a

o

g

m

[

m

w

s

t

s

s

b

p

a

a

K

t

j

t

F

v

b

T

t

p

b

u

t

t

n

s

l

a

g

t

h

0

. Introduction

Finding out whether a given garment fits to the body (we will

all this part the fit evaluation ) and being able to modify the panels

o customize the fit (we will call this part the panel modification )

re both fundamentally and practically important in the production

f clothes. This paper is about customizing the fit of the given 3D

arment to the given 3D avatar based on fit evaluation and panel

odification.

In the clothing field, several drafting schemes such as ESMOD

1] , BUNKA [2] have long been established for producing gar-

ents. From the given primary body sizes (e.g., waist back length,

aist circumference, etc. for the case of the bodice), each drafting

cheme specifies how to construct the panels that supposedly fit

he target body. (In this work, we use the term panel to mean the

ewing pattern.)

When a given garment which has been constructed targeting a

pecific body A ( Fig. 2 (a)) needs to be customized to fit a different

ody B ( Fig. 2 (b)), by employing a drafting scheme, a preliminary

anel resizing can be performed. In fact, we have programmed an

utomatic drafting scheme which takes the panels for A ( Fig. 2 (c))

nd resizes them for B ( Fig. 2 (d)) by examining the differences in

∗ Corresponding author.

E-mail addresses: [email protected] (W. Lee), [email protected] (H.-S.

o).

m

s

a

p

ttps://doi.org/10.1016/j.cag.2017.10.004

097-8493/© 2017 Elsevier Ltd. All rights reserved.

he two bodies. However, note that the drafting scheme provides

ust a general guideline. That is why we call it “preliminary”. With

he same primary body sizes, the body can have different shapes.

or example, the bust circumference does not tell the shape or

olume of the bust itself. The particular (bent, straight, or cam-

er) shape of the torso is not reflected in the primary body sizes.

herefore a drafting scheme itself can not be regarded as a fit cus-

omization.

This paper aims to develop a fit customization method using a

hysically-based clothing simulator. Fig. 1 (a) shows the front and

ack panels used for the bodice, Fig. 1 (b) shows their draping sim-

lation, and Fig. 1 (c) shows the result after the fit is customized to

he body with the proposed method.

As the input, the proposed fit customization method takes (1)

he 3D geometry of the avatar in a static pose and (2) the prelimi-

ary 3D garment (i.e., the panels comprising the garment with the

eams defined between the panels). It assumes that a number of

andmark points and landmark lines are pre-marked in both the

vatar body and the panels, as summarized in Appendices.

The proposed method is basically a loop. It firstly drapes the

iven garment on the given body by running the clothing simula-

or. It analyzes the fit of the simulated results, makes necessary

odifications to the panels (according to the method described

ubsequently), then the resultant garment is simulated again. The

bove are repeated until the desired fit is achieved.

The purpose of this paper is to develop a method that can be

otentially used for the tailoring task. This paper is not the first

Page 2: Heuristic misfit reduction: A programmable approach for 3D

2 W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13

Fig. 1. Customization of the fit to the given body: (a) before the simulation, (b) without fit customization, (c) with the proposed fit customization. The red cones/lines and

blue spheres/lines represent the body landmark points/lines and the garment landmark points/lines, respectively. (For interpretation of the references to color in this figure

legend, the reader is referred to the web version of this article.)

Fig. 2. (a) Body A, (b) body B, (c) patterns automatically generated by the drafting

rule for body A, (d) patterns automatically generated by the drafting rule for body

B.

o

i

fi

a

fi

p

t

a

fi

j

p

t

t

2

e

2

c

s

a

e

s

t

[

b

n

u

o

f

v

ne that tackles this problem. Differently from the previous stud-

es, however, this paper is based on the classical 2D pattern modi-

cation and tries to adopt the field schemes and practices as much

s possible in the development.

We limit the scope of this paper to the clothing features whose

t can be quantitatively analyzed with the misfit measures pro-

osed in this paper. For example, whether the collar fits well to

he given body cannot be well-defined. When there is an opening

s in the jacket, the fit cannot be defined according to the mis-

t measures given in this paper. Therefore, tops with the collar or

acket is out of the range. This paper will focus on the basic bodice

attern for the ladies, which is an excellent choice to see the po-

ential whether the proposed method can be used for automatizing

he fit customization.

The contribution of this paper can be summarized as:

• We define novel misfit measures that can measure the degree of

misfit based on the discrepancy between the landmark points

and landmark lines of the body and the garment. • With the above measures, we develop the heuristic misfit reduc-

tion method , which can customize the fit by sequentially reduc-

ing misfit measures one after another using iterative method,

and we propose various ways to execute the fit customization

program, namely, the step repetition and program repetition .

. Related work

We briefly review the previous work in the following three cat-

gories: fit evaluation, fit customization, and panel modification.

.1. Fit evaluation

The fit can be evaluated by creating the 3D versions of the

lothing and the body. In the clothing field, using a commercial

oftware, Apeagyei and Otieno [3] imported the panels for a jacket

nd a skirt to the software, then graded them into two differ-

nt sizes. The graded results were simulated on two differently

ized mannequins to analyze the horizontal cross-sectional fit at

he bust, waist, and hip. For tight garments, Apeagyei and Otieno’s

3] analysis is not applicable. In such cases, the cloth pressure can

e used as the measure of the fit. Seo et al. [4] proposed a tech-

ique to evaluate the cloth pressure from the physically-based sim-

lation with the intention of replacing the physical measurement

f cloth pressure. They calculated the pressure by taking the spring

orces exerted on each cloth vertex along its normal direction, di-

ided by the summed area of triangles adjacent to that vertex.

Page 3: Heuristic misfit reduction: A programmable approach for 3D

W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13 3

c

s

o

3

r

(

b

s

w

m

t

a

f

2

e

a

t

a

m

t

c

e

s

m

g

h

c

f

m

t

w

b

t

b

i

fi

c

d

k

t

e

r

i

t

e

p

m

t

t

p

m

B

a

f

f

t

Fig. 3. (a) BLP-frame and BLP-plane for Anterior Waist, (b) BLL-plane for the waist

line. (For interpretation of the references to color in this figure, the reader is re-

ferred to the web version of this article.)

2

m

i

o

t

e

t

b

p

m

t

3

d

a

b

s

w

G

t

3

l

i

f

r

a

N

e

w

v

i

p

N

1 In setting up the BLP-frames for the Lateral Shoulder and Lateral Neck, the ver-

tical up direction is not well defined. In those two cases, instead of the vertical up

direction, the direction that is orthogonal to the shoulder line is used for the y -axis.

As for the BLP-frames existing in the limbs, the center axis of the limb is used for

the y -axis.

Lee et al. [5] proposed an experiment that can evaluate the ac-

uracy of the simulated 3D virtual garments. For this, they (1) 3D-

canned subjects bodies, (2) constructed two identical garments,

ne in 3D and the other with real fabrics, then (3) compared the

D-clothing-to-avatar-fit and the real-clothing-to-subject-fit. The

esultant two fits were compared (1) with human vision, and

2) by calculating cross-sectional vacant space distance and area

etween the body and the garment. Measurement of the cross-

ectional vacant space between the subject and the real garment

as done by 3D-scanning (1) the body alone and (2) with the gar-

ent put on.

Although there have been various attempts to evaluate the fit,

o our knowledge, no misfit measure based on the landmark points

nd landmark lines has been proposed yet in explicit equational

orms as given in this paper.

.2. Fit customization

Cordier et al. [6] presented a web application architecture that

nables online fitting/resizing of the garment to the 3D avatar

long with the automatic adjustment of the 3D avatar according

o the customer’s body measurement.

From a given body and fitted garment, Wang et al. [7] gener-

ted a garment template whose features are related to the hu-

an model such that the template can generate the garment for

he novel body. Wang et al. [8] proposed a system for automatic

ustomization of the garment to the given body. They noted that

ach apparel product can be represented as a feature template pre-

erving its individual characteristics and styling. They proposed a

ade-to-measure clothing production technique by encoding the

arment feature template to have an equivalent structure as the

uman body feature template.

Meng et al. [9] extended Wang’s et al. [8] work with the shape

ontrol capability, which effectively preserved the original shape

eature in the automatic resizing.

Li et al. [10] proposed a method for fitting the given 3D gar-

ent drape onto various body shapes and poses. Their goal was

o find out pose-dependent transformation of the garment drape,

hereas our goal is to find out the customized patterns that may

etter fit the given body. The pose-dependent garment transforma-

ion is, in our case, taken care of by the physics-based simulator

eing used. Li and Lu [11] proposed another method for customiz-

ng the 3D garment (originally constructed for the source body) to

t the target body based on the establishment of cross-sectional

orrespondence between the source and target bodies and tetrahe-

ralization of the space around the body. Another approach of this

ind (i.e., garment construction from the body-to-garment rela-

ionship) called “volume parameterization” was proposed by Wang

t al. [12] , which related the body model and the garment by pa-

ameterizing the space around the body.

Brouet et al. [13] presented an automatic method for transform-

ng garments between the avatars with different body shapes. In

he development of the method, a number of factors were consid-

red including the preservation of the design details, (body) pro-

ortionality, and fit. Lee et al. [14] proposed a method that auto-

atically transfers and fits the 3D garments from source body to

arget body. Their method is applicable even when the source and

arget bodies have different poses. Huang et al. [15] proposed a

attern generation method which generates fit-ensured block gar-

ent. This method parameterized the body model into piecewise

-spline level curves to recognize the feature points of the body. By

dding the wearing ease to the above B-spline curves, they could

orm the desired block garment.

Except for Cordier’s et al. [6] work, the above methods trans-

orm the garment in 3D. In contrast, the garment customization of

he present work is based on the pattern modifications in 2D.

.3. Panel modification

Umetani et al. [16] presented a method in which the 3D gar-

ent and its constituent 2D patterns are coupled in such a way an

nteractive modification of one results in immediate modification

f the other. This work has similarity to our work in that 2D pat-

ern modification is applied to the 3D garment directly. Berthouzoz

t al. [17] presented automatic seam-matching system that ex-

racts seams from patterns and automatically finds seam-pairs to

e sewed based on a probabilistic approach. Bartle et al. [18] pro-

osed 3D pattern editing algorithm which is able to adjust gar-

ent patterns directly in 3D, and the 3D garment edits are applied

o the 2D patterns automatically.

. Defining the misfit measures

According to the fitting literature (e.g., [19] ), fitting is normally

one based on how well the garment landmark points (GLPs)

nd garment landmark lines (GLLs) match with the corresponding

ody landmark points (BLPs) and body landmark lines (BLLs), re-

pectively. This work adopts that principle. More specifically, this

ork defines three misfit measures based on how accurately each

LP/GLL coincides with BLP/BLL when the garment is draped on

he body.

.1. Landmark-point misfit

The misfit of a landmark point (LP-misfit) is defined for each

andmark point. For this, we first define the BLP-frame as shown

n Fig. 3 (a). The z -axis (the blue arrow in the figure) of the BLP-

rame is along the surface normal direction, y -axis (the green ar-

ow) is vertical up direction, and x -axis (red arrow) is determined

ccording to the right-handed rule. 1

The xy -plane of the BLP-frame is called the BLP-plane ( Fig. 3 (a)).

ow, we define the LP-misfit vector e LP

as,

LP = proj plane XY

( v B i − v G

i ) , (1)

hich is the 2D Euclidean vector between the corresponding BLP

B i

and GLP v G i

excluding the surface normal directional component,

.e., e LP

is the 2D vector obtained by projecting the GLP to the BLP-

lane then resolving the coordinates with respect to the BLP-frame.

ote that the misfit vector itself informs how the panel has to be

Page 4: Heuristic misfit reduction: A programmable approach for 3D

4 W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13

w

p

4

p

m

t

w

t

u

μ

T

t

d

m

b

s

s

f

f

s

p

u

x

g

o

i

g

5

fi

a

h

d

t

5

M

modified to enhance the fit. (As introduced in Section 5.4.1 , the

panel can be effectively modified by translating the problematic

GLP by the current LP-misfit.) The LP-misfit magnitude e LP

= || e LP ||

tells the amount of misfit.

3.2. Landmark-line misfit

The landmark line misfit (LL-misfit) is defined for each circum-

ferential landmark line. We define the plane that contains the BLL

as the BLL-plane. Fig. 3 (b) shows the BLL-plane for the waist line.

For the LL-misfit, we create 12 reference points to each BLL and

GLL. Fig. 3 (b) shows the reference points for the GLL with tiny

blue spheres. (Four of the reference points are identical to the LPs.)

Now, the LL-misfit vector e LL

is defined as

e LL

=

⎢ ⎢ ⎣

e RP 1

e RP 2

. . . e

RP 12

⎥ ⎥ ⎦

=

⎢ ⎢ ⎢ ⎣

v B RP 1

− v G RP 1

v B RP 2

− v G RP 2

. . . v B

RP 12 − v G

RP 12

⎥ ⎥ ⎥ ⎦

, (2)

where v B RP i and v G RP i

are the i th reference point in BLL and GLL, re-

spectively. e RP i

is the i th reference point misfit vector. The positive

direction is up/counterclockwise.

The LL-misfit is not defined for non-circumferential LLs. The

reasons for not defining LL-misfit for the non-circumferential LL

are (1) misalignment of the BLL and GLL reference points does not

directly indicate misfit, and (2) LP-misfit largely covers the misfit

associated with the non-circumferential LLs. For example, as pre-

sented in the later sections, the misfit along the shoulder line is

taken care of by resolving the LP-misfit of Lateral Shoulder and

Lateral Neck.

The LL-misfit average E( e LL ) is defined as

E( e LL ) =

e RP 1

+ e RP 2

+ · · · + e RP 12

12

, (3)

restriction of which to the up and down direction can tell how the

length of the panel should be adjusted in general to match the BLL.

When the GLL runs across multiple panels, E( e LL | Panel ) rather than

E( e LL ) can be more useful for the customization (see Section 5.4 ).

E( e LL | Panel ) is defined as the mean of e

RP i where the summation

(for the mean) is limited to the GLL reference points that lie within

the given panel. (If no GLL reference point lies in the panel, then

the two closest reference points are used for the calculation.)

The LL-misfit root mean square e LL

is the root mean square of the

LL-misfit vector, i.e.,

e LL

=

√ ∥∥e RP 1

∥∥2 +

∥∥e RP 2

∥∥2 + · · · +

∥∥e RP 12

∥∥2

12

, (4)

which tells how well the GLL is in general aligned to the BLL-plane.

3.3. Circumferential misfit

The circumferential misfit e LLC

is also defined only for the cir-

cumferential LLs. e LLC

is a scalar giving the difference in the cir-

cumferential length between the corresponding GLL and BLL. More

specifically, e LLC

is defined as

e LLC

= GLL length − ( BLL length + ε) , (5)

where ε is the ease between GLL and BLL. For all the experiments

in this paper, we set the ε to 3 cm. A positive/negative e LLC

means

the GLL circumference is longer/shorter than the BLL circumfer-

ence. This measure tells how tight or loose the circumferential GLL

is compared to the BLL. The vector version e LLC

of e LLC

, which is

called the circumferential misfit vector is defined as

e LLC

= e LLC

· e x , (6)

here e x is the horizontal (x directional) unit vector in the 2D

anel space.

. Formulation as an optimization problem

With the above misfit measures, the fit customization can be in

rinciple formulated as an optimization problem. For the given gar-

ent G and body B , we can think of the following objective func-

ion

(G, B ) =

n ∑

i =1

w i ∗ μi , (7)

here w i and μi are the weights and misfit terms , respectively. For

he case of the bodice, one possible set of misfit terms that can be

sed for the objective function are as follows:

• μ1 = e LP

( L. Lateral Shouder ) + e LP

( L. Lateral Neck ) +

e LP

( R. Lateral Shouder ) + e LP

( R. Lateral Neck ) , ( Fig. 9 (a)–(c)) • μ2 = e

LP ( Anterior Neck ) + e

LP ( Cervicale ) , ( Fig. 9 (d)–(f))

• μ3 = e LL ( Bust ) ( Fig. 9 (g)–(i))

• μ4 = e LLC

( Bust ) ( Fig. 9 (j)–(l)) • μ5 = e

LL ( Waist ) ( Fig. 9 (m)–(o))

• μ6 = e LLC

( Waist ) ( Fig. 9 (p)–(r))

In the above, μ1 is for the shoulder line, μ2 is for the neckline,

3 and μ4 are for the bust line, μ5 and μ6 are for the waist line.

he μi ’s involved in the objective function varies depending on

he garment under consideration. For example, for a certain bodice

esign, the neckline misfit can be omitted or a different formula

ight need to be used. The details of the objective function should

e determined with the aid of the fashion expert.

Note that the misfit measure appearing in each of the μi is

calar, but it involves multiple panels. For example, when we re-

olve the error e LP

( Lateral Shoulder ) , it ends up modifying both the

ront and the back panels that share the Lateral Shoulder GLP. In

act, resolving each of μ3 , μ5 calls for modifying all four panels

ince both the Bust and Waist landmark lines run across all four

anels (two in the front and two in the back).

The shape of a panel is determined by a number of (2D) artic-

lation points and (2D) curve control points. We denote them as

j , ( j = 1 , . . . , M) . For the optimization, we need to calculate the

radient ∂μi ∂ x j

.

We will call the above way of mathematically formulating and

ptimizing the objective function by introducing all the participat-

ng articulation and curve control points x j as the mathematical

radient method (MGM).

. Heuristic misfit reduction – our fit customization method

Although the above MGM sounds theoretically clear, some dif-

culties arise when to implement it. Therefore, this paper takes

different approach. We develop a heuristic method called the

euristic misfit reduction method (HMR), which is supposed to re-

uce the misfit such that we can apply the method iteratively until

he desired fit is achieved.

.1. Motivation

The motivations that led us to develop the HMR instead of

GM are:

• The misfit is measured in 3D whereas the panel modification

should be done in 2D. Although its computational procedure

(i.e., the simulation algorithm) is known, the formula that re-

lates the 2D panel shape to its 3D draped shape is not explicitly

given. Therefore, with the mathematical gradient method, it is

not clear how to calculate the gradient.

Page 5: Heuristic misfit reduction: A programmable approach for 3D

W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13 5

5

s

p

m

w

t

m

t

d

i

b

s

o

F

n

S

i

i

u

e

l

H

s

f

Fig. 4. Flow chart of HMR procedure.

Fig. 5. Panel modification operations: (a) point translation, (b) edge modification,

(c-1) spreading (vertical), (c-2) spreading (horizontal), (d-1) contraction (vertical),

(d-2) contraction (horizontal), (e) dart edit.

t

c

d

5

a

h

f

w

• The objective function tends to become complex and should be

formulated differently for each garment. Each panel can intro-

duce a number of variables, and for a complex garment which

consists of tens of panels, the formulation of the objective func-

tion itself can be a burdensome task. Another factor that in-

creases the complexity is that the terms in the objective func-

tions can be related to multiple panels. • There has been well-established pattern-making expertise that

tells how the misfit should be reduced. For example, if the GLL

and BLL do not match at the bust line (as shown in Fig. 1 (b)),

then extending the relevant panels by applying the Spreading

operation in Fig. 5 (c-1) to the upper (upper than the bust line)

part of the panels can produce the desired effect. Such kind

of modification is not limited to the bust line. In most misfit

cases, it is pretty clear how to proceed for the misfit reduc-

tion. Based on the above observation, we make modifications to

the MGM in the following two aspects to obtain HMR: (1) In-

stead of scalar misfits, HMR uses misfit vectors, which can tell

us how the modification should be performed without need for

calculating the gradients. (2) Whereas all the μi ’s appearing in

the objective function are handled in an arbitrary order, HMR

processes each misfit vector � μi separately in the programmed

order. • It turns out misfit reductions can be covered with only a hand-

ful of operations. For the customizations experimented in this

paper, the five panel operations introduced in Fig. 5 suffice.

.2. Formulation as a sequential misfit reduction program

We now formulate the HMR as follows. Differently from the

ummation of the misfit scalars ( Eq. (7) ), the HMR is basically a

rogram �( G , B ), which sequentially reduces the magnitude of the

isfit vectors � μ1 , � μ2 , . . . , � μn :

(G, B ) = ( � μ1 , � μ2 , . . . , � μn | �1 , �2 , . . . , �n ) , (8)

here each

� μi is e

LP , e

LL , or e

LLC . �i is supplementary information

hat is needed for reducing � μi . �i consists of (1) the type of the

isfit, (2) the GLP or GLL to be modified to reduce this misfit, (3)

he panel modification operation to be used for this reduction, (4)

etailed information for the post-processing.

Here, we highlight that the HMR is not an optimization, but it

s a programming. The program details, (i.e., (1) which

� μi ’s should

e included in the program, (2) which information �i should be

upplied for the reduction of the misfit vector � μi , and (3) the order

f � μi ’s) should be determined with the aid of the fashion expert.

or the case of the bodice, one possible HMR program is to set

= 10 and have the following misfit vectors:

� μ1 = e

LP ( L. Lateral Shoulder ) �

μ2 = e LP ( R. Lateral Shoulder )

� μ3 = e

LP ( L. Lateral Neck ) �

μ4 = e LP ( R. Lateral Neck )

� μ5 = e

LP ( Anterior Neck ) �

μ6 = e LP ( Cervicale )

� μ7 = e

LL ( Bust ) �

μ8 = e LLC

( Bust ) � μ9 = e

LL ( Waist ) �

μ10 = e LLC

( Waist )

(9)

In contrast to MGM, in the HMR, the misfit vectors defined in

ection 3 can tell the specific amount as well as direction of mod-

fication. No weights ( w i ) need to be given in HMR. (Note that giv-

ng wrong weights can affect the accuracy of the MGM.)

HMR is basically a repetition of the procedure shown in Fig. 4

ntil the desired fit is achieved. Upon receiving ( � μi , �i ) , HMR

valuates the actual misfit value � μ∗i

with the current draping simu-

ation. If � μ∗i

is above the threshold, based on the information in �i ,

MR performs panel modification and subsequent post-processing

teps automatically without any human intervention.

For the sake of clarity, the subsequent description will be done

or the female bodice shown in Fig. 1 . Note that it does not lose

he generality; the programmable nature of the proposed method

an apply to different types of garments such as pants, one-piece

ress, and so forth.

.3. Panel modification operations

The panel modification operations this paper proposes to use

re the following five, each of which are depicted in Fig. 5 , which

as been determined in consultation with a fashion expert. (In the

ashion industry, most panel modification task is in fact covered

ith the those five operations.)

• Point translation: this operation translates the designated

points by the given 2D vector as shown in Fig. 5 (a). The amount

should be determined based on the misfit. • Edge modification: this operation translates (with some possi-

ble rotation) a contour edge as shown in Fig. 5 (b). • Spreading: this operation extends a panel; It cuts the panel,

spreads them as shown in Fig. 5 (c-1) and (c-2), then fills the

gap so that the contour has a smooth shape.

Page 6: Heuristic misfit reduction: A programmable approach for 3D

6 W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13

Fig. 6. Panel modification for LP-Misfit: (a) initial, (b) target GLP, (c) point trans-

lation, (d) intermediate result, (e) the feature subject to post-processing, (e) post-

processing.

Fig. 7. Panel modification for LL-misfit: (a) initial, (b) target GLL, (c) horizontal

cut, (d) vertical spreading, (e) intermediate result, (f) the feature subject to post-

processing, (g) post-processing.

Fig. 8. Panel modification for circumferential-misfit: (a) initial, (b) determine the

vertical line where the cut should be made, (c) vertical cut, (d) horizontal spread-

ing, (e) intermediate result, (f) the feature subject to post-processing, (g) post-

processing.

C

C

s

5

w

h

a

t

t

t

5

c

l

c

u

c

t

i

r

a

r

• Contraction: this operation shortens a panel; It cuts the panel,

makes them overlap as shown in Fig. 5 (d-1) and (d-2), then re-

draws the contour so that the contour has a smooth shape. • Dart edit: this operation increases or decreases the dart amount

as shown in Fig. 5 (e). This operation may entail modification of

the shape of the dart-adjacent lines.

5.4. Panel modification algorithm

In Section 4 , we noted that, each misfit term

� μi is in fact an LP-,

LL-, and circumferential misfit. Instead of calculating the mathe-

matical gradient, the proposed HMR performs panel modification

that are expected to reduce the corresponding misfit vector di-

rectly. In this section, we describe which/how panel modification

operations are automatically performed for the reduction of each

type of misfit.

5.4.1. Panel modification for LP-misfit

Suppose that an LP-misfit vector � μi along with �i is given. Let

P be the landmark point ( Fig. 6 (b)) for which the misfit vector is

defined. Note that a number of panels can share P . In such a case,

we will say that those panels are adjacent to P . LP-misfit is usually

resolved with the point translation operation. (All those informa-

tion including P itself, all the adjacent panels to P , the operation to

resolve the misfit, post-processing steps are included in �i .)

This section explains how the given LP-misfit is resolved with

the point translation operation. Suppose that the LP-misfit is given

as shown with the arrow in Fig. 6 (c). (Although e LP

is a 3D vector,

since it is tangent to the body surface, the vector can be made to

appear with the 2D panel.) Panel modification for the LP-misfit is

done in the following steps:

1. Locate the GLP under consideration.

2. Translate the GLP by � μi .

3. Locate the feature to post-process ( Fig. 6 (e)).

4. Perform the necessary post-processing as summarized in �i

( Fig. 6 (f)). (For example, for the case of the bodice, the point

translation operation for the LP-misfit resolution can result in

increasing/decreasing the total armhole length, which should

be cancelled out to keep the relationship with the neighboring

panels (e.g., sleeves) valid. The post-processing is discussed in

more detail in Section 5.4.4 .)

Since the 3D vector e LP

is directly used in association with the

2D panel, translating by e LP

in 2D may end up overshooting or un-

dershooting. The above can be reduced by performing relaxation,

i.e., after the first round of Point Translation (followed by static

simulation and fit evaluation), we run the second round of Point

Translation to the first result, and so forth.

5.4.2. Panel modification for LL-misfit

Suppose that an LL-misfit μi = E( e LL | Panel ) is given along with

�i . The LL-misfit in general measures how much the GLL is

offset from the BLL vertically. Therefore, to resolve it, Spread-

ing/Contraction operation is performed in the vertical direction af-

ter cutting the panel horizontally. We call them vertical Spreading/

ontraction to distinguish them from the horizontal Spreading/

ontraction in Section 5.4.3 .

More specifically, the LL-misfit is resolved in the following

teps:

1. Locate the GLL under consideration ( Fig. 7 (b)).

2. Cut the panel horizontally ( Fig. 7 (c)).

3. Spread/contract the panel vertically by the amount of � μi

( Fig. 7 (d)).

4. Locate the feature to post-process ( Fig. 7 (f)).

5. Perform post-processing ( Fig. 7 (g)).

.4.3. Panel modification for circumferential misfit

Suppose that a circumferential misfit μi = e LLC

is given along

ith �i . This section explains how to resolve it by performing a

orizontal Spreading/Contraction by e LLC

as shown in Fig. 9 (l). The

mount of Spreading/Contraction should be distributed to the par-

icipating panels. In this work, we use the even-distribution rule:

he amount is evenly distributed. So that the resulting garment has

he proper circumference (with some necessary ease) as the body.

1. Count the number of participating panels. Let it be N .

2. For each participating panels, do the following:

3. Determine the vertical line where the cut should be made

( Fig. 8 (b)).

4. Cut the panel vertically ( Fig. 8 (c)).

5. Spread the panel horizontally by the amount of � μi

N ( Fig. 8 (d)).

6. Locate the feature post-process ( Fig. 8 (f)).

7. Perform post-processing ( Fig. 8 (g)).

.4.4. Post-processing after panel modification

Note that a panel modification operation can create unwanted

hanges elsewhere. For example, in the bodice, adjusting the bust

ine circumference with a horizontal spreading/contraction may

ause unwanted changes in the waistline and shoulder line. Those

nwanted changes have to be cancelled to preserve the originally

ustomized results at those locations.

In some cases, a panel modification operation calls for adjusting

he line shape. In the case of neck line modification

� μ2 ( Fig. 9 (f)),

f translation is applied only to Anterior Neck and Cervicale, the

esulting neckline curve will have a strange shape. As another ex-

mple, with the bust circumference modification

� μ8 ( Fig. 9 (l)), the

esulting shoulder line may not have a straight shape any more. To

Page 7: Heuristic misfit reduction: A programmable approach for 3D

W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13 7

Fig. 9. Fit customization for the bodice. The left/center column shows before/after the fit customization, respectively. The right column shows the panels after the panel

modification operations are applied. (a)–(c) fitting the shoulder line, (d)–(f) fitting the neck line, (g)–(i) fitting the bust line, (j)–(l) fitting the bust circumference, (m)–(o)

fitting the waist line, and (p)–(r) fitting the waist circumference.

Page 8: Heuristic misfit reduction: A programmable approach for 3D

8 W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13

6

p

p

p

l

r

t

c

o

e

1

avoid such anomalies, in such cases, we adjust the control points

of the curve or redraw straight line, according to the pattern draft-

ing rule.

In programming, when a misfit vector � μi goes into the pro-

gram, the fashion expert specifies the details of the necessary post-

processing steps into �i . On the other hand, in executing the pro-

gram, the post-processing is performed without human interven-

tion. Note that (1) the decision on the steps to be done in the

post-processing and (2) putting those steps into �i are done man-

ually in the programming stage. However, the execution of the so-

determined post-processing steps are performed automatically in

the execution stage.

For the modified panels, the triangulation should be re-done,

thus the seams need to be re-created, which are taken care of

automatically. There might occur self-intersections when the sim-

ulation restarts with the modified panels. The collision handling

module we use in this work is equipped with modern discrete

collision detection (DCD) based collision resolution algorithms, in-

cluding the global intersection analysis [20] and intersection con-

tour minimization [21] , by which the potential inter-penetrations

are gradually resolved as the simulation progresses.

Program 1 : Program to customize the fit of the bodice.

1: Fit the shoulder line;

2: Fit the neck line;

3: Fit the bust line;

4: Fit the bust circumference;

5: Fit the waist line;

6: Fit the waist circumference;

Program 2 : Program to customize the fit of the sleeve.

1: Fit the bicep line;

2: Fit the bicep circumference;

3: Fit the elbow line;

4: Fit the elbow circumference;

5: Fit the wrist line;

6: Fit the wrist circumference;

5.5. Execution of HMR program

In Program 1 , we will call the whole program as the HMR pro-

gram (or simply as program ), and each step as the HMR step (or

simply as step ). The above program for the bodice consists of six

steps. When all the steps comprising the program are (sequen-

tially) executed, we call it a cycle . For the purpose of obtaining

more accurate fit, multiple cycles may be executed repetitive.

We use HMR( i , j ) to denote various ways of executing the fit

customization program. The parameters i and j represent the step

repetition and program repetition, respectively. If an HMR program

is composed of four steps, e.g., adjusting (a) shoulder line, (b) neck

line, (c) bust line, and (d) waist line, then HMR(2, 3) executes the

steps in the order of (a, a, b, b, c, c, d, d), (a, a, b, b, c, c, d, d), (a,

a, b, b, c, c, d, d).

5.6. Tracing the fit customization HMR(1,1)

To help understand how the proposed heuristic misfit reduc-

tion method works, we trace how it executes for the bodice.

More specifically, this section will show how HMR(1,1) exe-

cutes �(G, B ) = ( � μ1 , � μ2 , . . . , � μ10 | �1 , �2 , . . . , �10 ) , where the

� μi ’s listed in Eq. (9) for the program. Note that, at this point, it

is assumed that both the garment and body are already given with

all the landmark points/lines are labeled, and initial draping simu-

lation has already been done.

• Fitting the shoulder line

From the draped result, HMR evaluates the LP-misfit of the Lat-

eral Shoulder and Lateral Neck for both left and right and per-

forms Point Translation operation to both LPs. See Fig. 9 (a)–(c). • Fitting the neck line

This step resolves the misfit at the Anterior Neck and Cervicale

by executing the Point Translation operation to the two LPs. Af-

ter the point translation, HMR re-draws the neckline passing

through the 4 LPs, namely, the Left/Right Lateral Neck , Anterior

Neck , and Cervicale . Fig. 9 (d) and (e) shows the situation before

and after the fit customization for the neck line, respectively.

Program 3 : Program to customize the fit of the pants.

1: Fit the waist circumference;

2: Fit the hip line;

3: Fit the hip circumference;

4: Fit the knee line;

5: Fit the knee circumference;

6: Fit the ankle line;

Program 4 : Program to customize the fit of the skirt.

1: Fit the waist circumference;

2: Fit the hip line;

3: Fit the hip circumference;

4: Fit the knee line;

5: Fit the skirt circumference;

• Fitting the bust line

HMR makes the horizontal cut at the level of (or a little above)

the Nipple and performs either Spreading or Contraction opera-

tion by y value of E( e LL ) as shown in Fig. 9 (i). Fig. 9 (g) and (h)

shows before and after the customization, respectively. • Fitting the bust circumference

As shown in Fig. 9 (j), based on e LLC

, HMR performs either the

Spreading or Contraction operation (with a vertical panel cut)

to all the participating panels. This operation may destroy the

shoulder line shape, so post-processing step will restore the

original shape of shoulder line. • Fitting the waist line

Similarly to the bust line, HMR performs vertical Spreading or

Contraction operation with horizontal cut. Note that, to pre-

serve the original length of the side line, the Dart Edit oper-

ation has to be performed ( Fig. 9 (o)). Fig. 9 (m) and (n) shows

before and after the customization, respectively. • Fitting the waist circumference

It does not perform Spreading/Contraction to resolve the waist

circumference, since with it, the bust circumference which has

been resolved above may not be correct any more. To resolve

the waist circumference, the Dart Edit operation is performed

for the waist circumference as shown in Fig. 9 (r). Fig. 9 (p) and

(q) shows before and after the customization, respectively. Ex-

amining the waist line in Fig. 9 (p) and (q), we can observe the

circumferential fit is clearly better in Fig. 9 (q).

. Programming guidelines

We recommend a general guideline how the HMR should be

rogrammed. When customizing the fit for the bodice, for exam-

le, according to the classical tailoring practice, customization is

erformed so that the misfit is reduced in the shoulder line, neck

ine, bust line, and waist line. Additionally, the fit is adjusted to

educe the circumferential misfit in the bust and waist.

The order of the customization matters. Suppose that the fit has

o be adjusted for the bust and waist. Option 1 is to perform the

ustomization to the waist first then to the bust. Option 2 is the

ther way around, i.e., to the bust first then to the waist. In gen-

ral, the tailors take Option 2. The reason they do not take Option

is because, as they make customization to the bust, it tends to

Page 9: Heuristic misfit reduction: A programmable approach for 3D

W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13 9

Fig. 10. LP-misfit e LP

before, during, after HMR(3,1): the blue curve shows the mis-

fit before the fit customization. The orange, violet and green curves show the misfit

after the first, second, and third repetition of the steps, respectively. (For interpre-

tation of the references to color in this figure legend, the reader is referred to the

web version of this article.)

Table 1

LP-misfit e LP

before, during, after HMR(3,1).

GLP/BLP Initial e LP

1st repeat 2nd repeat 3rd repeat

Rt_Lateral_Shoulder 1.24300 0.11878 0.11878 0.11878

Lt_Lateral_Shoulder 1.39355 0.42485 0.42485 0.42485

Rt_Lateral_Neck 1.18665 0.42264 0.42264 0.42264

Lt_Lateral_Neck 1.21878 0.53364 0.53364 0.53364

Anterior Neck 1.21538 0.39548 0.34479 0.28379

Cervicale 1.34751 0.45913 0.37831 0.33018

Rt_Nipple 2.29070 0.63194 0.59217 0.44337

Lt_Nipple 2.27191 0.60502 0.48388 0.30767

Anterior_Waist 4.94377 0.82717 0.59917 0.52348

Posterior_Waist 2.35554 0.37216 0.29573 0.04496

Rt_Lateral_Waist 3.03493 1.11455 0.94340 0.86102

Lt_Lateral_Waist 3.08580 1.10132 0.99375 0.76533

a

w

p

o

j

t

i

b

F

a

f

t

7

o

a

s

K

T

s

p

a

e

fi

m

Fig. 11. LP-misfit e LP

before, during, after HMR(1,3): the blue curve shows the misfit

before the fit customization. The orange, violet and green curves show the misfit

after the first, second, and third cycle, respectively. (For interpretation of the refer-

ences to color in this figure legend, the reader is referred to the web version of this

article.)

Table 2

LP-misfit e LP

before, during, after HMR(1,3).

GLP/BLP Initial e LP

1st repeat 2nd repeat 3rd repeat

Rt_Lateral_Shoulder 1.24300 0.25350 0.25350 0.25350

Lt_Lateral_Shoulder 1.39355 0.16783 0.16783 0.16783

Rt_Lateral_Neck 1.18665 0.27654 0.27654 0.27654

Lt_Lateral_Neck 1.21878 0.30888 0.30888 0.30888

Anterior Neck 1.21538 0.22315 0.13559 0.10482

Cervicale 1.34751 0.20223 0.13570 0.16011

Rt_Nipple 2.29070 1.11068 0.89758 0.86569

Lt_Nipple 2.27191 0.94568 0.80739 0.70648

Anterior_Waist 4.94377 1.05868 0.73907 0.55934

Posterior_Waist 2.35554 0.62881 0.42506 0.37538

Rt_Lateral_Waist 3.03493 1.50263 1.21424 1.21325

Lt_Lateral_Waist 3.08580 0.98268 0.86393 0.74708

Fig. 12. LL-misfit e LL

of the waist line before, during, after HMR(1,3): the blue curve

shows the misfit before the fit customization. The orange, violet and green curves

show the misfit after the first, second, and third cycle, respectively. (For interpre-

tation of the references to color in this figure legend, the reader is referred to the

web version of this article.)

e

c

m

T

t

T

d

G

ffect the fit at the waist. This paper adopts the above wisdom,

hich we will call the top-to-bottom principle . The top-to-bottom

rinciple applies to all kinds of garments. For example, for the skirt

r pants, the fit should be adjusted for the waist first, then the ad-

ustment should go down from it. For the case of the bodice, the

op-to-bottom principle leads us to have the HMR program listed

n Program 1 .

Even though the presentation so far was done only for the

odice, our current implementation can cover other garments.

ig. 13 shows the results when HMR is applied to a blouse, pants,

nd skirt. The programs listed in Programs 2–4 are the ones used

or the customization of the sleeve (only), pants, and skirt, respec-

ively.

. Results

We implemented the proposed garment customization method

n a 3.40 GHz Intel Core(TM) i7-6700 processor with 8GB memory

nd a Nvidia GeForce GTX 1060 video card.

We used a physically-based clothing simulator which uses the

tretch and shear models from Baraff and Witkin [22] and Choi and

o [23] , and hinge-based bending model from Grinspun et al. [24] .

he draping simulation was run on the static reference pose.

We ran the HMR program listed in Section 6 to the panels

hown in Fig. 9 . To observe the effect of the step repetition and

rogram repetition, we ran the program in two ways: HMR(3,1)

nd HMR(1,3).

Fig. 10 and Table 1 show the LP-misfit before, during, after the

xecution of HMR(3,1). As Fig. 10 shows, the first execution of each

t customization step significantly reduced the misfit, then the

isfit reduction in the subsequent step repetitions was minor.

Fig. 11 and Table 2 show the LP-misfit before, during, after the

xecution of HMR(1,3). As Fig. 11 shows, the first cycle of the fit

ustomization program significantly reduced the misfit, then the

isfit reduction in the subsequent cycles was minor.

We note the reason why the third column (i.e., 1st repeat) of

able 1 is different from that of Table 2 . It is because, in Table 2 ,

he misfit was measured after the program repetition, whereas, in

able 1 , the misfit was measured after each step repetition.

Fig. 12 and Table 3 show the LL-misfit of the waist line before,

uring, after HMR(1,3). Along the waist line, there are 12 points (4

LPs and 8 reference points) to be evaluated. As in the LP-misfit,

Page 10: Heuristic misfit reduction: A programmable approach for 3D

10 W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13

Fig. 13. The proposed fit customization applied to the sleeved blouse, pants and skirt: (a) sleeved blouse before, (b) sleeved blouse after, (c) pants before, (d) pants after, (e)

skirt before, (f) skirt after.

Fig. 14. Fit customization applied to the various body sizes: the rows represent the height (from top to bottom, short, medium, and tall). The columns represent the fatness

(from left to right, skinny, medium, and fat). For each body size, the left/right images show before/after the execution of the proposed fit customization, respectively.

Page 11: Heuristic misfit reduction: A programmable approach for 3D

W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13 11

Fig. 15. The proposed fit customization method applied to a shirt and pants: (a) before, (b) before (wireframe), (c) before sideview, (d) before sideview (wireframe), (e) after,

(f) after (wireframe), (g) after sideview, (h) after sideview (wireframe).

Table 3

LL-misfit e LL

of the waist line before, during, after HMR(1,3).

GLLRPs Initial e LL

1st repeat 2nd repeat 3rd repeat

Anterior_Waist 4.629330 0.485207 0.328720 0.300865

RP_1 4.694990 0.646416 0.467384 0.437149

RP_2 3.578280 0.256493 0.085480 0.063713

Rt_Lateral_Waist 2.933810 0.453438 0.232857 0.219887

RP_3 2.473740 0.479454 0.325577 0.327187

RP_4 2.313640 0.286163 0.137405 0.135765

Posterior_Waist 2.548260 0.350502 0.330757 0.325074

RP_5 2.593440 0.381554 0.341072 0.341591

RP_6 2.497630 0.233269 0.169167 0.166809

Lt_Lateral_Waist 3.117200 0.350502 0.230125 0.215881

RP_7 3.642250 0.207504 0.214770 0.215881

RP_8 4.688150 0.607826 0.455521 0.426414

Anterior_Waist 4.629330 0.482330 0.328720 0.300865

t

d

c

A

H

t

T

H

r

t

t

i

r

F

i

S

F

F

F

t

f

e

t

o

8

o

t

t

i

t

c

l

t

t

c

e

s

b

o

m

m

b

a

f

s

i

m

s

c

i

t

t

A

g

he first cycle significantly reduced the misfit, then the misfit re-

uction in the subsequent cycles was minor.

According to the experiments, after three iterations, further

hange was not significant for both HMR(3,1) and HMR(1,3).

lthough the differences are subtle, the authors recommend

MR(1, n ) rather than HMR( n , 1). We note that the fit customiza-

ion method we proposed in this paper produced expected results.

he accompanying video shows that just one iteration of HMR, i.e.,

MR(1,1) produces quite satisfactory customization.

Although the video shows each fitting step interactively, we can

un the whole customization session in batch, which takes about

wo minutes for the bodice alone. (The time taken for the fit cus-

omization depends on the simulator performance. Since the phys-

cal simulation itself is not the contribution of this paper, we omit

igorous time analyses here.)

We extended the experiment to various body sizes.

ig. 14 shows the fit customization results for the nine bod-

es, SS, SM, SF, MS, MM, MF, TS, TM, and TF, where the first letters

, M, T stand for Short, Medium, Tall and the second letters S, M,

stand for Skinny, Medium, Fat.

We experimented the proposed method to two sample clothes.

ig. 15 shows the application of HMR to a T-shirt and jeans.

ig. 15 (a)–(d) and Fig. 15 (e)–(h) show before and after the fit cus-

omization, respectively. Reduction of the LL-misfit and the circum-

erential misfit (thus the improved fit in the length and circumfer-

ntial fit) is clearly noticeable. (The circumferential misfit reduc-

ion for the T-shirt at the waist was performed with enough ease

n purpose.)

. Conclusion

Based on physically-based clothing simulation, this paper devel-

ped a new method to customize the fit of the given garment to

he reference body. The method is basically a loop that consists of

hree steps: (1) fit evaluation, (2) panel modification, and (3) drap-

ng simulation. This work proposed three misfit measures, namely,

he LP-misfit, LL-misfit, and circumferential misfit, based on the

orrespondence in 3D between the landmark points and landmark

ines of the body and garment.

In this work, the fit customization was achieved by modifying

he 2D panels. We defined a number of panel modification opera-

ions based on the traditional tailoring practice. Five panel modifi-

ation operations were enough to cover the basic garment patterns

xperimented in this paper.

The proposed method can work harmoniously with drafting

chemes. It can apply to the basic patterns created from an ar-

itrary drafting scheme whether it be the sleeves, pants, skirts,

r one-piece. In the future work, we plan to extend the proposed

ethod so that the it can be applied to more general range of gar-

ents.

Note that the variations created for the design purpose cannot

e covered with the proposed method. Often the basic patterns

re varied to create different designs. There is no typical principle

or the variation; It entirely depends on the intention of the de-

igner; A garment might be designed to have loose and long waist,

n which case the proposed method will find that there is a large

isfit.

Although this work is based on the physically-based clothing

imulation, the approach it adopted for the fit customization is

lassical. It resembles the fitting procedure that has been practiced

n the custom tailor shop. The present work learns from old , which

urns out to produce predictable and expected customization of

he fit.

cknowledgments

This work was supported by the Basic Science Research Pro-

ram through the National Research Foundation of Korea (NRF)

Page 12: Heuristic misfit reduction: A programmable approach for 3D

12 W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13

Fig. A.1. BLPs and BLLs on the body: (a) BLPs, (b) BLLs.

p

B

e

b

n

i

s

fi

m

M

T

f

T

a

b

o

s

i

p

A

w

F

p

u

funded by the Ministry of Education, Science and Technology

(MEST) (No. 2015R1A2A1A10055178 ), in part by the Brain Korea 21

Project 2017 funded by the Ministry of Education (MOE) (No. 5264-

20170100), and in part by ASRI (Automation and Systems Research

Institute at Seoul National University).

Appendix A. Landmark points/lines of the body and garment

ASTM D5219-09 [25] and ISO 8559:1989 [26] define the land-

mark points and landmark lines (they call them “features”) of the

human body.

This paper adopts the features defined in the above standards,

but decides to call those feature points and lines as the body land-

mark points and body landmark lines, respectively.

The body landmark points (BLPs) and body landmark lines

(BLLs) mark the key locations and lines on the surface of the body

as shown in Fig. A.1 . On the other hand, the garment landmark

points (GLPs) and garment landmark lines (GLLs) are the key lo-

cations and lines of the garment as shown in Fig. A.2 , which are

supposed to coincide with the BLPs and BLLs. Therefore, the dis-

crepancies in (BLPs, GLPs) and (BLLs, GLLs) can be used as a mea-

sure that tells the degree of misfit.

Appendix B. Determination of BLPs and BLLs of a novel body

We can think of two kinds of novel bodies: (1) 3D scanned bod-

ies and (2) program-generated bodies from the input of primary

body sizes. In this work, our novel body handling is limited to the

program-generated bodies. (Note that most 3D body scan systems

Fig. A.2. GLPs and GLLs used for the bodice: (a) GLPs, (b) GLLs.

t

P

F

T

rovide primary body sizes.) Below, we describe how the BLPs and

LLs are determined for program-generated novel bodies.

Throughout our fit customization, we use the so-called param-

terized body. For the parameterized body, standard sized 3 × 3 set

odies (SS, SM, SF, MS, MM, MF, TS, TM, and TF (see Section 7 ))

eed to be prepared first. Their preparation is done in the follow-

ng steps: (1) from the given primary body size specs for the 3 × 3

et bodies, the 3D modeler is asked to create the MM body mesh

rst. (2) On this MM body, an anthropometrical expert is asked to

ark the BLPs and BLLs. (3) The 3D modeler is asked to modify the

M body mesh so that all the BLPs come to the mesh vertices. (4)

he 3D modeler is asked to modify the MM body mesh resulting

rom step (3) to create the rest 8 bodies (i.e., SS, SM, SF, MS, MF,

S, TM, and TF), while keeping the mesh topology same. (5) The

nimator is asked to complete the skeleton and rigging.

Then, creation of an arbitrarily sized body can be done by com-

ining the above 3 × 3 set bodies (interpolation-based), the details

f which are omitted in this paper. Since the mesh topology is the

ame for all 3 × 3 set bodies, in the novel body which is created by

nterpolating the set bodies, the BLPs and BLLs come at reasonable

laces.

ppendix C. Reference pose

An important decision in quantitative analysis of the fit is,

hich pose should be used for the fit evaluation. As shown in

ig. C.1 , geometrical shape of the shoulder varies significantly de-

ending on the arm height. In the real garment fitting, the fit eval-

ation is done mainly in the H-pose. This work also decides to use

he H-pose as the reference pose. More specifically, we use the H-

ose with the armpit = 15 ° and crotch = 20 °.

ig. C.1. Poses for the fit evaluation: (a) V-pose, (b) T-pose, (c) A-pose, (d) H-pose.

his work uses the H-pose.

Page 13: Heuristic misfit reduction: A programmable approach for 3D

W. Lee, H.-S. Ko / Computers & Graphics 71 (2018) 1–13 13

S

f

R

[

[

[

[

[

[

upplementary material

Supplementary material associated with this article can be

ound, in the online version, at 10.1016/j.cag.2017.10.004 .

eferences

[1] ESMOD . Become a pattern drafter: women’s garments, vol. 1. ESMOD Editions;

2012 . [2] College BF . Bunka fashion series garment design textbook 1 – fundamentals of

garment design. Bunka Fashion College; 2009 . [3] Apeagyei PR, Otieno R. Usability of pattern customising technology in the

achievement and testing of fit for mass customisation. J Fash Mark Manag Int

J 2007;11(3):349–65. doi: 10.1108/13612020710763100 . [4] Seo H, Kim S-J, Cordier F, Hong K. Validating a cloth simulator for measuring

tight-fit clothing pressure. In: Proceedings of the 2007 ACM symposium onsolid and physical modeling (SPM ’07). New York, NY, USA: ACM; 2007. p. 431–

7. ISBN 978-1-59593-666-0. doi: 10.1145/1236246.1236308 . [5] Lee J, Nam Y, Cui MH, Choi KM, Choi YL. Fit evaluation of 3D virtual

garment. In: Proceedings of the second international conference on usabil-

ity and internationalization (UI-HCII’07). Berlin, Heidelberg: Springer-Verlag;2007. p. 550–8. URL http://dl.acm.org/citation.cfm?id=1769821.1769888 . ISBN

978-3-540-73286-0 . [6] Cordier F, Seo H, Magnenat-Thalmann N. Made-to-measure technologies for an

online clothing store. IEEE Comput Graph Appl 2003;23(1):38–48. doi: 10.1109/MCG.2003.1159612 .

[7] Wang CC, Wang Y, Yuen MM. Feature based 3D garment design through 2D

sketches. Comput-Aided Des 2003;35(7):659–72. doi: 10.1016/S0010-4485(02) 0 0 091-X .

[8] Wang CCL, Wang Y, Yuen MMF. Design automation for customized apparelproducts. Comput Aided Des 2005;37(7):675–91. doi: 10.1016/j.cad.2004.08.

007 . [9] Meng Y, Wang CCL, Jin X. Flexible shape control for automatic resizing of ap-

parel products. Comput Aided Des 2012;44(1):68–76. doi: 10.1016/j.cad.2010.11.

008 . [10] Li J, Ye J, Wang Y, Bai L, Lu G. Technical section: fitting 3D garment models

onto individual human models. Comput Graph 2010;34(6):742–55. doi: 10.1016/j.cag.2010.07.008 .

[11] Li J, Lu G. Customizing 3D garments based on volumetric deformation. ComputInd 2011;62(7):693–707. doi: 10.1016/j.compind.2011.04.002 .

[12] Wang CCL, Hui KC, Tong KM. Volume parameterization for design automationof customized free-form products. IEEE Trans Autom Sci Eng 2007;4(1):11–21.

doi: 10.1109/TASE.2006.872112 . [13] Brouet R, Sheffer A, Boissieux L, Cani M-P. Design preserving garment transfer.

ACM Trans Graph 2012;31(4) . 36:1–36:11. doi: 10.1145/2185520.2185532 [14] Lee Y, Ma J, Choi S. Technical section: automatic pose-independent 3D garment

fitting. Comput Graph 2013;37(7):911–22. doi: 10.1016/j.cag.2013.07.005 . [15] Huang HQ, Mok PY, Kwok YL, Au JS. Block pattern generation: from parameter-

izing human bodies to fit feature-aligned and flattenable 3D garments. Comput

Ind 2012;63(7):680–91. doi: 10.1016/j.compind.2012.04.001 . [16] Umetani N, Kaufman DM, Igarashi T, Grinspun E. Sensitive couture for interac-

tive garment modeling and editing. ACM Trans Graph 2011;30(4) . 90:1–90:12.doi: 10.1145/2010324.1964985 .

[17] Berthouzoz F, Garg A, Kaufman DM, Grinspun E, Agrawala M. Parsing sewingpatterns into 3D garments. ACM Trans Graph 2013;32(4) . 85:1–85:12. doi: 10.

1145/2461912.2461975 .

[18] Bartle A, Sheffer A, Kim VG, Kaufman DM, Vining N, Berthouzoz F. Physics-driven pattern adjustment for direct 3D garment editing. ACM Trans Graph

2016;35(4) . 50:1–50:11. doi: 10.1145/2897824.2925896 . [19] Veblen S . The complete photo guide to perfect fitting. Minneapolis, MN, USA:

Creative Publishing International; 2012 . Complete Photo Guide. 20] Baraff D, Witkin A, Kass M. Untangling cloth. ACM Trans Graph

2003;22(3):862–70. doi: 10.1145/882262.882357 .

[21] Volino P, Magnenat-Thalmann N. Resolving surface collisions through intersec-tion contour minimization. ACM Trans Graph 2006;25(3):1154–9. doi: 10.1145/

1141911.1142007 . 22] Baraff D, Witkin A. Large steps in cloth simulation. In: Proceedings of the

twenty-fifth annual conference on computer graphics and interactive tech-niques (SIGGRAPH ’98). New York, NY, USA: ACM; 1998. p. 43–54. ISBN 0-

89791-999-8. doi: 10.1145/280814.280821 .

23] Choi K-J, Ko H-S. Stable but responsive cloth. ACM Trans Graph2002;21(3):604–11. doi: 10.1145/566654.566624 .

24] Grinspun E , Hirani AN , Desbrun M , Schröder P . Discrete shells. In: Proceed-ings of the 2003 ACM SIGGRAPH/eurographics symposium on computer ani-

mation (SCA ’03). Aire-la-Ville, Switzerland, Switzerland: Eurographics Associ-ation; 2003. p. 62–7. ISBN 1-58113-659-5 .

25] ASTM. Standard terminology relating to body dimensions for apparel sizing.

ASTM, D5219-09. West Conshohocken, PA, USA: ASTM International; 2009.doi: 10.1520/D5219-09 .

26] ISO. Garment construction and anthropometric surveys – body dimensions.ISO, 8559:1989. Geneva, Switzerland: International Organization for Standard-

ization; 1989 . URL http://www.iso.org .