A customized garment collaborative design process by using virtual reality and sensory evaluation on garment fit

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

Computers & Industrial Engineering journal homepage: www.elsevier.com/locate/caie

A customized garment collaborative design process by using virtual reality and sensory evaluation on garment fit ⁎

Xuyuan Tao , Xiao Chen, Xianyi Zeng, Ludovic Koehl Université Lille 1 Sciences et Technologies, 59655 Lille, France GEMTEX, ENSAIT, 2 allée Louise et Victor Champier, 59056 Roubaix Cedex 1, France

A R T I C L E I N F O

A B S T R A C T

Keywords: Virtual garments Fit preference Collaborative design Sensory evaluation Computational model

As the successful implantation of CAD (computer-aided-design) technology in garment industry, the 3D virtual garment technology has attracted a great attention of textile/garment companies. However, wearer’s perception on virtual garment in terms of fitting and comfort has never been systematically studied. In this paper, we propose an original customized 3D garment collaborative design process by integrating interactions between the designer and the specific consumer. In this context, a normalized sensory evaluation on garment fitting effects will be organized in a virtual environment, in order to enhance communications of the concerned actors on perception of products. Also, by learning from the measured distances between the garment surfaces and the mannequin (input) and sensory descriptors (output), we model the relationship between the garment design parameters and the human perception on fit of the finished virtual products. Using this model, we can estimate the fit perception for a specific garment size on a specific body shape without any real try-on experience. In practice, the proposed collaborative design process will permit to develop an online recommendation system for garment size selection and fit estimation. Furthermore, it will permit to recursively modify the initial garment patterns according to the consumer’s fit preferences so that they can obtain a real personalized garment. In this way, the consumer will be directly involved in the product design process by performing a series of sensory evaluations on virtual garment fit in order to obtain a desired finished product. This process has been validated by creating a collection of T-shirts meeting requirements of various customers.

1. Introduction In modern society, under the economic pressure and faced to more and more demanding and various requirements of consumers in terms of fashion style, comfort and functionality, classical garment design and production need to be extensively innovated in order to deliver rapid and personalized products and services of high quality with low costs. The quick progress of information technologies such as virtual reality technology, e-business, mass customization and collaborative design, provide new opportunities for restructuring the entire textile/garment/ distribution supply chain in a more optimized way. The sizing system of a garment involves a series of grading rules and operations to enlarge or reduce a basic pattern to one or several sizes (Gilewska, 2008). In a classical garment design process, the designer first creates all the patterns for a dedicated size, called the basic size. Then they integrate the grading rules into each feature point of these patterns in order to proportionally increase or decrease the outlines of the original patterns and generate garments of the same style but with different sizes. Generally, the grading rules are developed from ⁎

empirical knowledge of the garment designer by interpreting the body measurements of standard human models and ease allowance on different parts of the body. The grading rules can also vary with garment style and nature of materials used. Based on generation of physical prototypes, the classical product design process is rather long, leading to high costs in human resource and production. In this context, the concept of collaborative design or co-design has been widely introduced in many industrial sectors such as automobile, furnishing, architecture, manufacturing and civil engineering for quickly delivering user-centered new products to the market. Collaborative design represents the experience or process of a collaborative construction of new content in a virtual environment represented as avatars by multiple users through their communication and shared exploration of 3D visualizations (Luo & Yuen, 2005). The basic element of collaborative design is the Internet-based visual platform showing interactions between the product, the designer and the customer (García-García, Chulvi, & Royo, 2017). Based on 3D visual models, new design methodologies, such as scenario-oriented design or

Corresponding author at: 2 allée Louise et Victor Champier, BP 30329, 59056 Roubaix Cedex 1, France. E-mail address: [email protected] (X. Tao).

http://dx.doi.org/10.1016/j.cie.2017.10.023

0360-8352/ © 2017 Elsevier Ltd. All rights reserved.

Please cite this article as: Tao, X., Computers & Industrial Engineering (2017), http://dx.doi.org/10.1016/j.cie.2017.10.023

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 1. Basic functions of the collaborative garment design process.

Fig. 2. Feature points in the patterns of a T-shirt. Fig. 3. Scheme of movements of a feature point for different sizes.

context-oriented design (Martinez-Maldonado et al., 2017), customized product design (Hosun, Istook, & Cassill, 2009), are developed for dealing with different application contexts. They permit to realize the product design by visualizing a virtual person using the virtual product in a virtual scenario. Collaborative design is very significant in garment industry because, compared with other domestic products, garments and their accessories are more human-centered in terms of body fitting, comfort and aesthetics. Conventional CAD technology using the size table of standard body measurements can provide a short product development cycle. However, it cannot produce accurately fitting garment for individual customers. 3D virtual garment design using specific CAD software can be considered as an optimal combination of designers, computer technology and animation technology, permitting to realize and validate design ideas and principles within a very short time (Tao & Bruniaux, 2013; Volino & Magnenat-Thalmann, 2012). New knowledge and design elements on garments can be obtained from human-machine interactions. The application of virtual technology can effectively accelerate new product development, reduce design and production cost, and increase product quality in order to enhance the competitiveness of garment companies in the worldwide market (Chan, Yang, Wong, Chan, & Lam, 2015; Vuruskan, Ince, Bulgun, & Guzelis, 2015; Wang, Zeng, Koehl, & Chen, 2015). The most appreciated CAD systems generating 3D virtual garments

are based on mechanical models (Provot, 1995). These models, built according to the mechanical properties of real cloth measured on devices such as KES and FAST (Philippe, Schacher, Adolphe, & Dacremont, 2004), can effectively simulate fabric deformable structures and be accurate enough to deal with nonlinearities and deformations occurring in cloth, such as folds and wrinkles. Moreover, they can strongly interact with wearers or human body models. In practice, the perception of virtual garments can be modified by adjusting the fabric technical parameters and garment patterns of the corresponding garment CAD software so that their visual and tactile effects are as close as possible to those of real garments. In this way, consumers and designers can control or enhance some sensory criteria such as softness, smoothness, rigidity, draping and fitting effects in virtual products according to their requirements. This approach is particularly interesting for designing new personalized products. In this paper, we propose a customized 3D garment collaborative design process by integrating interactions between the designer and the specific consumer in order to recursively control garment fitting to converge to the visual effect desired by the consumer. The human perception on garment fitting effects is quantitatively characterized by a normalized sensory evaluation procedure. By learning from experimental data measured on a set of representative garment samples, a mathematical model is built in order to characterize the relationship 2

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Table 1 Pearson’s correlation coefficient for all the feature points in x direction. Direction x

dx1 dx2 dx3 dx4 dx5 dx6 dx7 dx8 dx9 dx10 dx11 dx12 dx13

dx1

dx2

dx3

dx4

dx5

dx6

dx7

dx8

dx9

dx10

dx11

dx12

dx13

1 1 0.99 0.99 0.99 −0.97 −0.97 −0.97 −1 1 1 −1 −1

1 0.99 0.99 0.99 −0.96 −0.96 −0.96 −1 1 1 −1 −1

1 1 1 −0.99 −0.99 −0.99 −0.98 0.99 0.99 −0.99 −0.99

1 1 −0.99 −0.99 −0.99 −0.98 0.99 0.99 −0.99 −0.99

1 −0.99 −0.99 −0.99 −0.98 0.99 0.99 −0.99 −0.99

1 1 1 0.95 −0.97 −0.97 0.97 0.97

1 1 0.95 −0.97 −0.97 0.97 0.97

1 0.95 −0.97 −0.97 0.97 0.97

1 −1 −1 1 1

1 1 −1 −1

1 −1 −1

1 1

1

Table 2 Linear relationship of movements between the feature points and the key points between each adjacent size. Direction x: key movement dx3 dx1 dx2 dx4 dx5 dx6 dx7 dx8 dx9 dx10 dx11 dx12 dx13

dx1 = 0.493·dx3 − 1.012 dx2 = 0.975·dx3 − 1.013 dx4 = dx3 dx5 = dx3 dx6 = −0.507·dx3 − 1.012 dx7 = −0.507·dx3 − 1.012 dx8 = −0.507·dx3 − 1.012 dx9 = −0.482·dx3 + 0.002 dx10 = 0.401·dx3 + 0.801 dx11 = 0.242·dx3 − 0.595 dx12 = −0.249·dx3 + 0.419 dx13 = −0.403·dx3 − 0.889

Table 3 Relationship between feature lines and corresponding 3D garment measurements.

Direction y: key movement dy5, dy11

3D garment measurement gmi

fGMi

dy1 dy2 dy3 dy4 dy6 dy7 dy8 dy9 dy10 dy12 dy13

Neck size (gm1) Shoulder size (gm2) Breast size (gm3) Waist size (gm4) Hip size (gm5) Torso height (gm6) Biceps size (gm7) Armhole depth Arm length (gm9)

gm1 = 2 * l1 gm2 = 2 * l2 gm3 = 2 * l3 gm4 = 2 * l4 gm5 = 2 * l5 gm6 = l6 gm7 = l7 gm8 = l8 gm9 = l9

0 dy2 = 0.027·dy5 + 0.112 dy3 = 0.617·dy5 + 3.401 dy4 = dy5 dy6 = dy5 dy7 = dy5 dy6 = 0.617·dy5 + 3.401 dy7 = 0.028·dy5 + 0.115 dy8 = 0.423·dy5 + 3.73 dy12 = 0.993·dy11 − 0.006 dy13 = 0.423·dy5 + 3.73

order to generate a new virtual garment, whose fitting effect is closer to that desired by the consumer. In this way, the proposed collaborative design process will permit to recursively modify the initial garment patterns according to the consumer’s fit preferences so as to obtain a real personalized garment. And the consumer can directly be involved in the product design process by performing a series of sensory evaluations on virtual garment fit in order to obtain a desired finished product. The effectiveness of the proposed design process has been validated by creating a collection of T-shirts meeting requirements of various customers. This paper is organized as follows. In Section 2, we give the general scheme describing the structure and principles of the proposed collaborative design process. In Section 3, the basic concepts and data are formalized mathematically and the key features of garment patterns are extracted. Based on these key features, the garment design parameters, describing the distance between the garment surface and the human body surface, are determined. In Section 4, the virtual garment samples used in our study and the corresponding sensory evaluation procedure are described. In Section 5, the procedure for modeling the relation between garment design parameters and the normalized sensory descriptors describing garment fits is established. Finally, examples of application are presented. For simplicity, this paper only deals with the influence of garment patterns on fitting effects. The control of fabric parameters on perception of finished garments has been discussed in our previous paper (Chen, Tao, Zeng, Koehl, & Boulenguez-Phippen, 2015).

dy5 = −0.7862·dx3 − 4.7562

Fig. 4. Definition of the feature lines on the T-shirt patterns.

between the garment design parameters, namely the distance between the garment surface and the mannequin, and the normalized sensory descriptors describing the garment fit. Using this model, we can estimate the fit perception for a specific garment size on a specific body shape without any real try-on experience. Also, according to the difference between the consumer’s perception on fitting effect and his/her desired value, this model will permit to adjust the garment patterns in

2. General scheme of the collaborative design process In our previous study, a collaborative garment design process (platform) was proposed. Its basic functions are given in Fig. 1 as follows. In this platform, the learning data are first collected by performing 3

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 5. Body measurements of the female mannequin.

rules. Although one garment can be composed of a number of patterns characterized by enormous feature points, the movements of these feature points are generally not independent but have strong correlation among them. During grading, they always have the same trend leading to enlarge or reduce the pattern without changing its basic shape. In this context, we firstly explore the relationship between these movements in order to identify the independent feature points. The selected T-shirt consists of 4 patterns, i.e. one front piece, one back piece and two symmetric sleeve pieces. The front and back pieces are centrosymmetric. The silhouette of the front piece is almost the same as the back piece except for the curve shape at the neckline. In order to simplify the complexity of the study, we only focus on the back piece and left sleeve. Fig. 2 shows the back and sleeve patterns for the T-shirt. The green pattern represents the silhouette of the basic size. The other contours with different colors1 represent the pattern silhouettes for different sizes. In each pattern, there is an alignment point, considered as the common departure of contours for all sizes. Usually, its coordinates are taken as original point on the plan of the pattern. Hence, the remaining points on the pattern can be considered as the feature points of the garment. Particularly, the selection of feature points ignores the symmetry feature of back piece in order to show that this method could be applied in other more general patterns. During the grading process for different sizes, the movements of a feature point can be decomposed into horizontal movement (direction x) and vertical movement (direction y), as shown in Fig. 3. Let dx i and dyi be the movement of the point Pi in x and y directions from the basic size tk . The coordinates of the feature point Pi can be defined as follows.

three sensory experiments (Experiments I, II and III). Experiment I permits to estimate the most appropriate values of the fabric parameters for each specific sample. Combined with a specific garment pattern, these values constitute the inputs to the garment CAD software for generating the corresponding virtual garment. Experiment II aims at extracting the normalized tactile and visual sensory descriptors for quantitatively characterizing the consumers’ or designers’ perception on the concerned virtual fabric. These works are already detailed in our previous study (Chen, Tao, Zeng, Koehl, & Boulenguez-Phippen, 2014; Chen et al., 2015). This paper mainly focuses on Experiment III, a normalized sensory evaluation procedure for acquiring human perception on the virtual garment fitting effects of a collection of garment samples. Meanwhile, the garment design parameters are quantitatively extracted from the patterns. From the acquired quantitative learning data, a model has been set up to characterize the relationship between the pattern design parameters and the consumer’s fitting perception on the corresponding virtual garment by controlling different ease allowance values. This model has two major functions. First, it will be able to estimate the garment fitting on a new wearer from his/her basic body measurements without any real try-on experience. Second, it will automatically adjust the initial parametric patterns according to a fitting requirement desired by the consumer. 3. Mathematical formalization of garment patterns In this study, T-shirt patterns of all sizes are taken as representative garment samples. The sizes are XXS, XS, S, M, L, XL, XXL, XXXL, among which S is the basic size. Next, we will extract feature points and key points of these patterns.

x i (t j ) = x i (tk ) + dx i

1 ⩽ i ⩽ n, 1 ⩽ j ⩽ m

(1)

3.1. Pattern feature points and key points

yi (t j) = yi (tk) + dyi 1 ⩽ i⩽ n, 1 ⩽ j⩽ m

Let t = {t1,…,tk,…,tm} be a set of m different sizes where tk denotes the basic size. Let P = {P1,…,Pn} be a set of n feature points in the pattern of size tk. Let xi(tj) and yi(tj) be the x and y coordinates for the feature point Pi in size tj. In garment design, the feature points are basic elements on the pattern, determining the general silhouette of the garment. Different pattern sizes can be interpreted as ‘movements’ of these feature points from their original positions in the basic pattern by following grading

The grading movement is a special movement by following the traditional grading rules. Without loss of generality, the modification of a pattern can be interpreted as the proportional movements of the feature points from the basic pattern. Let dx and dy be the set of grading movements of all feature points in x and y directions.

(2)

1 For interpretation of color in Figs. 2 and 5, the reader is referred to the web version of this article.

4

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 6. Creation of sewing lines on the patterns (a) creation of patterns in the specific software and (b) creation of sewing in the 3D simulation software.

Let dy∗ = {dyc },dyc ∈ dy be the grading movements for the key points Pc in y direction. The movement of any other feature point Pq can be linearly derived from its correlated key feature point by the following equation:

As mentioned previously, in order to keep the silhouette of pattern, the movements of feature points are not totally independent with each other. Some independent features points can be identified by calculating the correlation coefficients between all the features points. These independent feature points are defined as key points, denoted as Pc. Their independent movements are defined as key movements. Let dx ∗ = {dx c },dx c ∈ dx be the grading movements for the key points Pc in x direction.

dx q = a0 + a1·dx c

(3)

dyq = b0 + b1·dyc

(4)

where a 0 , a1, b0 and b1 are the coefficients. 5

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

the alignment point and one feature point. Let P li be the set of the feature points of the associated feature line li . They could be the points with the same x or y coordinates of the extremity points or pass-by points on the feature line li . Let L = {l1,…,lq} be the set of the feature lines, where li = fLi (Pj,Pk,…),Pj,Pk ∈ P li , be a feature line on the pattern, fLi the feature line function, and Pj, Pk the feature points. On the T-shirt patterns, 9 feature lines, l1 … l9, are defined as shown in Fig. 4. Conforming to the principle of simplicity, we use the x or y direction straight distance to present the feature lines. For example, l 7 = fL7 (P13,P10) = |x13−x10 | and l6 = fL6 (P6) = |y6 |. By performing a pattern grading or modification, the feature line li varies to li′. We obtain the following relations: Let dli = fDLi (dl xi,dl yi ) be the variation of length from li to li′, where fDLi is a function to calculate it. Let dl xi = fDLXi (dx j,dxk,…),Pj,Pk ∈ P li be the variation in x direction where fDLXi is a function to calculate it. Let dl yi = fDLYi (dyj,dyk,…),Pj,Pk ∈ P li be the variation in x direction where fDLYi is a function to calculate it. From Figs. 2 and 4, we can derive the relationship between the variation of the feature lines l3 and the movements of its feature points for the T-shirt patterns as follows.

P l3 = {P3,P8}

(5)

dl x 3 = fDLX 3 (dx3,dx 8) = dx3−dx 8

(6)

dl y3 = fDLX 3 (dy3,dy8) = 0

(7)

dl3 = fDL3 (dl x 3,dl y3) = dl x 3 = dx3−dx 8

(8)

A garment measurement in 3D can be derived from one or several corresponding feature lines, corresponding to the body measurement at the same position, such as chest size, arm length, etc. Let GM = {gmi },gmi = fGMi (l j,lk,…) be a set of garment measurements, where fGMi is the garment measure function and l j,lk the corresponding feature lines. Let BM = {bmi} be a set of human body measurements, in which each bmi corresponds to one gmi . Table 3 shows the garment functions for the T-shirt patterns. The ease allowance between the garment and the human body at the body position i is the difference between gmi and bmi . Let MD = {md1,…,mds} be a set of differences between the human body measurement and the 3D garment measurement. Each element mdi (i ∈ {1,…,s} ) has a bijective correspondence in the set BM and GM. It is calculated by mdi = gmi −bmi .

Fig. 7. Arrangement of the patterns around the avatar.

Table 1 shows the Pearson’s correlation coefficients for movements of the feature points in x direction of the T-shirt patterns. From this table, we can find that the movements of all the feature points in x direction have strong correlation between them, whose coefficient values are higher than a predefined threshold τ (τ = 0.95). According to the same principle, the movements in y direction have the similar result: the dy1,…,dy10 are correlated between them while dy11 and dy12 are correlated as well. As a result, the 26 variables (dx1,…,dx13,dy1,…,d y13) derived from the 13 feature points can be reduced to only 3 variables (dx3,dy5 and dy11) . We obtain Pc = {P3,P5,P11} , dx ∗ = {dx3} and dy∗ = {dy5,dy11} . Moreover, there may exist correlations between movements in x and y directions. For instance, among the 3 identified key movements, dx3 and dy5 are strongly correlated with a coefficient of 0.98. Therefore, dy5 can also be described by a linear function of dx3 as well. Table 2 shows the linear relationship of movements between feature points and key points for T-shirt patterns. These relationships will be used during the pattern grading procedure when the designer wants to proportionally enlarge or decrease the patterns according the consumer’s fitting requirement. Generally, the grading of all feature points can be derived from the grading of dx ∗ and dx ∗. Based on this analysis, we establish the relationship between the pattern silhouette and garment measurements in the next section.

4. Acquisition of the learning data on fitting perception of virtual garments 4.1. Virtual garment generation In our study, a sensory experiment is carried out in order to quantitatively characterize these garments’ fit on different positions of the specific body in a virtual environment. For this purpose, one standard female mannequin is used as support of the garment visualization. Her principle body measurements are shown in Fig. 5. The red lines represent the projections of feature lines on the body. The patterns of garment (T-shirt in this study) are generated in the specific pattern making software and then imported into the simulation software. Different grading sizes of garment will be individually imported to generate different garments with identical silhouettes but different sizes. All these garments will be virtually sewn in the simulation software as shown in Fig. 6. Then every piece of pattern will be aligned to the corresponding position of the body as shown in Fig. 7. After determining the material parameters and texture of fabric, the simulation of virtual try-on will be realized. Different poses or animations can be imported into the

3.2. Pattern feature lines and 3D garment measurements The feature points characterize of the patterns’ silhouettes while the feature lines characterize their measurements. A feature line is a distance between two feature points. It could also be the distance between 6

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 8. Virtual simulation effect of the garments in (a) static scenario and (b) dynamic scenario.

simulation software in order to generate the static or dynamic scenarios (Fig. 8). The simulation results can be captured as images or videos for the sensory evaluation experiments.

Table 4 List of sensory descriptors. Ref.

Description of sensory descriptors

D1 D2 D3 D4 D5 D6 D7 D8

Neckline width Shoulder width Chest Waist Garment length Sleeve length Arm restriction in the movement Torso restriction in the movement

4.2. Acquisition of the learning data on perception of virtual garment fit In our sensory experiment, 8 sensory descriptors are selected by the experts of garment design for describing the garment fit on different positions of the wearer in both static and dynamic scenarios (Table 4). Meanwhile, a scale of 9 evaluation scores, varying from −4 to 4, is used for evaluating each sample on each sensory descriptor. The sign ‘−’ indicate the garment is too loose/big/long while ‘+’ signifies the

Table 5 Score scale of sensory descriptor and their semantic descriptions. Score

−4

−3

−2

−1

0

1

2

3

4

Description

Extremely tight

Very tight

Quite tight

A little tight

Perfect

A little loose

Quite loose

Very loose

Extremely loose

7

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 9. Scheme of the model characterizing the relation between garment design parameters and sensory descriptors.

5.1. Formalization of the sensory evaluation data

Table 6 The inputs and outputs of the sub-models for the T-shirt samples. Sub-model i

Inputs MDi

Outputs Eval Di

1

md1

Eval D1

2

md2

Eval D2

3

md3

Eval D3

4

md4

Eval D4

5

md6

Eval D5

6

md2, md9

Eval D6

7

md7, md8

Eval D7

8

md2, md3, md4

Eval D8

The sensory evaluation data involved in this study are formalized as follows. Let PL = {pl1,…,plr } be a set of r panelists participating in our sensory experiment. Let D = {D1,…,Dp} be a set of p sensory descriptors on garment fit perception. For the descriptor Di, its evaluation value from the sensory experiment is denoted as:

Eval Di = {di (t j )}, i ∈ {1…p} and j ∈ {1…m}

(9)

where di (t j ) is the averaged value of the scores given by all the panelists for the garment of size t j related to the descriptor Di. It varies from −4 to 4. It is calculated by the following equation:

opposite meanings. The sign ‘0’ shows a perfect fit on the wear for a specific descriptor. Each of these scores is defined semantically in Table 5. One panel of 6 consumers is recruited for realizing the sensory experiment. They are students or researchers with the textile and apparel knowledge background. A training session is organized before the evaluation sessions in order to help the evaluators to observe and understand fitting perception in a systematic way. In this session, a number of virtual garment images with typical fits are shown to the evaluators. Then, during each evaluation session, each evaluator starts by giving an overall fitting perception on the photos and videos of the virtual garments displayed. This step can help the evaluators not only to catch an overview about the garment fitting but also ‘pre-position’ each virtual garment in the ‘scale’ for each sensory descriptor. Next, each evaluator is free to compare, weigh and judge differences between any pair of the garments, and then give a linguistic score from Table 5 for each sensory descriptor.

r

di (t j ) =

1 · ∑ di (t j,plk ) i ∈ {1…p} and j ∈ {1…m} r k=1

(10)

where di (t j,plk ) is one evaluation value on the garment of size t j , given by the panelist plk . 5.2. Modeling of the relation between garment ease allowance and sensory descriptors There exist various definitions of the ease allowance of a garment. In our study, the ease allowance at the position i is represented by mdi , i.e. the difference value between the 3D garment feature line gmi and the human body measurement bmi . For example, the chest circumference is 90 cm, and the garment contour in the chest position is 92 cm. The ease allowance equals to 2 cm. For a specific person, the fit perception can be adjusted by the modification of the garment patterns. In this section, we develop a model, permitting to quantitatively characterize the relationship between sensory descriptors on garment fit and the ease allowance. Using this model, a consumer can predict the fit perception of a given virtual garment from his/her own body measurements. Furthermore, this model can generate the new patterns according to the consumer’s specific fit preferences. As discussed previously, a fitting effect can be generated by adjusting the distance between the 3D garment surface and the human body surface. In the context, the key issue is to model the relationship between the ease allowance MD for different body positions (inputs) and evaluation scores on different sensory descriptors EvalD (outputs). In the modeling process, we firstly select the relevant inputs for each output according to experts’ knowledge. In practice, a specific sensory

5. Modeling of the relation between design parameters and sensory descriptors on garment fitting The fitting effect is strongly related to the garment ease allowance, i.e. the distance between the garment surface and the body surface. The following text introduces the methods for extracting the sensory data on garment fit and related garment ease allowance.

8

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 10. Fitting curves of the sub-models for 8 descriptors (D1-D8). Fig. 11. Schema of application of fitting perception on a new body morphotype and garment.

all the p descriptors, in order to obtain a general model (Fig. 9). Table 6 shows the inputs and outputs of each sub-model for the Tshirt samples. According to the experts’ knowledge, the relationship between the fitting evaluation and the garment size roughly follows a sigmoid function. Each sub-model is derived by using the nonlinear

descriptor is related to a few number of body measurements only. For example, the descriptor “contact with waist” is strongly relevant to the distance around waist, but not to that of neck at all. Secondly, based on the selected relevant inputs, we create a sub-model for each sensory descriptor. Finally, we combine all the p sub-models, corresponding to 9

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 12. Simulation result of new garment size and body morphotype in the 3D simulation software.

8

least square algorithm (Moré & Sorensen, 1983). Fig. 10 shows the submodels for the sensory descriptors D1 to D8 . ‖MDi ‖

Eval Di = fEVAL (MDi ) =

∑ j=1

⎞ ⎛ min ⎜∑ βi × Eval D2 i ⎟ , ∀ βi ⩾ 0 and ⎠ ⎝ i=1

8 αij ⎛ −4⎞, mdij ⎝ 1 + e−kij (mdij − cij) ⎠

∑ βi = 1

(13)

where βi is a coefficient of evaluation score on different position. This value can be given by the customer according to their preferences.

Ni

∈ MDi , αij > 0 and

∑

αij = 1

j=1

6. Applications of the model

(11)

6.1. Prediction of the virtual garment fitting for a body morphotype

where kij is the steepness of the sigmoid curve and cij the midpoint of sigmoid, ‖MDi ‖ is the count of mdij in MDi . The cij can be considered as the local optimal point for the mdij . When mdij = cij , the Eval Di = 0 . That means the fitting effect is perfect. All the experimental results have validated the previous sigmoid hypotheses because the determination coefficient R2 are very high (> 0.90). For the sensory descriptor, D7, as md9 does not vary for the Tshirt, the measure of distance, md9, is not considered in the model. For the descriptors D8, as the number of data is close or less than the number of the input parameters, PCA (Principle Component Analysis) is used for reducing the number of the inputs. The result shows that the first principle component contains 99% of the original data. Therefore, we can establish a linear model by using the first principle component only, denoted as pc, instead of md2, md3 and md4.

With the previous sub-models for each descriptor, we can estimate the fit perception of a virtual garment on a specific morphotype of body. When we receive a new data, including the new patterns (new size of the T-shirt) expressed by a set of garment measurement GM∗, and a new wearer with a set of body measurements at the same position BM∗, we calculate the ease allowance DM∗. Then we introduce this result into the p sub-models in order to obtain the fit perception on each sensory descriptor Dk (k ∈ {1,…,p} ). This process can be expressed by the following scheme. For example, for a special size garment GM∗ = {368.6, 822.4, 920, 920, 920, 746.2, 477, 235, 600} and a new body morphotype with the corresponding body measurement BM∗ = {338, 786, 930, 905, 930, 590, 274, 118, 576}, we calculate MD∗ = {30.6, 36.4, −10, 15, −10, 156.2, 203, 117, 24}. We obtain the fitting perception for descriptor D1 to D8 as followed: D1 = 0.014, D2 = 0.627, D3 = −0.155, D4 = −0.782, D5 = 2.669, D6 = 1.844, D7 = −0.034 and D8 = 0.013. These results give a general fit perception of this special size on the new body morphotype. In the neck (D1), chest (D3), arm movement (D7) and torso movement (D8), the fitting is almost perfect. However, in the shoulder (D2) and the waist position (D4), the garment is a little large. The garment length (D5) is considered as very large compared with the avatar. The sleeve (D6) is also very long for the avatar. The simulation result is shown in Fig. 12, which proves the fitting perception.

T

⎡ (md2−13.07)/145.8 ⎤ ⎡ 0.574 ⎤ pc = ⎢ (md3−68.50)/148.9 ⎥ × ⎢ 0.579 ⎥ ⎢ ⎥ ⎣ 0.579 ⎦ ⎣ (md4−282.50)/148.9 ⎦

(12)

5.3. Optimization for the global fitting The global fitting evaluation can be considered as a function of all mdi. However, it’s difficult to find a linear or nonlinear function to fit the global fitting effect. In order to get an optimal global fitting effect, a promising method is to take full advantage of the derived local submodels to obtain a global optimum. In all sub-models, some md are independent with each other for the different Eval Di , such as md1 and md6 . Some of them are used for several Eval Di in common, such as md2 , md3 . If the customer wants to have a global optimum fitting result, the genetic algorithm could be used to get the global optimum, MD opt , by minimize the objective function as follows (see Fig. 11).

6.2. Modification of virtual garment according to the fitting requirement This model can also be used to modify the existed pattern in order to satisfy the clients’ fitting requirements. It could be a choice of adapted size for a customer or a partial modification to create a new style. 10

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 13. Procedure for modifying the pattern in order to choose an adapted size or create a new style of the pattern.

feature points are moved according to the movement of key points, the new grading size can be achieved by keeping the silhouette. For the designed fitting scores EvalD∗ , we can obtain the corresponding difference measurements from the model MD∗. Compared with the body measurements, the new garment measurements GM∗ can be achieved. From Table 3, the new feature lines L∗ can be calculated. Compared to the original feature lines L, the movement of concerned feature points, dx i and dyi can be obtained. For different descriptors, the movements of different feature points are achieved. As a result, a new style of garment can be derived. If the customer wants to have a new garment size by following the grading rules to modify all feature points according to the requirement on one descriptor, the key points can be employed to obtain the movement for all the feature points. If we want to change the style of garment in Fig. 12 by making the shoulder position smaller and garment length shorter, we can make desired evaluation scores as Eval D∗ 2 = 0 for the sensory descriptor D2. According to the sub-model (Table 7), the new ease allowance

Table 7 Identified coefficients of the sigmoid function-based sub-models.

i

MDi

αij

kij

cij

R2

1 2 3 4 5 6 7

md1 md2 md3 md4 md6 md2 md7 md8 pc

α11 = 1 α21 = 1 α31 = 1 α 41 = 1 α51 = 1 α61 = 1 α71 = 0.61 α72 = 0.39 α81 = 1

k11 = 0.01858 k21 = 0.008686 k31 = 0.007735 k 41 = 0.009975 k51 = 0.02443 k 61 = 0.007839 k71 = 0.045 k72 = 2.826 k81 = 0.9611

c11 = 30.22 c21 = 0 c21 = 0 c21 = 54.7 c21 = 90.25 c21 = −90.82 c 71 = 171.6 c 72 = 160.7 c81 = −1.258

0.9226 0.9720 0.9919 0.9817 0.9551 0.9603 0.9954

8

0.9824

For an existing garment, based on the fit evaluation data and the customer’s fitting requirement, the model can also permit to modify the initial patterns in order to obtain the new personalized/customized patterns. The general procedure can be expressed in Fig. 13. If all the

11

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

Fig. 14. Virtual try-on for T-shirt and its pattern in 3D simulation software (left) original pattern, (middle) modified the should length according to the customer’s requirement, (right) new size by applying the grading rules.

0.5

Best: 0.139742

0.45

md 2∗ = c2 = 0 . The initial value for md2 is 36.4. Hence, we can obtain that the modification of feature line dl2 = (0−36.4)/2 = −18.2 . That means the pattern will be reduced about 18.2 mm at the shoulder position. The simulation results are shown in Fig. 14 (left and middle). The significant change can be observed. The new silhouette of T-shirt is more suitable in the shoulder position for the mannequin. Through the key points, the movement of feature can be applied to the other feature points by following the grading rules. As a result, a new size can be achieved for the customer. For example, a customer wants to have a new size with a little loose visual effect in chest position (Eval D3 changes from −0.155 to 0.5). Fig. 14 (right) shows the result of virtual try-on simulation. With the grading, we find that the new size is a little loose not only in the chest position but also in other positions such as shoulder, waist and garment length.

Best fitness Mean fitness

Fitness value

0.4 0.35 0.3 0.25 0.2 0.15 0.1

0

20

40

60

80

100

120

Generation

(a)

6.3. Modification of virtual garment according to the global fitting requirement If the customer requires having a global optimal effect, for the independent md, their values could be the local optimal values in Table 7. For instance, md1 = c11 and md6 = c51. The other md could be obtained by the genetic algorithm. Fig. 15(a) shows the genetic algorithm simulation result with a population of 500 and an adaptive feasible mutation function. We assume that βi = 1/8 for all the individual evaluation positions. The objective function value is 0.14 after 112 generation iterations. We obtain MDopt = {33.03, −38.42, 4.14, 57.19, 57.19, 90.25, 171.6, 160.7, 24}. From the given body measurements BM∗ = {338, 786, 930, 905, 930, 590, 274, 118, 576}, we can create garment with global optimal fitting effect GMopt = {368.22, 747.58, 934.14, 962.19, 987.19, 680.25, 445.6, 278.7, 600}. Fig. 15(b) shows the virtual try-on simulation result by using the optimal garment measurements. 6. Conclusion In this study, we have developed a customized 3D garment collaborative design process by integrating the customers’ fitting requirements. The human perception on garment fitting effects is quantitatively characterized by a normalized sensory evaluation procedure. By learning from experimental data measured on a set of representative garment samples, the mathematical sub-models for local positions are proposed. For each sub-model, the garment measurements and body anthropometric measurements are considered as input data. The

(b) Fig. 15. (a) Genetic algorithm result for the global fitting evaluation and (b) 3D simulation for the optimal global fitting effect.

12

Computers & Industrial Engineering xxx (xxxx) xxx–xxx

X. Tao et al.

García-García, C., Chulvi, V., & Royo, M. (2017). Knowledge generation for enhancing design creativity through co-creative Virtual Learning Communities. Thinking Skills and Creativity, 24, 12–19. Gilewska, T. (2008). Modélisme de mode T1 Coup à plat les bases (Vol. 1). EYROLLES. Hosun, L., Istook, C. L., & Cassill, N. L. (2009). Advanced mass customization in apparel. Journal of Textile and Apparel, Technology and Management, 6(1), 1–12. Luo, Z. G., & Yuen, M. M. F. (2005). Reactive 2D/3D garment pattern design modification. Computer-Aided Design, 37(6), 623–630. http://dx.doi.org/10.1016/j.cad.2004.09. 005. Martinez-Maldonado, R., Goodyear, P., Carvalho, L., Thompson, K., Hernandez-Leo, D., Dimitriadis, Y., ... Wardak, D. (2017). Supporting collaborative design activity in a multi-user digital design ecology. Computers in Human Behavior, 71, 327–342. Moré, J. J., & Sorensen, D. C. (1983). Computing a trust region step. SIAM Journal on Scientific and Statistical Computing, 4(3), 553–572. http://dx.doi.org/10.1137/ 0904038. Philippe, F., Schacher, L., Adolphe, D. C., & Dacremont, C. (2004). Tactile feeling: Sensory analysis applied to textile goods. Textile Research Journal, 74(12), 1066–1072. http:// dx.doi.org/10.1177/004051750407401207. Provot, X. (1995). Deformation constraints in a mass-spring model to describe rigid cloth behavior. Paper presented at the graphics interface 95, Quebec City. Tao, X., & Bruniaux, P. (2013). Toward advanced three-dimensional modeling of garment prototype from draping technique. International Journal of Clothing Science and Technology, 25(4), 266–283. http://dx.doi.org/10.1108/09556221311326301. Volino, P., & Magnenat-Thalmann, N. (2012). Virtual clothing: Theory and practice. Springer-Verlag Book Series. Vuruskan, A., Ince, T., Bulgun, E., & Guzelis, C. (2015). Intelligent fashion styling using genetic search and neural classification. International Journal of Clothing Science and Technology, 27(2), 283–301. Wang, L. C., Zeng, X. Y., Koehl, L., & Chen, Y. (2015). Intelligent fashion recommender system: Fuzzy logic in personalized garment design. IEEE Transactions on HumanMachine Systems, 45(1), 95–109.

sensory evaluation scores are output data. By using these models, we can estimate the fit perception for a specific garment size on a specific body shape without any real try-on experience. Through these local sub-models, customers can choose an adaptable garment or deign a new garment according to their fitting requirements. Through the genetic algorithm, the global optimal fitting effect can be achieved by modifying the silhouette of garment based on local sub-models. By integrating with the previous research about the materials design (Chen et al., 2015), an efficient collaborative design platform can be developed in order to permit customers involve in the collaborative design process and quickly identify personally material and fitting favorable requirements through a series of interactions with fashion designers. References Chan, A. P. C., Yang, Y., Wong, F. K. W., Chan, D. W. M., & Lam, E. W. M. (2015). Wearing comfort of two construction work uniforms. Construction Innovation, 15(4), 473–492. Chen, X., Tao, X., Zeng, X., Koehl, L., & Boulenguez-Phippen, J. (2014). Optimization of human perception on virtual garments by modeling the relation between fabric properties and sensory descriptors using intelligent techniques. Paper presented at the IPMU, Montpellie. Chen, X., Tao, X., Zeng, X., Koehl, L., & Boulenguez-Phippen, J. (2015). Control and optimization of human perception on virtual garment products by learning from experimental data. Knowledge-Based Systems, 87, 92–101. http://dx.doi.org/10.1016/ j.knosys.2015.05.031.

13