The construction of a three-dimensional human body based on semantic-driven parameters to improve virtual fitting

Abstract

The construction of 3D (three-dimensional) human model is the key link to the development of intelligent technology in garment engineering. In order to establish more convenient and efficient 3D human reconstruction methods based on semantic-driven parameters, we study semantic parameter prediction and 3D human reconstruction. Based on the self-collection database, we adopt the semantic parameter prediction method combining strong and weak correlations, and innovatively realize the automatic prediction of the remaining 12 parameters by entering the two parameters of height and weight. Then, the mapping relationship between the 14 semantic parameters and the 3D human template model is established to complete the personalized 3D human body reconstruction. Results show that the reconstruction error of each part of personalized 3D human body reconstruction is basically within ±1.5 cm, which meets the requirements of 3D human modeling for clothing size design. Further studies can use these results for the development and intellectualization of digital garment engineering

Keywords

Garment engineering semantic-driven parameter prediction personalization human reconstruction

The clothing industry is developing from the traditional production and sales mode in the digital and intelligent direction, such as in online virtual fitting, personalized clothing customization, etc. The efficient and accurate establishment of a three-dimensional (3D) human model that conforms to the size of the human body is the key to show the effect of fitting. Therefore, a precise 3D human body reconstruction method must be established that is suitable for the garment industry and capable of restoring its own characteristics.^1,2

Existing 3D human modeling methods involve 3D scanning technology, specific software, visual imaging, and semantic parameter (human body size data) driven parametric modeling.^3
–5 The equipment used to establish a 3D human model based on 3D scanning technology has a large volume and comes at high cost. Building a 3D human model based on specific software requires sufficient expertise and time. Furthermore, the establishment of a 3D human model based on visual images has multiple requirements for the shooting background and subject. In short, the above three methods not only rely on specific equipment, but also have high requirements for the used methods, so the efficiency of establishing a human body model is low, making it difficult to meet the rapid development needs of virtual fitting and other technologies in the clothing industry. A personalized 3D human model based on semantic-driven parameters can be established and customized through high-quality body shape dataset learning models and the input of body size data. This represents a convenient and low cost method for consumers. Multiple researchers have conducted studies on this technology. For example, Liu et al.⁶ used the traditional optimization algorithm depth frame to fit the Shape Completion and Animation of People (SCAPE) parameter space to generate the 3D human model. Yao et al.⁷ drove the pose and shape parameters of the skinned multi-person linear (SMPL) model to deform the 3D human body based on the Kinect sensor. The above methods efficiently deform the human posture and shape as well as muscle details, although the operation time is long. Huang et al.⁸ used the principal component analysis (PCA) space of the fully connected network regression characteristic curve to reconstruct a 3D human body based on a female model dataset. Several scholars have designed a mechatronics sensor network based on the characteristic curve to further optimize and establish the upper torso model of the human body.⁹ Although the error of these two methods is small, their use population is limited. Xie et al.¹⁰ reconstructed a 3D human body using five anthropometric parameters based on the graph convolution neural network algorithm. Among them, the circumference parameter was obtained based on a human slice ring to improve the reliability of the reconstructed data; however, a small number of dimensional parameters are difficult to accurately express in a personalized 3D human body.

Based on the above research, the current semantic parameter-driven 3D human body modeling method is relatively convenient in hardware, but there are still problems such as complex parameter acquisition and inadequate personalized representation of human body modeling. Human body modeling based on semantic parameters generally requires more size parameters to drive more accurate deformation of the human body, while in fact, most consumers only know their height and weight, and do not know many other parameters. Therefore, this article proposes a semantic parameter prediction method combining strong and weak correlations, and realizes the automatic prediction of the remaining 12 parameters by entering the two parameters of height and weight. Then, based on the SPRING 3D human body dataset, the mapping relationship between the 14 semantic parameters (for the 14 parameters, please refer to the annotation in Figure 1) and the 3D human template model is established to complete the personalized 3D human body reconstruction and deformation. Finally, it is verified by the comparison of model dimensions. This method will provide theoretical support for 3D human reconstruction technology for clothing.

Figure 1.

The 14 semantic parameters of the human body.

Construction of a parameterized three-dimensional human body model

Semantic-driven 3D human body model

Parametric human reconstruction consists of two parts, one is semantic parameters and the other is a 3D human dataset that can drive deformation. Human body reconstruction is used to establish the correlation between semantic parameters and the 3D human model to realize the deformation of the human model. For semantic parameters, according to the collection method in China's current national standards,¹¹ we manually collected 14 size parameters of 130 young men aged 18–23 living in northern China, and then trained the relationship between each parameter. The results of the training can realize semantic parameter prediction, so that users can enter the two parameters of height and weight to obtain the other 12 semantic parameters for human body reconstruction. The mapping relationship between the body part sizes of young men is obtained. This relationship is used for the linkage between body parts when the 3D model is deformed. Currently, European and American human body datasets are widely used,^12
–14 and there is no practical Asian human body dataset. Nevertheless, the datasets are universal, and the human body structure remains the same, with only small differences in body shape due to regions and habits. Therefore, we use the SPRING body model datasets as the source of the statistical analysis of the 3D human body model. These are high-resolution model datasets constructed by the combination of the SCAPE and Civilian American and European Surface Anthropometry Resource (CAESAR) datasets The datasets have point-to-point correspondence, and have been processed by loopholes and noise. The dataset of the human body model in a standard standing posture is represented by triangular mesh with consistent mesh topology,¹⁵ as shown in Figure 2.

Figure 2.

Human model of each dataset. CAESAR: Civilian American and European Surface Anthropometry Resource; SCAPE: Shape Completion and Animation of People.

Based on Python programming, the 3D human body model and corresponding size data can be used for virtual fitting, intelligent garment factories, and mass customization to realize an entirely data-driven process. The workflow of the personalized 3D human reconstruction system driven by semantic values is shown in Figure 3. The semantic parameter prediction and 3D human reconstruction is introduced in detail below.

Figure 3.

Flow chart of the three-dimensional (3D) human reconstruction system.

Semantic parameter prediction

The semantic parameter prediction takes 100 sets of size data from the self-collection database as the training set, and carries out prediction through strong correlation prediction and weak correlation prediction methods, respectively. This is performed to obtain the remaining 12 human semantic parameters required for modeling by only inputting the body height and weight.

To control the generation of the 3D human model by inputting dimension parameters, the correlation among the dimension parameters must be established first. From the manually collected data, 100 sets of data were randomly selected as sample sets for semantic parameter training to obtain the relationship among the 14 semantic parameters, while the remaining 30 sets are used as the verification sets to measure the accuracy of the predicted semantic parameters and the established 3D human model, The semantic parameters required for specific modeling are shown in Figure 1. The name of each part in the figure refers to the current national standard GB/T 16160-2017, Anthropometric definitions and methods for garment.¹⁶

Before establishing the linear regression model for prediction and analysis, we first must establish the linear relationship among the size parameters. If $A_{j}^{i}$ is used to represent the jth semantic parameter of the ith human body, the size parameters of different human bodies i are expressed as vectors $A^{i} = [A_{1}^{i}, A_{2}^{i}, A_{3}^{i}, \dots A_{14}^{i}] \in R^{14}$ . The Pearson coefficient ${Cor}_{k j}$ is used to measure the linear relationship between k and j among the 14 dimensional parameters,¹⁷ which is expressed as the following formula:

C o r_{k j} = \frac{\sum_{i = 1}^{n} (A_{k}^{i} - {\bar{A}}_{k}) (A_{j}^{i} - {\bar{A}}_{J})}{\sqrt{\sum_{i = 1}^{n} {(A_{k}^{i} - {\bar{A}}_{k})}^{2}} \sqrt{\sum_{i = 1}^{n} {(A_{j}^{i} - {\bar{A}}_{J})}^{2}}}

(1)

where n is the number of human body models input for training and

{\bar{A}}_{k}

and

{\bar{A}}_{J}

represent the measured average values of size parameters k and j. The range of the Cor_kj output value is (–1, 1), and zero means that there is no correlation between size parameters k and j. At threshold ρ = 0.8, a Cor_kj value greater than ρ or less than –ρ indicates a strong or weak linear correlation between size parameters, respectively.¹⁸

By establishing the correlation between the two parameters by Equation (1), Pearson correlation coefficient, it can be found that the chest girth, hip girth, and waist girth have a strong linear relationship with the body weight and body mass index (BMI), while other size parameters showed a weak linear relationship. Some correlation coefficients are shown in Tables 1 and 2.

Table 1.

Pearson coefficient of strongly correlated parameters

Position	Chest girth	Waist girth	Hip girth
Weight	0.9142	0.8449	0.9352
BMI	0.8814	0.8586	0.8704

Note: BMI is the human body size index, which can represent human body shape, BMI = body weight/height².

Table 2.

Pearson coefficient of some weak correlation parameters

Position	Neck-base girth	Thigh girth	Knee girth
Weight	0.4946	0.6301	0.6812
BMI	0.2783	0.5413	0.5437
Back shoulder width	0.5206	0.3470	0.3618
Inside leg length	0.1092	0.1663	0.2720
Arm length	0.1711	0.3003	0.2970

Note: BMI is the human body size index, which can represent human body shape, BMI = body weight/height².

After obtaining the linear relationship among the semantic parameters to more accurately predict the semantic parameters, we use the combination of strong and weak correlation prediction and use the linear regression algorithm to predict the chest girth, waist girth, and hip girth based on the two parameters of height and weight, and then use K-nearest neighbor (KNN) algorithm to predict the remaining nine parameters based on these five parameters, so as to establish the correlation among the 14 parameters and realize the prediction of various size parameters.

Strong correlation prediction

The strong correlation prediction associates the height, weight, chest girth, waist girth, and hip girth into a linear system, and uses the linear regression method to fit the semantic value function of the size parameters of the chest, waist, and hip girth. To achieve a better prediction effect, we select two parts in Table 1, where the number of the Pearson correlation coefficient is closer to 1, as the regression model for linear fitting. Hence, the chest girth and weight, hip girth and weight, waist girth and BMI are combined into three linear relationships, respectively. The regression formula ${\hat{y}}_{i} = α + β x_{i}$ is used to represent the linear relationship, where ${\hat{y}}_{i}$ is the estimated value of the mean E(y_i) of the dependent variable corresponding to the independent variable $x_{i}$ , α is the constant term of the sample regression equation, and β is the linear regression coefficient obtained from the experimental data training. The coefficient of the regression equation is estimated according to the minimum sum of squares of errors in the following formula, that is, the least squares method:

\sum_{i = 1}^{n} (y_{i} - {\hat{y}}_{i})^{2} = \sum_{i = 1}^{n} [y_{i} - (α + β x_{i})]^{2}

(2)

Finally, the linear regression prediction model of the chest girth, hip girth, and waist girth can be obtained, as shown in Table 3.

Table 3.

Linear regression prediction model of the chest, waist, and hip

Position	Linear regression prediction relation	R ²
Chest girth	${\hat{y}}_{x} = 44.881 + 0.621 x_{1}$	0.901
Waist girth	${\hat{y}}_{y} = 27.396 + 2.398 x_{2}$	0.925
Hip girth	${\hat{y}}_{t} = 58.878 + 0.557 x_{1}$	0.940

Herein, x₁ denotes weight, x₂ represents the BMI value, and the P-value indicates the significance of the regression model (the regression model is considered valid when the P-value is less than 0.05).¹⁸ After linear regression analysis, the value of the judgment coefficient R² is larger than 0.9, and the significance of the P-value is below 0.05. Therefore, the regression model is considered effective.

At this point, the mapping relationship among the chest, waist, and hip girth of Asian people and their height and weight has been established. By inputting the height and weight, the strongly correlated girth dimension parameters of the chest, waist, and hip can be obtained, and these five dimension parameters are represented by the dataset A_in.

Weak correlation prediction and verification

For other size parameters with a weak linear relationship, represented by the dataset A_out, the complete training sample set can be represented by $A^{i} = A_{i n} \cup A_{out}$ . When A_in is known, we use the KNN algorithm to predict and interpolate the missing variable A_out to form the optimal set of human size semantic parameters.¹⁹

Interpolation is used to predict the missing variables in the dataset through statistical analysis of the dataset. This paper implements and evaluates five common interpolation methods applicable to this dataset based on the Python language, as follows: the mean interpolation and variable interpolation in the machine learning (scikit-larn) package, IterativeImputer algorithm, KNN algorithm, multiple imputation by chained equations (MICEs) algorithm, and SoftInput algorithm in the fancysimple package.

We use the drawing tool Matplotlib to display the mean square error of each algorithm to intuitively evaluate the accuracy of each interpolation algorithm on the dataset; the results are shown in Figure 4.

Figure 4.

Comparison of five interpolation methods.1: Neck-base girth; 2: inside leg length; 3: back shoulder width; 4: arm length; 5: upper-arm girth; 6: waist girth; 7: chest height; 8: knee girth; 9: thigh girth. KNN: K-nearest neighbor.

It can be seen from the comparison in Figure 4 that the KNN algorithm has high stability on the dataset. Therefore, this article uses the KNN algorithm to perform weak correlation prediction on nine parameters, including the neck-base girth and inside leg length.

If the predicted parameters are discrete variables, the KNN classifier is used to vote for the most parameters in K neighbors for filling. If the prediction parameter is a continuous variable, the KNN regression is used to fill it with the average value of the variable in K neighbors.²⁰ Because the anthropometric size parameters, such as height, weight, and circumference, are continuous variables, we choose the KNN regression for prediction and interpolation. This process is based on the statistical analysis of the training sample set Aⁱ. The algorithm is as follows.

Input – feature space:

T = {(x_{1}, y_{1}), (x_{2}, y_{2}), \dots, (x_{n}, y_{n})}

In the training sample set, each data has a label, such that the corresponding relationship between the data in the sample set and the classification is clear. In the above feature space T, $x_{i} \in X \subseteq R^{n}$ is the feature vector of the training sample set and $y_{i} \in Y = {c_{1}, c_{2}, \dots c_{k}}$ is the category of the feature, where i = 1,2,…n. In the training phase of the algorithm, only the stored feature vectors and the labels of the training sample set are included.

Output – the predicted value of the vacancy characteristic variable to which a label belongs.

To illustrate the process, we build a model where X is a complete data matrix, and X_incomplete has the same value as X, except for several numeric variables that are replaced by NaN. Take the numerical variable x as an example in the following.

Firstly, according to the selected distance measure, k samples closest to x are found in the training set. The related neighborhood of x covering these k points is denoted as N_k(x).

Determine x in N_k(x) to category y according to the classification rules:

y = \arg \max_{c_{j}} \sum_{x_{i} \in N_{k} (x)} I (y_{i} = c_{j})

(3)

where I is the indicator function, I is 1 when y_i =c_j; otherwise, I is 0, i = 1, 2,···,n.

3. The variable values of K samples are used for the weighted average to estimate the missing data of samples. This method is the same as the mean interpolation and is single value interpolation. The difference is that it first uses the hierarchical clustering model to predict the type of missing features, and then fills the predicted values into the original dataset.

From the above process, we see that there are three important factors in the prediction and interpolation of missing values using the KNN algorithm: the selection of the K value, distance measurement, and decision rules.

Selection of the K value: the selection of the hyperparameter K has a significant influence on the algorithm. If the value of K is too small, the complexity of the model increases and the estimation error increases, leading to the over-fitting of the prediction results. The selection of a large K value can reduce the estimation error of training, but it will lead to a large training sample neighborhood, and distant samples will also participate in the decision-making and prediction process, which increases the approximate error and leads to under-fitting.²¹ To solve the above problems, we adopted the cross-validation method to determine the appropriate K value, as shown in Figure 5. Firstly, the entire training set X is divided into k disjoint subsets. Assuming that the number of training samples is m, each subset has m/k training samples. Then, one sample is selected from the subsets as the test set each time, such as the striped square in Figure 5, and the other k–1 as the training set, such as the blank square in Figure 5. Finally, we tested the test set based on a five-fold cross-validation, and constantly changed the K value to re-model, re-train, and evaluate it to obtain the score of the five models. The optimal K value of 11 was obtained by averaging the five scores and the corresponding score was 0.909890, as shown in Figure 6.

Figure 5.

The cross-validation process.

Figure 6.

Optimal K value and corresponding score.

Distance measurement: our continuous variable selects the Euclidean distance as the measurement, assuming that the two points in the n-dimensional space are X and Y:

X = {x_{1}, x_{2}, x_{3}, \dots, x_{n}}

Y = {y_{1}, y_{2}, y_{3}, \dots, y_{n}}

Then, the Euclidean distance is as follows:

D_{(x, y)} = \sqrt{\sum_{k = 1}^{n} {(x_{k} - y_{k})}^{2}}

(4)

Decision rules: the weighted average method is used for prediction interpolation:

f (x) = \frac{\sum_{i = 1}^{n} D_{i} W_{i}}{\sum_{i = 1}^{n} W_{i}}

(5)

where D_i represents the distance between the nearest neighbor i and the value to be predicted x, W_i represents its weight, and f(x) is the numerical result of the prediction.

After determining the factors of the strong correlation prediction and weak correlation prediction methods, the nine parameters obtained by the KNN algorithm and the previous five parameters will be used to drive the deformation of the 3D human modeling subsequently. Finally, based on the test set, the size of the deformed human modeling will be compared with the size of the test set, and the size error of the established 3D human modeling will be analyzed to verify the accuracy of the strong and weak correlation prediction methods.

Three-dimensional human body reconstruction

After obtaining 12 human body sizes from the semantic parameter prediction part, and combining the two parameters of height and weight, we obtain the mapping relationship between the semantic parameters and the model grid based on the 3D human body grid and the corresponding human body size and completed the semantic-driven 3D human deformation. The human body geometric mesh is composed of multiple triangular surfaces, where each surface is represented by V, the deformation of a single triangular surface is represented by the 3 × 3 linear transformation matrix Q, and the human body geometric mesh is represented by S_i. Currently, there are two classical methods to represent human body shape changes. One is the relationship between dimension parameters obtained by Ji²² using linear mapping and the PCA weight, namely the global mapping method. The process is expressed in the following equations:

U_{i} = S_{i} - \bar{S}

(6)

S = \bar{S} + w_{1} * D_{1} + w_{2} * D_{2} + \dots + w_{m} * D_{m}

(7)

W C = M

(8)

In Equation (6), vector U_i represents the variation of human body geometry and $\bar{S}$ represents the average vector of 3D human body geometry. To compress the data, PCA is used to reduce U_i to m dimensions to obtain a novel representation of human body geometry. For example, D_i represents the principal direction of PCA in Equation (7), $i \in 1, 2, \dots m$ . In Equation (8), W = {w₁,w₂,…,w_m} represents the PCA weight, and a novel human body geometry can be obtained by setting w. To more intuitively represent human body geometry, the linear mapping between size parameter C = {c₁,c₂,…,c_m} and W is used for setting, but w has no actual semantics, resulting in the neglect of the human body topological structure in the deformation process, such that the effect of a precise deformation of a single part cannot be achieved. Therefore, the local mapping method based on the global mapping idea completes the accurate deformation of a single part²³:

N ϕ ({[f_{1} \dots f_{L} ​ 1]}^{T}, i) = s_{i}

(9)

where

ϕ

represents multiple dimensional parameters that affect the ith triangle. Setting

ϕ

enables a single part to deform; however, this approach requires manually binding the dimensional parameters to the rigid part of the SPRING body based on prior knowledge. We use a similar method to learn the linear regression model between the deformation matrix Q and the most relevant parameters of each face to achieve 3D human body local mapping. In contrast, to evaluate the importance of semantic parameter features to each triangular face, obtain the dominant semantic features and automatically bind them to the relevant parts, we use the random forest algorithm.

The random forest algorithm measures the importance of feature according to Gini index and OOB error calculation rate.²⁴ Because there are both continuous and classification variables in our sample training set, and as there is correlation between continuous variables, we choose the OOB error calculation rate $V I M_{l}^{(OOB)}$ to observe the importance of feature l. Taking the importance of feature l as an example:

V I M_{l}^{(OOB)} = \frac{\sum_{p = 1}^{n_{o}^{i}} I (Y_{P} = Y_{P}^{i})}{n_{o}^{i}} - \frac{\sum_{p = 1}^{n_{o}^{i}} I (Y_{P} = Y_{P, π_{l}}^{i})}{n_{o}^{i}}

(10)

where

n_{o}^{i}

is the observation number of OOB in the ith tree, the exponential function is represented by I(g), and the value is 0 or 1. Here,

Y_{P} \in {0, 1}

is the true result of the lth observation; we replace the original eigenvalue with a uniform random value. The prediction result of feature l by the ith tree before the replacement is

Y_{P}^{i} \in {0, 1}

, and the prediction result after the replacement is

Y_{P, π_{l}}^{i} \in {0, 1}

. The importance of the final feature

l

V I M_{l}^{(OOB)}

. Similarly, the feature selection is carried out based on the importance of all features, and the importance of each feature is arranged in descending order. Then, according to the importance of features, the elimination ratio is determined, and corresponding features are eliminated to obtain new features. Finally, the above process is repeated in the new feature collection, and the feature set with the lowest OOB error calculation rate is selected. We take the grid at the hip girth as an example to evaluate the importance of the random forest. As shown in Figure 7, the order of importance according to size characteristics is weight, chest height, waist girth, etc.

Figure 7.

Ranking of feature importance in a grid at the hip. The meanings of each designation in the abscissa are, from left to right: weight, chest height, waist girth, chest girth, arm length, upper-arm girth, wrist girth, inside leg length, neck-base girth, height, back shoulder width, thigh girth.

Similarly, the importance score is obtained by the random forest algorithm for each face, and t parameter vectors F′ that are most relevant to each face are obtained in descending order:

F^{'} = {(f_{1}, f_{2}, \dots f_{t})}^{T}

(11)

Finally, we learn the linear regression model Q′ = MF′ between the deformation matrix M and the most relevant parameter F′ of each face. The deformation ${S^{'}}_{i} = {({Q^{'}}_{i, 1}, {Q^{'}}_{i, 2}, \dots, {Q^{'}}_{i, m})}^{T}$ of n human meshes with m plane numbers is obtained, and the local mapping of the 3D human body is realized.

Results and discussion

Analysis of results

To verify the accuracy of using semantic values to establish 3D human models based on our method, the verification is divided into two aspects: on the one hand, we use the Anthroscan M4 Plus 3D human scanner to scan 30 men in the verification set to obtain 3D human models. We also compare the 3D human model obtained by scanning with the 3D human model established by our method (taking two people with different body types as an example), as shown in Figures 8 and 9. On the other hand, based on the method in this paper, the height and weight of 30 men in the verification set are input, and the remaining 12 dimensions are obtained. The body size parameters of the two examples are listed in Table 4, so as to realize the comparison between the semantic-driven model size and the original size data of the verification set. Finally, the distribution of Size errorr epresentation of various parts of the human body in 30 groups is shown in Figure 10, and the predicted values are statistically analyzed in Table 5.

Figure 8.

Comparison between the scanned and established three-dimensional (3D) human body (example 1: normal body shape).

Figure 9.

Comparison between the scanned and established three-dimensional (3D) human body (example 2: obese body shape).

Table 4.

Input and predicted dimension parameters

Example		Example 1			Example 2
Input value	Height(cm)	183			170
Input value	Weight(kg)	70			97
Value	Position	True value	Predicted value	Absolute error	True value	Predicted value	Absolute error
Predictive value (cm)	Neck-base girth	39.214	37.93	1.284	44.684	44.13	0.554
	Chest girth	88.352	89.73	1.378	105.118	103.60	1.518
	Hip girth	97.868	96.39	1.478	112.907	112.25	0.657
	Inside leg length	86.776	86.35	0.426	77.970	77.47	0.500
	Back shoulder width	40.609	39.67	0.939	42.586	42.24	0.346
	Waist girth	77.574	77.12	0.454	108.837	108.91	0.073
	Arm length	65.329	64.09	0.239	60.905	60.23	0.675
	Upper-arm girth	34.104	33.26	0.844	40.409	40.31	0.099
	Wrist girth	17.778	16.89	0.888	18.837	18.62	0.217
	Chest height	131.595	132.21	0.615	125.180	125.05	0.130
	Knee girth	39.211	39.59	0.379	42.519	42.58	0.061
	Thigh girth	60.207	59.25	0.957	69.209	69.57	0.361

Table 5.

Analysis of the prediction error of various parts of the body

Part name	Mean (cm)	Standard deviation (cm)	CV value (%)
Neck-base girth	37.679	2.457	6.52
Chest girth	88.860	4.542	5.11
Hip girth	98.141	3.989	4.06
Inside leg length	72.954	3.205	4.39
Back shoulder width	41.346	1.632	3.95
Waist girth	80.631	5.075	6.29
Arm length	61.513	1.716	2.79
Upper-arm girth	33.937	1.733	5.11
Wrist girth	17.749	0.668	3.77
Chest height	126.482	4.292	3.39
Knee girth	39.288	1.528	3.89
Thigh girth	60.710	3.269	5.39

Note: the error analysis method refers to Gu et al.²⁵

From the comparison in Figures 8 and 9, it can be found that the semantic human model and the real human model have a high similarity in the overall posture. For the Table 4 statistics data and the Figure 10 data, the accuracy rate of each part of the human body within ±1.5 cm reaches an average of 91%, which basically meets the anthropometric requirements and can represent a personalized 3D human model.^26,27

Figure 10.

Size error representation of various parts of the human body in 30 groups.

Limitations and further improvements

Since this study is based on the semantic human model generated by the European and American human body parametric models, it is still different from the real Asian human model. Therefore, the contours that can best reflect the shape features of the two human models are compared, and the Figure 11 shows the 3D human contour information extracted based on the contour detection algorithm.

Figure 11.

Comparison of the outer contours of the human body.

From the figure, it can be seen that although there are differences in the human standing posture, differences can also be seen in the chest and waist, indicating that even if the predicted circumference data has few errors, the shape of the human body is still different. Subsequent corrections to the human form will not be discussed in this article. In addition, this study will continue to collect human data of different body shapes and different age groups, increase the amount of data, and improve the scope of application of this research method. The important parameters affecting the fit of clothing are analyzed and discussed to improve the applicability of the virtual fitting of clothing.

Conclusion

Aiming at the problem that human parameters are difficult to obtain and that it is difficult to establish personalized human models with a small number of parameters, this paper proposes the research idea of predicting the remaining 12 parameters with the two parameters of height and weight, so as to realize the accurate modeling of multiple parameters. Using the correlation among human body size parameters, this study employs the concept of combining strong and weak correlation predictions, that is, the linear correlation and KNN algorithm, to predict the 12 human semantic body size parameters. Then, the mapping relationship between the semantic parameters and the 3D model mesh is established to realize the deformation of the 3D human model. The accuracy rate of each part of the human body within ±1.5 cm reached an average of 91%, the 3D human body reconstruction is well completed, and we can establish a 3D human model with highly restored user characteristics for the garment industry through a small number of parameters, which has a certain practical value.

Although we made some progress in the prediction of human body size parameters and the construction of the semantic-driven 3D human model, the method has room for improvement. On the one hand, the amount of data collected in this study is limited, and the next step will continue to collect human body data of different ages and different human shapes, and classify human body shapes to improve the accuracy of semantic parameter prediction. Furthermore, a modification part of the 3D human body model could be added to refine the 3D human body model and make it more reducible to the user's own characteristics to further improve the accuracy and applicability of the 3D human modeling system based on semantic driving.

Footnotes

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the 2021 Industrial Technology Basic Public Service Platform Project of the Ministry of Industry and Information Technology of the People's Republic of China – Building a Public Service Platform for Testing and Verification of Intelligent Manufacturing Standards “Building a Public Service Platform for Testing and Verification of Intelligent Manufacturing Standards Around the Textile Industry” (2021-0173-2-1) and the National Key Research & Development Program of China, “Key common technologies of robot automation production line for sports shoes and clothing industry” (project no. 2018YFB1308801).

ORCID iDs

Wenqian Feng

Xin Rong Li

Xingxing Li

Yuzhuo Li

Hansen Li

Jingqiu Fang

References

Zhou

Long

, et al. Computational design of shape-changing robotic mannequin based on 3D human models. Text Res J 2021; 91: 1065–1082.

Wang

Tian

BQ.

Research progress of personalized 3D human modeling in virtual fitting. Text Herald 2020; 39(07): 79–82.

Chi

Zeng

Bruniaux

, et al. A new parametric 3D human body modeling approach by using key position labeling and body parts segmentation. Text Res J 2022; 92: 3653–3679.

, et al. Design of a multi-sensor information acquisition system for mannequin reconstruction and human body size measurement under clothes. Text Res J 2022; 92: 3750–3765.

Wang

HB.

Finite element modelling of Chinese male office workers' necks using 3D body measurements. J Text I 2017; 108: 766–775.

Liu

Huang

, et al. Template deformation-based 3D reconstruction of full human body scans from low-cost depth cameras. IEEE T Cybernetics 2017; 47: 695–708.

Yao

Peng

Guo

, et al. A data-driven approach for 3D human body pose reconstruction from a kinect sensor. J Phys Conf Ser 2018; 1098(1): 354–371.

Huang

Kwok

Zhou

Parametric design for human body modeling by wireframe-assisted deep learning. Comput Aid Des 2019; 108: 19–29.

, et al. Measuring and modeling human bodies with a novel relocatable mechatronic sensor-net. Text Res J 2019; 89: 4131–4147.

10.

Xie

Zhong

, et al. Non-parametric anthropometric graph convolutional network for virtual mannequin reconstruction. IEEE Acc 2020; 8: 3539–3550.

11.

GB/T 30548:2014. Verification method of anthropometric data in clothing field-obtained by using 3D measurement instrument.

12.

Pei

Park

Ashdown

SP.

Female breast shape categorization based on analysis of CAESAR 3D body scan data. Text Res J 2019; 89: 590–611.

13.

Pishchulin

Wuhrer

Helten

, et al. Building statistical shape spaces for 3D human modeling. Pattern Recogn 2017; 67: 276–286.

14.

Wang

Zhang

, et al. PanoMan: sparse localized components–based model for full human motions. TOG 2021; 42: 19–36.

15.

Allen

Curless

Popovic

The space of human body shapes: reconstruction and parameterization from range scans. TOG 2003; 22: 587–594.

16.

GB/T 16160:2017. Anthropometric definitions and methods for garment.

17.

Xie

Zhong

YQ.

Structure-consistent customized virtual mannequin reconstruction from 3D scans based on optimization. Text Res J 2020; 90: 937–950.

18.

Zhang

Liu

XL.

Regression analysis of women's leg size. J Silk China, 2017; 54: 31–35.

19.

Gui

Research on key technology of human body model matching based on 3D measurement. Dissertation, Dalian Polytechnic University, 2020.

20.

Zhu

Wang

Fuzzy monotonic K-nearest neighbor versus monotonic fuzzy K-nearest neighbor. IEEE Trans Fuzzy Syst 2022; 30: 3501–3513.

21.

Cheng

Chen

Martin

, et al. Parametric modeling of 3D human body shape-a survey. Comput Graph 2017; 71: 88–100.

22.

Wang

, et al. Shape-from-mask: a deep learning based human body shape reconstruction from binary mask images. Graph 2018; 1: 1–11.

23.

Zhou

Mao

Zhang

, et al. Selection of breast features for young women in northwestern China based on the random forest algorithm. Text Res J 2022; 92: 957–973.

24.

Zhu

Duan

Animal fiber imagery classification using a combination of random forest and deep learning methods. J Eng Fibers Fabr 2021; 16: 1–8.

25.

Zhong

ZJ.

Girth fitting of young women based on body digital images. J Silk China 2019; 56: 46–51.

26.

Wang

QL.

Parametric human body modeling and its rapid body measurement technology application. Shanghai: Shanghai University of Engineering Science, 2021.

27.

Kuang

Jiang

, et al. Young male's girth fitting of remote measurement based on HSI color model. Knit Ind 2023; 51(07): 70–75.