Predicting ready-made garment dressing fit for individuals based on highly reliable examples

Highlights

• Generating highly reliable body-garment paired data by digitalizing the garment on a shape-changing robotic mannequin with the method of physically simulating the twin virtual garment with the guidance of vision.

• The paired data of a template garment are generalized to be suitable to garments with similar style by scaling the template garment-body ease-allowance, i.e., the extra measurement added to the body measurement, for different sized human bodies.

• Deformation characters of a garment template are transformed to daily garments to predict their dressing effects on different persons by training a CNN-based network, which can obtain plausible results.

Abstract

Predicting the dressing fit of a ready-made garment for different individuals is a challenging problem in online garment sales. At present, the main solution involves physics-based virtual garment simulation, which has limitations in terms of a lack of sufficient reality and high computational costs. In this study, we developed a novel example-based method to guarantee the high reality and efficiency of garment dressing fit prediction using two methods. First, highly reliable examples were captured with the assistance of a robotic mannequin, thereby ensuring the authenticity of the sample data. Second, a mapping was established between the body-garment ease allowance and garment deformations using a convolutional neural network-based network in order to address the problem of dressing fit prediction for ready-made garments on different individuals. This method can also be extended to predicting the dressing fit for ready-made garments with similar styles. Experiments showed our virtual clothes try-on system obtained acceptable precision with a good sense of reality, and it can potentially be used in many applications such as three-dimensional garment design and online clothes retail.

Introduction

In general, garments are classified as ready-made or custom-made, but ready-made garments still dominate the clothing market. Before buying a ready-made garment, people usually need to try it on to check whether it fits well. Due to the rapid development of clothing E-commerce, online sales now account for almost two-thirds of ready-made garments sales [1]. However, due to the lack of effective methods for helping customers to perceive the dressing fit on themselves, the return rate is high at up to 25% of online sales. Therefore, predicting the dressing fit of ready-made garments for individuals is an important but challenging problem in online garments sales.

One possible solution to this problem is physics-based virtual garment simulation [2], [3], which simulates the deformation of three-dimensional (3D) virtual garments on 3D avatars of customers. Great progress has been made since the first clothing simulation method was proposed in the late 1980s, but clothing simulation is still insufficient realistic [4], [5]. This is mainly because the physical models of clothing used in simulation, such as the mass-spring model[6], position-based model[7], or finite element methods [8], [9], are simplified from the real clothing materials, thereby leading to differences in the micro- and macro- mechanical characteristics of the virtual clothing and real clothing. Thus, the virtual simulation results usually deviate much from the real clothing deformation and they cannot reflect the real dressing fit well.

To improve the realism of clothing simulations, many studies have aimed to reconstruct the dynamic deformation of garments based on computer vision methods [10], [11]. These studies mainly focused on reconstructing the deformation of a garment on a moving human body. Using these data, the deformation of a garment caused by human body posture changes can be modeled in a data-driven manner [12], [13].

Theoretically, it is possible to extend previous methods to reconstruct the dressing fit of a garment on various human bodies, but it is impractical to ask a huge number of people to try on a garment and scan them because of the high costs incurred.

Thus, in the present study, we used a shape-changing robotic mannequin [14] to imitate different shapes of various human bodies. Consequently, the dressing fit of a garment on various human bodies could be simulated and reconstructed by deforming the garment with the robotic mannequin. The mapping between body-garment ease allowance and deformations of a garment were trained using the reconstructed data. Therefore, even without trying on a garment physically, its dressing fit on a specific person can still be estimated. However, it is still uneconomic to reconstruct every garment on sale. To address this issue, we developed a new strategy for estimating the dressing fit of a class of garments based on the data captured from a template garment using the methods proposed in this study, as shown in Fig. 1. The main contributions of this study summarized as follows:

• A highly reliable body-garment paired data acquisition approach is proposed. By dressing up a shape-changing robotic mannequin that simulates different shapes of various human bodies, the garment deformations are reconstructed based on physical simulations, where constraints on marker positions are computed by stereo vision.

• Instead of mapping directly between the garment deformations and human body shapes, we obtained the mapping indirectly by training the relationship between the garment deformations and body-garment ease allowance, thereby allowing good training results to be trained with fewer data. The map is set in a nonlinear manner rather than the linear manner used in most previously proposed approaches in order to fit better with the intrinsic nonlinear relationship between body shape and garment deformation.

The remainder of this paper is organized as follows. In Section 2, we review previous research into virtual try-on methods, 3D human modeling methods, and robotic mannequin. In Section 3, we explain the pipeline used in our study. After describing the position-constrained physical deformation of the template garment in Section 4, the method for mapping from body-garment ease allowance to garment states is explained in Section 5. The experimental results are presented in Section 6. We give our conclusions in Section 7.