基于图神经网络的三维服装点云原型表征  

Prototype Representation of 3D Clothing Point Cloud Based on Graph Neural Network

作　　者：周呈林 袁晔 ZHOU Chenglin;YUAN Ye(School of Arts and Sciences,Beijing Institute of Fashion Technology,Beijing 100029,China)

机构地区：[1]北京服装学院文理学院,北京100029

出　　处：《北京服装学院学报（自然科学版）》2023年第4期60-66,共7页Journal of Beijing Institute of Fashion Technology：Natural Science Edition

基　　金：北京服装学院2022年研究生教育质量提升专项及一般教学改革项目(项目编号:120301990132)。

摘　　要：数据驱动的三维数字化服装是数字化服装的前沿研究领域,具有重要的理论意义和工程价值。本文将图神经网络引入三维服装点云原型的表征,首先在服装人体特征识别的基础上,通过对人体特征点云添加服装放松量得到图神经网络的测试样本;然后将选取的T恤服装样本按照不同部位分割为若干曲面模型,对分片T恤曲面模型下采样形成训练图神经网络的子点云数据,通过图神经网络对所有训练样本的特征训练产生服装点云原型表征模型;最后应用人体特征点云上采样结果与参考点云之间的距离值分析图神经网络模型与同类结构的服装原型表征的效果。相较于其他模型构建的服装原型点云而言,本方法具有更好的三维服装点云原型表征效果。Data-driven 3D digital clothing is a frontier research field of digital clothing with important theoretical significance and engineering value.In this paper,graph neural network was introduced into the representation of 3D clothing point cloud prototype.On the basis of human body feature recognition,the test sample of the graph neural network was obtained by adding ease to the human body feature point cloud.Then,the selected T-shirt clothing samples were divided into several surface models according to different parts,and the training samples of the graph neural network were sub-point cloud data obtained by using down-sampling operations on these surface models.The prototype representation model of clothing point cloud was generated by the feature training of all training samples through the graph neutral network.Finally,the distance value between the upsampling result and the reference point cloud was used to analyze the representation effect of the graph neural network model and the clothing prototype of the same structure.Compared with the point cloud of garment prototype constructed by other models,this method has better representation effect of three-dimensiong point cloud prototype.

关 键 词：点云 服装放松量 图神经网络 点云上采样 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]