摘要
针对现有虚拟试衣软件不能实时仿真的问题,提出一种基于机器学习的静态衣袖虚拟仿真方法。首先根据袖窿、袖山高和袖肥的变化规律,均匀划分出20个二维衣袖样板,组合6类面料,在CLO3D中制作120个与实物面料悬垂指标一致的三维仿真衣袖样本;提出三角网格规则四边形化的转化方法,通过UV传递将衣袖模型的不规则三角网格转化为规则四边形网格;最后选取随机森林RF和极端梯度提升XGBoost 2种机器学习算法进行拟合并优化。结果表明:XGBoost算法模型的顶点坐标预测值的MAE值为1.81 mm,比RF模型降低了0.51 mm,MAPE值为11.03%,比RF模型降低了2.51%,预测一组数据平均耗时0.389 s,总体表现优于RF模型;当新衣袖样板数据输入时,XGBoost算法模型可快速输出对应的三维衣袖网格顶点坐标,该方法为研究静态虚拟服装的实时仿真提供了参考。
Aiming at the problem that the existing virtual fitting software cannot simulate in real time,a virtual simulation method for static sleeves based on machine learning was proposed.Firstly,202D sleeve samples were evenly divided according to the variation law of armholes,sleeve heights and sleeve fats.Six kinds of fabrics were combined to produce 1203D simulation sleeve samples consistent with the real fabric drape index in CLO3D.Then,a method of transforming triangular meshes into regular quadrilaterals was proposed,and the irregular triangular meshes of the sleeve model were transformed into regular quadrilateral meshes through UV transfer.Finally,two machine learning algorithms,random forest RF and extreme gradient boosting XGBoost,were selected to fit and optimize.The results show that the MAE value of the vertex coordinate prediction value of the XGBoost algorithm model is 1.81 mm,which is 0.51 mm lower than that of the RF model.The MAPE value is 11.03%,which is 2.51%lower than that of the RF model.It takes 0.389 s to predict a set of data on average.The overall performance is better than the RF model.When the new sleeve template data is input,the XGBoost algorithm model can quickly output the corresponding three-dimensional sleeve mesh vertex coordinates,which provides a reference for the study of real-time simulation of static virtual clothing.
作者
张惠
ZHANG Hui(Jiangxi Centre for Modern Apparel Engineering and Technology,Jiangxi Institute of Fashion Technology,Nanchang,Jiangxi 330201,China)
出处
《毛纺科技》
CAS
北大核心
2023年第2期89-97,共9页
Wool Textile Journal
基金
江西省教育厅科学技术研究项目(GJJ191086)。
