面向移动设备的三维人体重建方法  

Method of 3D Human Body Reconstruction for Mobile Device

作　　者：寿旭峰 卢书芳[1] 高飞[1] SHOU Xu-feng;LU Shu-fang;GAO Fei(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023

出　　处：《小型微型计算机系统》2022年第7期1452-1457,共6页Journal of Chinese Computer Systems

基　　金：浙江省自然科学基金面上项目(LY19F020027)资助.

摘　　要：针对移动端的计算资源和存储空间有限等问题,本文提出了一种轻量级的三维人体重建方法.首先,采用DeepLabV3+网络对人体正面和侧面图像进行分割,获得人体净身轮廓.其次利用SMPL人体模型对三维人体进行参数化表示,并对SMPL模型的正面和侧面进行投影,获得二值轮廓作为数据集.然后,构建并训练一个教师网络以预测二值轮廓图的SMPL参数.之后,构建一个轻量级的学生网络,通过知识蒸馏的方式,利用教师网络来提高学生网络预测的人体参数精度.最后通过学生网络预测的SMPL人体参数生成三维人体模型.实验结果证明,本文的方法可以在消耗较少的计算资源和存储空间的情况下获得相对较高精度的三维人体.Due to the limitation of computing resource and storage space in mobile device,we propose a lightweight method to reconstruct 3D human body.Firstly,we use DeepLabV3+to segment front-view and side-view image to get under-the-clothes profile.Then,we utilize SMPL model to parameterize the 3D human model.After it,the front-view and side-view of SMPL model are projected to a plane to generate binary silhouettes as a dataset.Next,we design a teacher network and a lightweight student network,we train teacher network to predict SMPL parameters of binary silhouettes.Then,we use teacher network to improve the accuracy of SMPL parameters predicted by student network through knowledge distillation.Finally,we reconstruct 3D human body by SMPL parameters.The experimental results illustrate our method could reconstruct relatively high quality 3D human body with less computing resource and storage space.

关 键 词：三维重建 知识蒸馏 轻量化神经网络 净身轮廓 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]