基于深度学习的非刚体三维重建  

Non-rigid 3D Reconstruction Based on Deep Learning

作　　者：胡情丰 HU Qing-feng(School of Information,Zhejiang University of Technology,Hangzhou 310018,China)

机构地区：[1]浙江理工大学信息学院,杭州310018

出　　处：《软件导刊》2021年第7期43-48,共6页Software Guide

摘　　要：非刚性运动结构的三维重建旨在从二维图像中提取出关键坐标,并恢复其对应的三维形状及姿态。然而,目前的传统方法无法在大规模场景中应用。为此,将卷积神经网络应用于非刚体运动的三维重建中,提出一个基于无监督学习的非刚体三维重建框架。在FacebookAI实验室提出的C3DPO基础上,选择iResNet为backbone,并采用Ranger优化器进行训练。实验结果表明,该框架不仅可以更快地收敛,而且在H36M、Pascal3D、S-Up3D数据集中的MPJPE(位置误差的绝对平均值)分别达到了92.2、37.1、0.067,均优于C3DPO。The three-dimensional reconstruction of non-rigid motion structure aims to extract the key coordinates from the two-dimensional images and restore the corresponding 3D shape and attitude.However,the current traditional approach is not available in large-scale scenarios.Therefore,convolutional neural network is applied to the 3D reconstruction of non-rigid motion,and a non-rigid 3D reconstruction framework based on unsupervised learning is proposed.Based on C3DPO proposed by FaceBook AI laboratory,iResNet is used as backbone and trained with Ranger optimizer.In the experiment,the framework used in this paper can not only converge faster,but also MPJPE(absolute mean per joint position error)in H36M,Pascal3D and S-Up3D respectively reaches 92.2,37.1 and 0.067,which are better than C3DPO.

关 键 词：非刚性运动估计 深度学习 三维重建 无监督学习 神经网络 优化器 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]