基于少量数据集的三维点云生成模型  

3D Point Cloud Generation Model Based on a Small Number of Datasets

作　　者：张渊杰 沈洋 许浩[2] 包艳霞 应震[2] ZHANG Yuanjie;SHEN Yang;XU Hao;BAO Yanxia;YING Zhen(School of Information Science and Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China;School of Mathematics and Computer,Lishui University,Lishui 323000,China;Zhejiang Zhangxin Media Technology Co.,Ltd.,Lishui 323000,China;Zhejiang Juxin Automation Equipment Co.,Ltd.,Lishui 323000,China)

机构地区：[1]浙江理工大学信息科学与工程学院,浙江杭州310018 [2]丽水学院数学与计算机学院,浙江丽水323000 [3]浙江掌信传媒科技有限公司,浙江丽水323000 [4]浙江聚新自动化设备有限公司,浙江丽水323000

出　　处：《软件工程》2024年第11期69-74,共6页Software Engineering

基　　金：浙江省自然科学基金项目(LY21F02004);丽水市公益性技术应用研究计划项目(2022GYX12)。

摘　　要：针对生成对抗网络(GAN)需要大量训练数据及点云数据稀缺且获取难度大的问题,提出一种基于少量数据集的三维点云生成模型。该模型首先通过重采样和水平旋转的方法实现数据增强,使第一级网络能够生成具有多样性的低分辨率点云;其次通过确保低分辨率点云与高分辨率点云之间的对应关系,实现点云的超分辨率生成;最后实现生成具有多样性的高分辨率点云。实验结果表明,在ShapeNet Part(ShapeNet Part Segmentation Dataset)数据集上,该模型的JS散度相较于Tree-GAN的JS散度下降了0.416,证明其性能优于Tree-GAN。Aiming at the challenges that Generative Adversarial Network(GAN)requires large training datasets and the scarcity and difficulty of acquiring point cloud data,this paper proposes a 3D point cloud generation model based on a small number of datasets.This model first employs resampling and horizontal rotation techniques for data augmentation,enabling the first-level network to generate diverse low-resolution point clouds.Secondly,by ensuring the correspondence between low-resolution and high-resolution point clouds,the model achieves super-resolution generation of point clouds.Finally,it generates diverse high-resolution point clouds.The experimental results show that on the ShapeNet Part Segmentation Dataset,the JS divergence of the proposed model decreases by 0.416 compared to that of the Tree-GAN,demonstrating its superior performance.

关 键 词：GAN 少量点云数据 重采样 数据增强 超分辨率 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]