| 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。 |
| 关键词: GAN;少量点云数据;重采样;数据增强;超分辨率 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 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) |
|
| 3D Point Cloud Generation Model Based on a Small Number of Dataset |
|
ZHANG Yuanjie1, SHEN Yang1,2, XU Hao2, BAO Yanxia2,3,4, YING Zhen2
|
(1.School o f In f ormation Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.School of Mathematics and Computer, Lishui University, Lishui 323000, China;
3.Zhejiang Zhangxin Media Technology Co., Ltd., Lishui 323000, China;
4.Zhejiang Juxin Automation Equipment Co., Ltd., Lishui 323000, China)
630512163@qq.com; tlsheny@126.com; oah_ux@126.com; 82240849@qq.com; 470712367@qq.com
|
| Abstract: 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. |
| Keywords: GAN; a small amount of point cloud data; resampling; data enhancement; super-resolution |
