| 摘 要: 针对点云数据高维、稀疏、非结构化的特性及现有处理方法面临的规模与结构复杂度挑战,提出一种稀疏结构感知的无监督三维点云表示学习方法。该方法融合几何形状与特征点显著信息学习高语义结构点,引入对
比学习与三元正则化机制优化特征学习与结构点生成,在三维点云形状分类、语义分割任务中验证了方法的高效性与稳定性。 |
| 关键词: 深度学习 三维点云 稀疏结构表示 对比学习 结构点学习 |
|
中图分类号: TP391
文献标识码: A
|
|
| Unsupervised Structural Point Learning for 3D Point Clouds Driven by Contrastive Learning |
|
ZHANG Xiangyi, DOU Feng, SUN Xinglu, WANG Hao
|
(Dalian Neusoft University of Information, Dalian 116023, China)
zhangxiangyi@neusoft.edu.cn; doufeng@neusoft.edu.cn; sunxinglu24@dnui.edu.cn; c2945457578@163.com
|
| Abstract: Aiming at the high-dimensional, sparse and unstructured characteristics of point cloud data, as well as the challenges posed by the eve-r increasing data scale and structural complexity to existing processing methods, this paper proposes an unsupervised sparse structure-aware representation learning method for 3D point clouds. The proposed method fuses geometric shape features with the saliency information of feature points to extract structural points with rich semantic information, and further introduces the contrastive learning mechanism and ternary regularization strategy to optimize the feature learning process and the quality of structural point generation.Experimental validations on the tasks of 3D point cloud shape classification and semantic segmentation demonstrate the high efficiency and good stability of the proposed method. |
| Keywords: deep learning 3D point cloud sparse structure representation contrastive learning structure point learning |
