基于残差多感知机与空间注意力的点云分割算法
Point Cloud Segmentation Algorithm Based on Residual Multi-perceptron and Spatial Attention
钦耀 1韩永国 1陈永辉 1王赋攀1
作者信息
- 1. 西南科技大学计算机科学与技术学院 绵阳 621010
- 折叠
摘要
现有的三维点云分割方法一般使用多层感知机作为点云的特征提取器,从而实现点云的分割.但是该特征提取器未能够考虑到点云中点之间的关系,导致提取点云特征的能力不强.为了充分学习点之间的联系,提高点云分割的精度,论文提出了一种融合残差多感知机与空间注意力的神经网络,实现三维点云的分割效果,并将该神经网络取名为Re-sPoint++.ResPoint++网络通过多个含有残差多感知机模型的特征提取模块来提取局部点云的几何与结构特征,并在此基础上引入三维空间注意力机制来学习局部点间的联系,优化网络训练,最终输出的结果是每个点在数据集中的所属类别.实验结果表明,采用ResPoint++的点云分割网络相比PointNet与PointNet++等网络具有更高的分割精度,验证了该网络具有良好的点云分割效能.
Abstract
Existing 3D point cloud segmentation methods generally use multi-layer perceptrons as point cloud feature extrac-tors to achieve point cloud segmentation.However,the feature extractor fails to take into account the relationship between points in the point cloud,resulting in a weak ability to extract point cloud features.In order to fully learn the relationship between points and improve the accuracy of point cloud segmentation,this paper proposes a neural network that combines residual multi-perceptron and spatial attention to achieve the segmentation effect of 3D point cloud named ResPoint++.The ResPoint++network extracts the geometric and structural features of the local point cloud through multiple feature extraction modules containing residual multi-per-ceptron models,and on this basis,introduces a three-dimensional spatial attention mechanism to learn the relationship between lo-cal points and optimize network training.The final output is the category of each point in the dataset.The experimental results show that the point cloud segmentation network using ResPoint++has higher segmentation accuracy than PointNet and PointNet++,which verifies that the network has good point cloud segmentation performance.
关键词
点云/空间注意力/分割/残差多感知机/深度学习Key words
point cloud/spatial attention/segmentation/residual MLP/deep learning引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据