基于串行关注机制的3D点云分类算法

A 3D point cloud classification algorithm based on the serial attention mechanism

目的:优化3D点云的分类任务中邻域点的选取和聚合,平滑局部邻域内的结构信息,并使卷积神经网络关注较为有用的信息。方法:首先,提出多空间索引来优化点云中邻域点的选取。同时,通过关注机制学习得分权重聚合邻域点的特征。接着,设计多通道关注模块对特征进行细化,提高特征的融合力和可辨别性。最后,将细化的特征放入分类器中,完成点云的分类任务。结果:在ModelNet40和ScanObjectNN两个具有挑战性的基准数据集上分别达到了93.4%和81.7%的分类精度。结论:基于串行关注机制的卷积神经网络在点云分类任务中取得良好性能。

Aims:This paper aims to optimize the selection and aggregation of neighborhood points in the classification task of 3D point cloud,to smooth the structural information in local neighborhood,and to divert the convolutional neural network to the more useful information.Methods:Firstly,a multi spatial index was proposed to optimize the selection of neighborhood points in the point cloud.At the same time,the characteristics of neighborhood points were aggregated by learning the score weight through the attention mechanism.Then,a multichannel attention module was designed to refine the features to improve the fusion and distinguishability of the features.Finally,the refined features were put into the classifier to complete the classification task of points.Results:The classification accuracy was 93.4%and 81.7%respectively on the two challenging benchmark data sets of ModelNet40 and ScanObjectNN.Conclusions:The convolutional neural network based on the serial attention mechanism has achieved good performance in point cloud classification tasks.