面向局部学习的点云分割分类网络

A point cloud segmentation classification network for local learning

在各种3D对象表示中,三维点云越来越受欢迎,其中基于点的方法在各种数据集上都表现出了良好的性能。针对PointNet++只关注了点本身的信息,并未关注相邻点的信息,同时它采用最大池化聚合局部信息,导致丢失次最大值信息。由此,文中提出Con-PointNet++网络,该网络利用增强局部信息模块,以关注相邻点的信息,从而增强局部信息特征提取;采用局部注意力机制下的融合池化模块,将最大池化与注意力池化特征信息融合,得到更为丰富的局部特征信息。本文方法在室内数据集S3DIS的Area_5区域上评估模型语义分割能力,mIoU达55.2%;在数据集ModelNet40上评估模型分类效果,OA达91.2%。与其他方法相比,所提模型性能均有提升,进一步证明了本文方法的有效性。

3D point clouds are becoming increasingly popular in various 3D object representations,where point-based methods have shown good performance on a variety of datasets.In response to PointNet++only focuses on the information of the points themselves and not the information of the neighboring points,while it uses max-pooling to aggregate local information,resulting in the loss of sub-maximal information.As a result,the Con-PointNet++network is proposed to make full use of the enhanced Local Information Module to focus on the information of neighboring points and thus enhance the local information feature extraction.Then,the fusion pooling module under the local attention mechanism is used to fuse the max-pooling and attention pooling feature information to obtain richer local feature information.The proposed method evaluates the model semantic segmentation ability on Area_5 of indoor dataset S3DIS,with mIoU of 55.2%;and the model classification effect on dataset ModelNet40,with OA of 91.2%.Compared with other methods,the performance of the proposed model is improved,which further demonstrates the effectiveness of the proposed method.