摘要
为增强捕捉细粒度局部特征能力以进一步提高复杂场景点云语义分割精度,将自注意力机制引入PointNet++构建点云语义分割网络SSA-PointNet++.首先将采样点邻域的自注意力明确分为中心自注意力和邻域自注意力两部分,综合两者并结合不同空间编码方式增强网络模型对采样点邻域拓扑结构的学习;然后构建注意力池化模块以强化重要信息在网络的有效传递,并通过差异性池化函数整合注意力池化、最大池化提取的多个全局特征以提高点云语义分割结果的鲁棒性.对公开数据集S3DIS,Semantic3D的场景语义分割实验表明,所提网络模型数据集分割精度mIoU较基准模型提升效果显著,在室内数据集S3DIS上的mIoU较PointNet++提升达6.6%,在室外数据集Semantic3D上的mIoU高出MSDeepVoxNet约3%;与公开数据集上其他网络模型的分割结果相比,所提模型性能均有不同程度的提升,具有更强的泛化性能和良好的应用价值.
To improve the accuracy of semantic segmentation of point cloud in complex scenes,a novel convolutional neural network(CNN)called SSA-PointNet++by imposing the self-attention mechanism onto traditional PointNet++network is proposed.Firstly,the self attention of the neighborhood of the sampling point is divided into two parts:the central self-attention and the neighborhood-attention.Two kinds of self-attention mechanism are then combined to improve the network’s ability in capturing fine-grained local features.Secondly,one attention pooling module is constructed based on adaptive selection of features from the attention mechanism for the effective transmission of important information in the network.The global features extracted from the attention pooling and the maximum pooling is effectively fused to improve the robustness of the point cloud semantic segmentation results.Experiments on public data sets S3 DIS and Semantic3 D show that the proposed network outperforms are improved compared with the benchmark model in terms of both the overall accuracy and mIoU.For indoor dataset S3 DIS,the mIoU of the proposed SSA-PointNet++is good and 6.6%higher than PointNet++.For outdoor dataset Semantic3 D,the mIoU of the proposed SSA-PointNet++is about 3%higher than MSDeepVoxNet.Compared with the segmentation results of other networks on the public datasets,proposed network is more general on different datasets and high application potential.
作者
吴军
崔玥
赵雪梅
陈睿星
徐刚
Wu Jun;Cui Yue;Zhao Xuemei;Chen Ruixing;Xu Gang(School of Electronic Engineering and Automation,Guilin University of Electronic Technology,Guilin 541000;Ningbo Institute of Materials Technology and Engineering,Chinese Academy of Sciences,Ningbo 315000)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2022年第3期437-448,共12页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(41761087,41801233)
浙江省宁波市“科技创新2025”重大专项(2020Z019,2020Z013)
广西自然科学基金(2020GXNSFBA159012)
广西自动检测技术与仪器重点实验室项目(YQ20104)。
关键词
点云语义分割
深度学习
卷积神经网络
自注意力机制
point cloud semantic segmentation
deep learning
convolutional neural network
self-attention mechanism
