高级检索
熊伟, 娄政浩, 徐敏夫, 袁和金. 集多头点注意力与边卷积的点云分类分割模型[J]. 计算机辅助设计与图形学学报. DOI: 10.3724/SP.J.1089.2023-00126
引用本文: 熊伟, 娄政浩, 徐敏夫, 袁和金. 集多头点注意力与边卷积的点云分类分割模型[J]. 计算机辅助设计与图形学学报. DOI: 10.3724/SP.J.1089.2023-00126
Wei XIONG, ZhengHao LOU, MinFu XU, HeJin Yuan. Edge Convolution with Multi-Head Point Attention for Point Cloud Classification and Segmentation[J]. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/SP.J.1089.2023-00126
Citation: Wei XIONG, ZhengHao LOU, MinFu XU, HeJin Yuan. Edge Convolution with Multi-Head Point Attention for Point Cloud Classification and Segmentation[J]. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/SP.J.1089.2023-00126

集多头点注意力与边卷积的点云分类分割模型

Edge Convolution with Multi-Head Point Attention for Point Cloud Classification and Segmentation

  • 摘要: 针对DGCNN只在局部尺度上独立提取点特征, 未将局部点互相关联的问题, 提出了一种集多头点注意力与边卷积的点云分类分割模型MHPAEC. 首先, 设计单头点注意力模块分别计算点云的注意力特征与邻域注意力特征, 学习点云的旋转不变性, 使用多头机制将单头点注意力模块进行聚合, 构建多头点注意力模块, 赋予邻域内不同点相应的注意力系数; 其次, 设计加权金字塔池化模块进行特征融合, 获得更加丰富的特征信息. 最后, 针对点云分类分割任务中存在的难分类样本和类别不平衡问题, 提出结合交叉熵损失和焦点损失的联合损失函数, 使样本得到充分训练. 在ModelNet40数据集和ShapeNet数据集上分别进行了点云分类与分割实验, 在ModelNet40数据集上, MHPAEC网络的总体精度达到了94.1%; 在ShapeNet数据集上, MHPAEC网络的平均交并比达到了86.3%, 有效地提升了网络模型的分类分割性能.

     

    Abstract: A point cloud classification and segmentation model MHPAEC, which combines multi-head point attention and edge convolution, is proposed to address the problem of DGCNN only extracting point features independently at local scales without correlating local points. Firstly, design a single-head point attention module to calculate the attention features of the point cloud and the neighborhood attention features, learn the rotation invariance of the point cloud, use a multi-head mechanism to aggregate the single-head point attention modules, construct a multi-head point attention module, and assign corresponding attention coefficients to different points in the neighborhood; Secondly, design a weighted pyramid pooling module for feature fusion to obtain richer feature information; Finally, aiming at the problem of difficult to classify samples and category imbalance in the task of point cloud classification and segmentation, a joint loss function combining cross entropy loss and focal loss is proposed to make samples fully trained point cloud classification and segmentation experiments were conducted on the ModelNet40 dataset and ShapeNet dataset, respectively. On the ModelNet40 dataset, the overall accuracy of the MHPAEC network reached 94.1%; On the ShapeNet dataset, the mean intersection over union of the MHPAEC network reached 86.3%, effectively improving the classification and segmentation performance of the network model.

     

/

返回文章
返回