摘要
局部几何形状的描述能力,对不规则的点云形状表示是十分重要的.然而,现有的网络仍然很难有效地捕捉准确的局部形状信息.在点云中模拟深度可分离卷积计算方式,提出一种新型的动态覆盖卷积(dynamic cover convolution, DC-Conv),以聚合局部特征. DC-Conv的核心是空间覆盖算子(space cover operator, SCOP),该算子通过在局部区域中构建各向异性的空间几何体覆盖局部特征空间,以加强局部特征的紧凑性.DC-Conv通过在局部邻域中动态组合多个SCOP,实现局部形状的捕捉.其中,SCOP的注意力系数通过数据驱动的方式由点位置自适应地学习得到.在3D点云形状识别基准数据集ModelNet40, ModelNet10和ScanObjectNN上的实验结果表明,该方法能有效提高3D点云形状识别的性能和对稀疏点云的鲁棒性.最后,也提供了充分的消融实验验证该方法的有效性.开源代码发布在https://github.com/changshuowang/DC-CNN.
The ability to describe local geometric shapes is very important for the representation of irregular point cloud.However,the existing network is still difficult to effectively capture accurate local shape information.This study simulates depthwise separable convolution calculation method in the point cloud and proposes a new type of convolution,namely dynamic cover convolution(DC-Conv),to aggregate local features.The core of DC-Conv is the space cover operator(SCOP),which constructs anisotropic spatial geometry in a local area to cover the local feature space to enhance the compactness of local features.DC-Conv achieves the capture of local shapes by dynamically combining multiple SCOPs in the local neighborhood.Among them,the attention coefficients of the SCOPs are adaptively learned from the point position in a data-driven way.Experiments on the 3D point cloud shape recognition benchmark dataset ModelNet40,ModelNet10,and ScanObjectNN show that this method can effectively improve the performance of 3D point cloud shape recognition and robustness to sparse point clouds even in the case of a single scale.Finally,sufficient ablation experiments are also provided to verify the effectiveness of the method.The open-source code is published at https://github.com/changshuowang/DC-CNN.
作者
王昌硕
王含
宁欣
田生伟
李卫军
WANG Chang-Shuo;WANG Han;NING Xin;TIAN Sheng-Wei;LI Wei-Jun(Institute of Semiconductors,Chinese Academy of Science,Beijing 100083,China;Center of Materials Science and Optoelectronics Engineering&School of Microelectronics,University of Chinese Academy of Sciences,Beijing 100049,China;Cognitive Computing Technology Joint Laboratory,Wave Group,Beijing 100083,China;Nanfang College Guangzhou,Guagnzhou 510970,China;Zhuhai Fudan Innovation Institute,Zhuhai 518057,China;School of Software,Urumqi 830091,China;Beijing Key Laboratory of Semiconductor Neural Network Intelligent Sensing and Computing Technology,Beijing 100083,China)
出处
《软件学报》
EI
CSCD
北大核心
2023年第4期1962-1976,共15页
Journal of Software
基金
国家自然科学基金(61901436)
广东省重点领域研发计划(2019B010107001)
中国科学院B类先导科技专项培育项目(XDPB22)。
关键词
点云分类
动态覆盖卷积
空间覆盖算子
局部邻域
注意力系数
point cloud classification
dynamic cover convolution
space cover operator
local neighborhood
attention coefficient
