A classification algorithm for point cloud based on the dual-branch fusion and structure sampling
Aims:This paper aims to improve the accuracy in the point cloud classification task by using the dual-branch fusion and structure sampling to extract adequate local features and retain global structure information at different scales.Methods:Firstly,a dual-branch fusion module was proposed to extract the local features from different angles.Then the fusion framework was added to eliminate the repeated features,highlight the salient features,and make the local features more discriminating.Secondly,a structure sampling unit was designed to focus on the global structure information at different scales to learn the global structure information of the point cloud by learning from each sampling point to enhance the sampling stability and smooth the structure information.Finally,the local-global features were used to classify and complete the classification task of the point cloud.Results:The classification accuracy was 93.6% and 86.6% in ModelNet40 and ScanObjectNN benchmark datasets,respectively.Conclusions:The network model based on the dual-branch fusion and structural sampling is advanced in the point cloud classification task.
deep learningpoint cloud classificationdual-branch fusionstructure sampling
国家科技期刊平台
NSTL
万方数据
