A Global-and-Local Feature Fusion Network for Point Cloud Classification

Abstract

This paper presents a global-and-local convolutional neural network architecture (GLCNN) for point cloud classification tasks. Its core lies in the construction of a manifold learning module and a global-and-local convolution structure (GLConv). The former projects the point cloud into the two-dimensional planes in different views to learn sufficient low-dimensional information. Then the global feature learning is obtained by constructing attention mechanisms from Transformer and local feature learning is captured by convolution and pooling operators, thereby forming the GLConv by fusing these two features to further explore the long-distance dependency of point clouds. Experiments show that GLCNN has excellent performance on two public datasets. Particularly, the overall accuracy and class average accuracy of the ModelNet40 dataset reached 93.3% and 90.8%, respectively.