Point Cloud Classification by Integrating Attention Mechanism with Sparse Convolutional

Point cloud data is one of the essential data sources for 3D reconstruction, and a major challenge in 3D reconstruction based on point clouds lies in the fine-grained classification of sparse point clouds. However, for large-scale objects such as buildings, the point sets can be extremely sparse, making it difficult to fully describe the morphological characteristics of the objects represented by the sampled point sets. In this paper, we propose a 3D point cloud classification network based on Minkowski sparse convolution, and introduce a sparse convolution module that integrates an attention mechanism. The aim is to fully represent the sparsely expressed features and enhance their expressive ability. To validate the effectiveness of our method, experiments were conducted on the ISPRS and 2019DFC datasets. The experimental results show that, compared to mainstream methods, our approach achieves higher point cloud classification accuracy and demonstrates good generalization ability.