Local-non-local complementary learning network for 3D point cloud analysis

Point cloud analysis is integral to numerous applications, including mapping and autonomous driving. However, the unstructured and disordered nature of point clouds presents significant challenges for feature extraction. While both local and non-local features are essential for effective 3D point cloud analysis, existing methods often fail to seamlessly integrate these complementary features. To address this limitation, we propose the Local-Non-Local Complementary Learning Network (LNLCL-Net), a novel framework that enhances feature extraction and representation. Leveraging partial convolution, LNLCL-Net divides the feature map into distinct local and non-local components. Local features are modeled through relative positional relationships, while non-local features capture absolute positional information. A Complementary Interactive Attention module is introduced to enable adaptive integration of these features, enriching their complementary relationship. Extensive experiments on benchmark datasets, including ModelNet40, ScanObjectNN, and ShapeNet Part, demonstrate the superiority of our approach in both quantitative and qualitative metrics, achieving state-of-the-art performance in classification and segmentation tasks.