Dual-focusing network: enhancing aerial point cloud semantic segmentation with adaptive neighborhood refinement

Point cloud data, particularly aerial point clouds, presents challenges due to its unstructured nature and low density. This paper introduces DFNet, a dual-focusing network designed to enhance aerial point cloud semantic segmentation. DFNet incorporates a shape-focus neighborhood refinement module to adaptively determine optimal neighborhood shapes and a local context focus module to prioritize discriminative local context features. Experimental results on the STPLS3D and ISPRS Vaihingen 3D datasets demonstrate state-of-the-art performance, where DFNet achieves outstanding mean intersection over union scores of 54.38% and 56.65%, respectively, to surpass existing methods. Our approach offers a robust solution for complex urban environment segmentation. The source code will be available at https://github.com/xianyu-wang/DFN.