A Local-Global Feature Fusing Method for Point Clouds Semantic Segmentation

In recent years, the abundance of information in 3D data has made the semantic segmentation of 3D point clouds a topic of great interest. However, current methods often rely solely on the original three-dimensional coordinates of the point cloud as input geometric features, leading to poor generalization performance. Additionally, occlusion of the point cloud data can negatively impact segmentation accuracy when only local information is considered. To address these issues, this paper proposes a network named LGFF-Net. To fully utilize the original information of point clouds, we designed a Local Feature Aggregation (LFA) module that treats geometric and semantic information equally and preserves the original properties while cross-augmenting them. On the other hand, we proposed a simple and effective Global Feature Extraction (GFE) module to extract global features. Finally, we hierarchically fuse local and global features using a U-shaped segmentation structure. Compared to state-of-the-art networks, our method achieves competitive results on several benchmark datasets, including Semantic Topographic Point Labeling-Synthetic 3D, Toronto_3D, Stanford Large 3-D Indoor Space, and ScanNet. We also conduct multiple ablation experiments to validate the efficacy of LGFF-Net.