| Title |
Exploiting Topological Characteristics in 3D Point Cloud Semantic Segmentation |
| Authors |
레득틴(Thinh D. Le) ; 유명식(Myungsik Yoo) |
| Keywords |
3D LiDAR data; Topological data analysis; Semantic segmentation |
| Abstract |
Semantic segmentation of large-scale LiDAR point clouds is a fundamental problem in 3D scene understanding for autonomous systems and robotic perception. Existing approaches predominantly rely on local geometric cues and neighborhood-based representations, which are often insufficient to capture the global structural organization of semantic classes. In real-world environments, object categories such as roads, buildings, and vegetation exhibit characteristic connectivity patterns and structural arrangements that are inherently topological in nature. This work introduces a topology-aware feature extraction pipeline based on Topological Data Analysis (TDA) to explicitly model the structural properties of semantic regions in 3D point clouds. Class-wise point subsets are first extracted from semantic predictions, and multi-scale filtrations are constructed to analyze their connectivity patterns. Persistent homology is then employed to capture topological features such as connected components and loop structures, which are summarized in persistence diagrams. To enable integration with learning-based frameworks, persistence diagrams are subsequently transformed into fixed-length feature vectors using a learnable PersLay layer. Experimental results on a public LiDAR benchmark demonstrate that different semantic classes exhibit distinctive topological signatures and that incorporating topology-aware features enhances the discriminative capability of 3D semantic segmentation models. The findings indicate that topological representations provide a complementary cue to conventional geometric features for improved 3D semantic segmentation. |