MeshSegNet: a local to global feature fusion-based semantic segmentation approach for urban meshes

Mesh semantic segmentation is essential for 3D scene understanding, with applications in urban planning, autonomous navigation, and smart city. However, the irregular structure of meshes limits the extraction of global information, and the noise also poses a challenge to high-precision semantic segmentation of small objects. To address this, we propose MeshSegNet, a novel method for accurate semantic segmentation of urban meshes. MeshSegNet incorporates a local feature extractor and a global feature aggregator to effectively integrate features from local to global. The local feature extractor captures intrinsic and extrinsic attributes of mesh vertex to express local features. The global feature aggregator can adaptively change the diffusion time to achieve local-to-global feature aggregation, effectively suppressing noise while maintaining boundary accuracy. Moreover, MeshSegNet attains superior performance across all metrics on the SUM and H3D datasets, validating its effectiveness and reliability.