LMSeg: An end-to-end geometric message-passing network on barycentric dual graphs for large-scale landscape mesh segmentation

The introduction of LMSeg marks a significant advancement in the field of semantic segmentation for large-scale 3D landscape meshes, a critical area for geospatial analysis. By leveraging the BudjBim Wall (BBW) dataset, which consists of high-resolution LiDAR scans from a UNESCO World Heritage site in Victoria, Australia, LMSeg aims to overcome persistent challenges in scalability and accuracy, particularly for small and irregular objects. The network employs a barycentric dual graph representation and a novel Geometry Aggregation+ module to enhance feature extraction and segmentation accuracy. Initial experiments demonstrate promising results, with LMSeg achieving 62.4% mean Intersection over Union (mIoU) on the BBW dataset, 78.4% overall accuracy on H3D, and 75.1% mIoU on SUM, all while maintaining a lightweight architecture with 2.4 million parameters. This development not only supports advancements in AI but also contributes to the preservation and understanding of underrepresented…