LMSeg: an end-to-end geometric message-passing network on barycentric dualgraphs for large-scale landscape mesh segmentation

A novel deep learning approach, LMSeg, achieves state-of-the-art performance in segmenting large-scale 3D landscape meshes, including complex cultural heritage structures like the UNESCO-listed Budj Bim dry-stone walls. Unlike existing methods, LMSeg combines geometric message-passing with hierarchical-local dual pooling, enabling accurate detection of small and occluded objects while maintaining scalability. Benchmark results on three datasets demonstrate its superiority over point- and graph-based baselines, with 62.4% mean Intersection-over-Union (mIoU) on the BudjBim Wall dataset despite using only 2.4 million lightweight parameters. This advancement addresses longstanding challenges in mesh segmentation for cultural heritage preservation and environmental monitoring.

Huang, Z., Khoshelham, K., & Tomko, M. (2025). LMSeg: an end-to-end geometric message-passing network on barycentric dualgraphs for large-scale landscape mesh segmentation. GIScience & Remote Sensing, 62(1). https://doi.org/10.1080/15481603.2025.2584806