Unleashing the power of OpenStreetMap tags: a graph neural network approach for efficient LiDAR point cloud classification
- 15 November 2023
- conference paper
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12815, 62-739
Abstract
When employing Graph Neural Networks (GNN) for point cloud classification, global information is passed through the iteration of hidden states of graph nodes. However, this method is often associated with high memory and time costs. To address this issue, we propose a strategy that integrates prior OpenStreetMap (OSM) tags into LiDAR point cloud classification, which enhance iteration speeds by fixing the initial hidden states of certain graph nodes. First, we perform over-segmentation of the LiDAR point cloud to obtain super points and embedded features. Then, OSM tags are affixed to the super points and incorporated into the GNN to update other super point embedded features. Our results demonstrate that coupling OSM with GNN significantly improves classification accuracy for roof, tree, and impervious surface categories.Keywords
This publication has 5 references indexed in Scilit:
- Assessing OSM building completeness for almost 13,000 cities globallyInternational Journal of Digital Earth, 2022
- LiDAR-OSM-Based Vehicle Localization in GPS-Denied Environments by Using Constrained Particle FilterSensors, 2022
- Deep Learning for 3D Point Clouds: A SurveyIEEE Transactions on Pattern Analysis and Machine Intelligence, 2021
- ConvPoint: Continuous convolutions for point cloud processingComputers & Graphics, 2020
- Improving LiDAR classification accuracy by contextual label smoothing in post-processingISPRS Journal of Photogrammetry and Remote Sensing, 2019