Back to home
Investigating the Impact of Point Cloud Density on Semantic Segmentation Performance using Virtual Lidar in Boreal Forest
Olivier Stocker, Reza Mahmoudi Kouhi, Eric Guilbert, Antonio Ferraz, Thierry Badard
Virtual LiDAR scan (VLS) serves as a powerful tool for the replication of real world conditions and can assist with the calibration of LiDAR systems. In this study, we utilize HELIOS++, a VLS software, to investigate the impact of point cloud density on the semantic segmentation performance of a well-established Deep Learning (DL) method for point clouds, KPConv. Our experiment is focused on a typical Quebec boreal forest composed of Abies balsamea and Picea mariana. We generated 10250 structurally diverse forest plots to train 10 DL models on a wide range point cloud densities to assess their effect on the semantic segmentation. Densities varied from 23 points/m2 to 225 points/m2, replicating point clouds output from classic airborne LiDAR scanning and high-density unmanned LiDAR scanning. Our results demonstrate that point cloud densification improves IoU score for both boreal tree species by an average of 0.3 percentage points per 10 points/m2.
Semantic segmentation resluts and plots vizualisations
- Plot 6853 : Tree and grounds models and semantic segmentation results and ground truth for set 10.
- Plot 5287 : Semantic segmentation results, ground truth and errors for all sets.