A Trade-Off Analysis of Point Cloud Density for Real-Time 3D Semantic Segmentation

Abstract The rapid growth of 3D LiDAR technology has significantly advanced environmental mapping capabilities; however, the dense point clouds generated present substantial computational and also storage challenges, particularly for real-time semantic segmentation in resource-limited environments. This study investigates the balance between point cloud density and segmentation accuracy by employing a baseline PointNet-based neural network on the extensive Semantic3D dataset. We systematically assess three different density levels (100%, 10%, and 1%) under the same model settings, training parameters, and hardware conditions. Results indicate that reducing density to just 1% of the original data maintains a robust overall accuracy of 77.7%, only slightly below the full-density accuracy of 78.8%, while greatly decreasing computational requirements by over 80%. Class-wise analysis identifies that more complex shaped and smaller objects, notably vehicles, experience greater accuracy degradation under severe subsampling, suggesting potential improvements through adaptive sampling methods. This research offers valuable insights for implementing real-time 3D semantic segmentation in low-cost drones, robots, and embedded systems, and paves the way for future exploration into adaptive sampling, efficient neural architectures, and multi-sensor fusion strategies.