How the Choice of SLAM Algorithm Impacts Point Cloud Classification: An Experimental Evaluation

Abstract. This paper investigates the impact of different Simultaneous Localization and Mapping (SLAM) algorithms on semantic point cloud classification, by analyzing how reconstruction characteristics affect downstream classification tasks. Four SLAM approaches (RTAB-Map, LIO-SAM, DLO, and LeGO-LOAM ) are used to process identical raw LiDAR and IMU data acquired by a sensorized mobile robotic platform, producing reconstructions with markedly different point density and geometric quality. A Random Forest classifier based on multi-scale geometric features is then applied to the SLAM-derived point clouds and to reference data acquired using Terrestrial Laser Scanning (TLS), assessing both in-domain classification performance and cross-dataset generalization. The results show that classification accuracy and class-wise reliability strongly depend on SLAM-induced point cloud characteristics, with denser reconstructions yielding higher performance. Cross-dataset experiments further reveal an asymmetric generalization behavior, whereby models trained on SLAM-derived point clouds transfer more robustly to denser datasets than models trained on TLS data. Feature importance analysis links this behavior to a shift toward coarser geometric descriptors as point density decreases. To support reproducibility and to encourage further investigations, including the evaluation of alternative classification strategies, the SLAM and TLS point clouds are released as an open-access dataset, together with manually annotated ground-truth labels.