Abstract. Accurate semantic segmentation of 3D railway point clouds is essential for enabling automated inspection and asset management. Although recent deep learning (DL) models achieve strong performance on large benchmark datasets, their ability to generalize to point clouds captured with different sensors and in different spatial environments remains insufficiently explored. This study investigates the cross-sensor robustness and spatial generalization of state-of-the-art DL architectures for 3D semantic segmentation in railway scenarios. Three advanced models, Point Transformer V3, MinkUNet, and Swin3D, were trained on the SemanticRail3D dataset and evaluated on a newly acquired railway section scanned using three heterogeneous LiDAR systems: a terrestrial laser scanner (Faro Focus S150+), and two handheld mobile mapping devices (CHCNAV RS10 and GeoSLAM ZEB Go). The test area was manually annotated to provide high-quality ground truth for quantitative assessment. Results show substantial performance variations across sensors, driven by differences in point density, noise levels, and scanning geometry. Domain-shift effects were evaluated directly from the model prediction outputs, including per-class IoU differences, uncertainty patterns, and cross-model agreement across sensors. To improve the robustness, an ensemble fusion strategy is evaluated to mitigate cross-sensor variability. The findings highlight the challenges of deploying DL models in real-world railway environments and provide insights for improving sensor-agnostic segmentation pipelines.
Ghasemlou et al. (Fri,) studied this question.