Semantic segmentation of imbalanced 3D point clouds in railway environments : comparative analysis of algorithms and training pipelines for semantic segmentation

Railway infrastructure is vital for modern transportation, and effective maintenance is crucial for ensuring safety and efficiency. Semantic segmentation of 3D point clouds can automate monitoring and support digital twins, yet it demands large amounts of training data. Existing large-scale datasets face challenges such as severe class imbalance, numerous unclassified points, and noise from distant objects, hindering model performance. In this paper, we address these issues using the SemanticRail3D dataset, which comprises 438 point clouds spanning 90 km of railway tracks with approximately 2.8 billion labeled points across 11 classes. We propose a novel preprocessing and training pipeline that employs PCA-based alignment, segmentation, a slight data augmentation, and tailored training configurations. Two state-of-the-art models, Swin3D and Point Transformer V3, are trained, achieving mean IoU scores of 82.47% and 82.95% on held-out test data, respectively. Our results demonstrate strong generalization and lay a solid foundation for automated railway infrastructure monitoring.