Reliability-aware superpoint graph learning for large-scale point cloud semantic segmentation: Evaluation on indoor, mobile, and airborne benchmarks

Large-scale point cloud semantic segmentation requires efficient contextual reasoning over complex 3D scenes. Superpoint-based methods reduce computation by operating on compact region graphs, but geometrically retained adjacency does not always imply semantically reliable interaction, which may cause cross-boundary feature leakage. Building on the hierarchical Superpoint Transformer, we develop RSGT as a reliability-aware extension that explicitly models the trustworthiness of retained inter-superpoint interactions. RSGT uses bounded, head-specific edge modulation to regulate contextual propagation without changing graph topology, together with selective decoder-side local correction for residual uncertain predictions. RSGT achieves 77.6% and 70.9% mIoU on S3DIS 6-fold and Area-5, 65.4% on KITTI-360 validation, and 81.1% on DALES test, improving upon the published Superpoint Transformer (SPT) results by 1.5–2.0 mIoU points across these four benchmark settings. It also remains competitive with strong point-based methods, reaching 77.6% versus 77.7% for PTv3 on S3DIS 6-fold and matching the literature-reported 81.1% of KPConv on DALES. Under the controlled S3DIS measurement protocol, these gains require a 5.96% increase in parameters and approximately 1.36% additional full-pipeline runtime relative to SPT. Edge-level, spatial, correction-based, and boundary-focused analyses further demonstrate that the learned reliability responses concentrate on semantically difficult interactions and contribute directly to improved boundary-sensitive predictions. The source code and benchmarking configurations are publicly available at https://github.com/Esperance-liren/RSGT.