ITU-Campus3D: A Campus-Scale Urban Mobile Mapping Benchmark for Semantic Segmentation

Automated semantic segmentation of the three-dimensional (3D) built environment remains a fundamental challenge in 3D computer vision and urban perception, supporting applications ranging from autonomous navigation to urban infrastructure management. Existing benchmarks are primarily focused on highway environments, limiting their ability to capture the geometric ambiguity and spatial complexity of mixed-use urban scenes. To overcome this limitation, ITU-Campus3D is introduced as a large-scale mobile laser-scanning dataset containing over 239 million annotated points collected in a complex campus environment. In addition to dataset construction, a robust benchmarking framework is established that incorporates an adaptive spatial split strategy and tile-level rare-class balancing to mitigate spatial correlation and severe class imbalance during evaluation. Using this framework, five state-of-the-art segmentation architectures, spanning point-based, voxel-based, and hybrid paradigms, are systematically evaluated. While modern architectures achieve competitive overall performance, reaching a mean intersection-over-union (mIoU) of 73.18% with an overall accuracy of about 91.55%, conventional evaluation metrics are shown to conceal failures in rare semantic classes. To explicitly assess segmentation reliability under long-tailed class distributions, the rare-class metric mIoU_rare is introduced. The evaluation reveals an approximately 8.00% performance gap between the mIoU and mIoU_rare, highlighting a critical limitation of current urban scene understanding models. Further analysis indicates that segmentation errors on structurally sparse urban objects, such as signs and poles, are largely driven by geometric similarity, occlusion, and complex neighborhood complexity. By explicitly exposing these failure modes, ITU-Campus3D serves as a diagnostic benchmark that advances urban semantic segmentation beyond aggregate accuracy metrics toward reliable real-world deployment.