Research map / Detection and segmentation / 3D detection and driving
MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MCD, a multi-campus robot perception dataset with CCS and NRE lidars, IMU, cameras, UWB, semantic labels on NRE lidar scans, and continuous-time survey-map ground truth, benchmarking SLAM and segmentation methods.
Problem
Existing well-annotated perception datasets are biased toward autonomous driving, use costly sensors, lack environment diversity and accurate ground truth, and have no semantic annotations for low-cost NRE lidar.
Approach
Collects multimodal data across three Eurasian university campuses with NRE and CCS lidars, UWB, and IMU, provides point-wise semantic labels for NRE lidar, and generates continuous-time ground truth by registering raw lidar-inertial data to survey-grade prior maps with B-spline trajectory optimization.
Key results
- Across 16 SLAM baselines on MCD sequences, no single method consistently excels; ATE often exceeds 0.5 m and many methods diverge on nighttime sequences.
- Eight 3D semantic segmentation methods suffer mIoU drops of 21-44% relative to SemanticKITTI when trained and tested on the NRE lidar NTU subset; WaffleIron and SphereFormer are the only ones above 50% mIoU.
- Cross-campus testing on TUHH drops mIoU by more than 70% relative to NTU for all segmentation methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Test semantic segmentation on NRE lidar instead of SemanticKITTI lidar | 3D semantic segmentation methods · MCD NTU | mIoU: -26.9% to -43.6% | Hurt |
| Test on a different campus (TUHH) than the NTU training campus | 3D semantic segmentation methods trained on MCD NTU · MCD TUHH | mIoU: more than 70% | Hurt |
| Use continuous-time formulation and multi-scale point-to-surfel association | SLICT · MCD SLAM sequences | ATE | Helped |
| Use visual factors to fill in when UWB line of sight is lost | VRO versus RO · MCD range-aided localization sequences | ATE | Helped |
| Evaluate visual-inertial SLAM in nighttime sequences | VINS, KIMERA, and OPENVINS · MCD nighttime sequences | ATE | Hurt |
| Use LIO-SAM on high-speed ATV sequences rather than HHS sequences | LIO-SAM · MCD NTU ATV sequences | ATE | Hurt |
Surprising findings
- LIO-SAM performs well on handheld sequences but loses its advantage on high-speed ATV sequences, while SLICT is consistent across both.
- Cross-campus segmentation performance collapses by more than 70% relative to same-campus results, indicating severe domain shift.
Limitations
- Ground truth relies on survey-grade prior maps and line-of-sight tracking in some prior datasets, not fully applicable to all settings
- Many SLAM methods diverge on nighttime and high-speed sequences
- Segmentation methods generalize poorly across campuses
Experimental setup
Datasets: MCD SemanticKITTI
Benchmarks: MCD SemanticKITTI
Metrics: ATE mIoU IoU
Baselines: LIO-SAM Fast-LIO DLIO SLICT VINS KIMERA OpenVINS RO VRO SalsaNext MinkowskiNet SPVCNN
Builds on: LIO-SAM Fast-LIO SLICT Newer College Dataset
Closest papers by meaning
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · similarity 0.87)
- Multi-Space Alignments Towards Universal LiDAR Segmentation (CVPR 2024 · similarity 0.87)
- USCILab3D: A Large-scale, Long-term, Semantically Annotated Outdoor Dataset (NeurIPS 2024 · similarity 0.87)
- HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.86)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.85)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.