Research map / 3D vision / 3D generation
MAN TruckScenes: A multimodal dataset for autonomous trucking in diverse conditions
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 MAN TruckScenes, a 747-scene multimodal dataset for autonomous trucking with 4 cameras, 6 lidars, 6 4D radars, and baseline 3D detection results for PETR, RadarGNN, and CenterPoint.
Problem
No large-scale multimodal perception datasets exist for autonomous trucks, which face truck-specific challenges such as trailer occlusion, novel sensor perspectives, and terminal environments.
Approach
A truck-specific multimodal dataset with 360° 4D radar coverage, long-range (230 m) manually annotated 3D boxes with tracking IDs, 34 scene tags, and a multi-objective optimization method (NSGA-II) for constructing balanced train/val/test splits.
Key results
- CenterPoint lidar baseline achieves 0.27 mAP and 0.41 NDS on the test split, outperforming PETR camera (0.02 mAP / 0.12 NDS) and RadarGNN radar (0.07 mAP / 0.11 NDS).
- Lidar detection performance drops from 0.48 mAP (0-25 m) to 0.27 mAP (0-150 m) for CenterPoint.
- The dataset comprises 747 scenes of about 20 s each, with 4 cameras, 6 lidars, 6 4D radars, 2 IMUs and RTK-GNSS, and 27 object classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate over a longer detection range (0–150 m instead of 0–25 m) | PETR · MAN TruckScenes test, 3D object detection | mAP: 0.06 -> 0.02 | Hurt |
| Evaluate over a longer detection range (0–150 m instead of 0–25 m) | RadarGNN · MAN TruckScenes test, 3D object detection | mAP: 0.09 -> 0.07 | Hurt |
| Evaluate over a longer detection range (0–150 m instead of 0–25 m) | CenterPoint · MAN TruckScenes test, 3D object detection | mAP: 0.48 -> 0.27 | Hurt |
Surprising findings
- Camera-based PETR baseline achieves very low performance (0.02 mAP), with many classes at zero AP.
- Radar-based RadarGNN performs poorly on non-metallic object classes and in tunnels and container terminals.
Limitations
- Data limited to public roads and logistics terminals in Germany recorded on a single autonomous test truck.
- Manual annotation is subject to human errors despite quality assurance.
- Radar calibration relies only on external measurements, leaving possible angular errors up to 0.03 degrees.
Experimental setup
Datasets: MAN TruckScenes
Benchmarks: MAN TruckScenes
Metrics: mAP NDS ATE ASE AOE AVE AAE
Baselines: PETR RadarGNN CenterPoint
Builds on: nuScenes PETR CenterPoint RadarGNN NSGA-II
Closest papers by meaning
- TruckDrive: Long-Range Autonomous Highway Driving Dataset (CVPR 2026 · similarity 0.89)
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · similarity 0.86)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.85)
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception (CVPR 2024 · similarity 0.84)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.84)
- OpenAD: Open-World Autonomous Driving Benchmark for 3D Object Detection (NeurIPS 2025 · similarity 0.83)
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.