Research map / Vision-language and multimodal / Human and embodied modeling
V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative 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 V2X-Radar, a real-world multi-modal dataset with 4D Radar, LiDAR and cameras for cooperative, roadside and single-vehicle 3D detection, with benchmarks across these three subsets.
Problem
Cooperative perception datasets focus on cameras and LiDAR and lack 4D Radar, which is robust in adverse weather, leaving no real-world V2X benchmark that includes 4D Radar.
Approach
A real-world V2X dataset collected with a connected vehicle platform and roadside unit, each equipped with LiDAR, multi-view cameras and 4D Radar, split into cooperative, roadside and single-vehicle sub-datasets with benchmarks for early, late and intermediate fusion under sync and async delays.
Key results
- On V2X-Radar-C with LiDAR, HEAL reaches 67.57 AP@0.7 (sync) versus 39.37 for late fusion and 7.13 for no fusion.
- Under 100 ms asynchronous delay, LiDAR HEAL drops from 67.57 to 57.76 AP@0.7, showing a substantial degradation.
- In adverse weather, 4D Radar-only PointPillars outperforms LiDAR-only on vehicles by about 1-2 AP at IoU 0.5 (48.35 vs 47.23 easy).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use cooperative LiDAR fusion rather than no fusion | V2X-Radar-C vehicle detection · Sync, overall | AP@IoU=0.7 / 0.5: 7.13 / 29.09 -> 67.57 / 83.00 | Helped |
| introduce 100 ms communication delay | HEAL with LiDAR on V2X-Radar-C · Overall | AP@IoU=0.7 / 0.5: 67.57 / 83.00 -> 57.76 / 79.66 | Hurt |
| use HEAL instead of no fusion | V2X-Radar-C with 4D Radar · Sync, overall | AP@IoU=0.7 / 0.5: 2.69 / 9.02 -> 12.50 / 29.04 | Helped |
| add 4D Radar to LiDAR (ablation) | Pointpillars on V2X-Radar adverse-weather frames · Vehicle Easy, IoU=0.5 | AP: 47.23 -> 53.61 | Helped |
| add Doppler information (ablation) | 4D Radar perception on adverse-weather V2X-Radar frames · Vehicle Moderate, IoU=0.5 | AP: 41.39 -> 44.80 | Helped |
| add Doppler information (ablation) | 4D Radar perception on adverse-weather V2X-Radar frames · Pedestrian Easy, IoU=0.25 | AP: 19.98 -> 20.14 | Helped |
| replace LiDAR with 4D Radar (ablation) | Pointpillars on V2X-Radar adverse-weather frames · Vehicle Easy, IoU=0.5 | AP: 47.23 -> 48.35 | Helped |
| replace LiDAR with 4D Radar (ablation) | Pointpillars on V2X-Radar adverse-weather frames · Pedestrian Easy, IoU=0.25 | AP: 21.12 -> 20.14 | Hurt |
| increase transmission delay | Cooperative perception methods on V2X-Radar-C · Async transmission delay from 0 ms to 400 ms | AP@IoU=0.5: 83.00 -> 69.52 | Hurt |
Surprising findings
- Camera-based and 4D Radar-based methods underperform LiDAR in normal conditions, but 4D Radar-only beats LiDAR-only under adverse weather.
- Communication delay causes severe performance drops in cooperative methods, with 0.7-IoU AP falling notably under async settings.
Limitations
- Limited temporal coverage and task diversity; no tracking or trajectory prediction tasks yet.
- Benchmarks use a small number of sequences (30 train, 5 val, 5 test) for cooperative split and evaluate on validation set.
- Models trained on this dataset may not generalize to other vehicle or roadside sensor platforms.
Experimental setup
Datasets: V2X-Radar-C V2X-Radar-I V2X-Radar-V
Benchmarks: V2X-Radar-C V2X-Radar-I V2X-Radar-V
Metrics: AP@IoU=0.7 AP@IoU=0.5 AM (transmission cost)
Baselines: PointPillars SECOND CenterPoint PV-RCNN SQDNet Fade3D SMOKE BEVDepth BEVHeight BEVHeight++ RDIoU RPFA-Net
Builds on: PointPillars LSS HEAL V2X-ViT CoAlign
Closest papers by meaning
- TUMTraf V2X Cooperative Perception Dataset (CVPR 2024 · similarity 0.89)
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.89)
- RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar (NeurIPS 2024 · similarity 0.88)
- RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception (CVPR 2024 · similarity 0.87)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.87)
- BEVSync: Asynchronous Data Alignment for Camera-based Vehicle-Infrastructure Cooperative Perception Under Uncertain Delays (AAAI 2025 · similarity 0.87)
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.