Research map / Multi-agent systems and games / Human-AI and goal-conditioned
UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections 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 UrbanIng-V2X, a real-world multi-vehicle, multi-infrastructure cooperative perception dataset from three Ingolstadt intersections with thermal cameras, plus LiDAR-only benchmarks and new intersection-aware splits that expose a generalization gap.
Problem
Existing real-world V2X cooperative perception datasets cover only a single intersection or a single vehicle, so benchmarks cannot measure generalization across diverse intersection layouts and multi-agent configurations.
Approach
A real-world dataset combining two vehicles and up to three infrastructure sensor poles (with thermal cameras) across three intersections, plus intersection-wise (SIS) and sequence-wise (EIS) splits to test generalization to unseen intersections.
Key results
- CoBEVT reaches 24.9 mAP@0.5 on the SIS 1/2vs.3 split (unseen intersection) versus 38.2 mAP@0.5 averaged over EIS splits, a drop of about 14 points.
- AttFuse achieves the best SIS result with 25.5 mAP@0.5, surpassing other intermediate fusion methods by at least 0.6 mAP@0.5.
- Late fusion is the weakest cooperative method at 11.1 mAP@0.5, below no fusion at 18.3 mAP@0.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use early fusion instead of no fusion | PointPillars cooperative perception · SIS₁/₂vs.₃ | mAP@0.5: 18.3 -> 22.8 | Helped |
| Use late fusion instead of no fusion | PointPillars cooperative perception · SIS₁/₂vs.₃ | mAP@0.5: 18.3 -> 11.1 | Hurt |
| Use intermediate fusion instead of no fusion | PointPillars cooperative perception · SIS₁/₂vs.₃ | mAP@0.5: 18.3 -> 24.2; 18.3 -> 25.5; 18.3 -> 22.5; 18.3 -> 23.8; 18.3 -> 24.9 | Helped |
| Use AttFuse rather than the other evaluated methods | PointPillars cooperative perception · SIS₁/₂vs.₃ | mAP@0.5: 25.5 | Helped |
| Evaluate CoBEVT on known intersections rather than unseen intersections | CoBEVT · EIS avg vs. SIS | mAP@0.5: 38.2; 24.2; 14.0 | Helped |
| Evaluate CoBEVT on the unseen Intersection 2 rather than the unseen Intersection 3 | CoBEVT · SIS₁/₃vs.₂ vs. SIS₁/₂vs.₃ | mAP@0.5: 24.6 -> 19.1 | Hurt |
| Evaluate CoBEVT on the unseen Intersection 1 rather than the unseen Intersection 3 | CoBEVT · SIS₂/₃vs.₁ vs. SIS₁/₂vs.₃ | mAP@0.5: 24.6 -> 28.9 | Helped |
| Use V2X-ViT rather than AttFuse | PointPillars cooperative perception · EIS single split | mAP@0.5: 28.8 -> 33.8 | Helped |
Surprising findings
- Late fusion underperforms even no fusion, indicating association of agent-specific object lists is a major challenge.
- Frame-wise splits are prone to leakage, so the paper favors sequence-wise and intersection-wise splits, which yield lower scores.
Limitations
- Only three intersections in Ingolstadt, Germany, with no adverse weather coverage.
- Annotations rely on LiDAR data, so objects visible only in cameras at long range may be unannotated.
- Per-agent recording duration (about 11 minutes) is shorter than single-agent AV datasets.
Experimental setup
Datasets: UrbanIng-V2X
Benchmarks: UrbanIng-V2X
Metrics: AP@0.3 AP@0.5 mAP@0.3 mAP@0.5
Baselines: No Fusion Early Fusion Late Fusion F-Cooper AttFuse V2X-ViT Where2Comm CoBEVT
Builds on: OpenCOOD PointPillars V2X-Real OpenPCDet
Closest papers by meaning
- HoloVIC: Large-scale Dataset and Benchmark for Multi-Sensor Holographic Intersection and Vehicle-Infrastructure Cooperative (CVPR 2024 · similarity 0.90)
- RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception (CVPR 2024 · similarity 0.89)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.89)
- TUMTraf V2X Cooperative Perception Dataset (CVPR 2024 · similarity 0.88)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.87)
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · 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.