Research map / Point clouds / Point cloud classification
RCooper: A Real-world Large-scale Dataset for Roadside 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.
Releases RCooper, a real-world large-scale roadside cooperative perception dataset (50k images, 30k point clouds, 10 classes, intersection and corridor scenes) with 3D detection and tracking benchmarks.
Problem
Single-infrastructure roadside perception suffers from limited sensing range and blind spots, and there is no real-world public dataset for roadside cooperative perception.
Approach
A real-world, multi-sensor roadside dataset with per-infrastructure 3D box annotations plus cooperative annotations generated by transforming to a world frame, Hungarian matching, and manual refinement, along with benchmarks for cooperative 3D detection and tracking-by-detection.
Key results
- In the corridor scene, CoBEVT reaches 57.2 AP@0.5 and 36.2 AP@0.7 versus 29.2 and 11.1 for no fusion, while early fusion reaches the best AP@0.3 (69.8).
- In the intersection scene, late fusion achieves the best AP@0.3 (65.1) and AP@0.5 (47.6) among cooperative methods, and early fusion drops to 50.0 AP@0.3, below no fusion (58.1).
- Removing the two MEMS LiDARs markedly boosts early fusion and feature fusion methods in the intersection scene, pointing to data heterogeneity as a key challenge.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use cooperative detection instead of No Fusion | Roadside perception · Corridor | AP@0.3: 40.0 -> 69.8 | Helped |
| use cooperative detection instead of No Fusion | Roadside perception · Corridor | AP@0.5: 29.2 -> 57.2 | Helped |
| use cooperative detection instead of No Fusion | Roadside perception · Intersection | AP@0.5: No Fusion 44.1; Late fusion 47.6; Early fusion 33.9 | Mixed |
| exclude point clouds from MEMS LiDARs (ablation) | Early fusion and feature-fusion detection methods · Intersection | detection performance | Helped |
| use cooperative tracking instead of No Fusion | Roadside tracking · Corridor | AMOTA: 8.28 -> 23.78 | Helped |
| use cooperative tracking instead of No Fusion | Roadside tracking · Intersection | AMOTA: No Fusion 18.11; Late fusion 21.57; AttFuse 11.84; F-Cooper -4.86 | Mixed |
Surprising findings
- In intersection scenes, early fusion performs worse than no fusion, and several SOTA intermediate fusion methods underperform simple late fusion.
- Excluding MEMS LiDAR data greatly improves cooperative methods, showing heterogeneous sensor data hurts fusion.
Limitations
- Benchmarks use only representative existing methods and do not address data heterogeneity or cooperative representation fully.
- Tracking relies on tracking-by-detection, so it is sensitive to unstable detections in intersections.
- Calibration noise and end-to-end approaches are left for future work.
Experimental setup
Datasets: RCooper
Benchmarks: RCooper
Metrics: AP@0.3 AP@0.5 AP@0.7 AMOTA AMOTP sAMOTA MOTA MT ML
Baselines: No Fusion Late fusion Early fusion AttFuse F-Cooper Where2Comm CoBEVT
Builds on: PointPillar AB3DMOT AttFuse F-Cooper Where2Comm CoBEVT
Closest papers by meaning
- UrbanIng-V2X: A Large-Scale Multi-Vehicle, Multi-Infrastructure Dataset Across Multiple Intersections for Cooperative Perception (NeurIPS 2025 · similarity 0.89)
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.87)
- TUMTraf V2X Cooperative Perception Dataset (CVPR 2024 · similarity 0.86)
- HoloVIC: Large-scale Dataset and Benchmark for Multi-Sensor Holographic Intersection and Vehicle-Infrastructure Cooperative (CVPR 2024 · similarity 0.86)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.85)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 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.