Skip to content

RealSim-CP

A high-fidelity, multi-modal cooperative-perception dataset for Japanese urban traffic.

RealSim-CP provides synchronized camera + LiDAR data captured from multiple connected vehicles and roadside units (RSUs) observing the same traffic scene, with 3D object annotations in the ASAM OpenLABEL 1.0.0 schema. Every scene is generated with the physics-based DIVP simulator (ray tracing + electromagnetic-wave sensor modeling), so the data stays close to real sensor physics while costing a fraction of real-world collection.

RealSim-CP dataset overview

/// caption Multiple vehicles and RSUs observe one shared traffic scene. Each variant ships synchronized camera images, LiDAR point clouds, calibration, per-frame poses, and OpenLABEL 3D annotations. ///

  • :material-rocket-launch: Quickstart Download a scene and explore it in the 3D visualizer in minutes.

  • :material-download: Download the dataset Get the data from the project's Google Drive.

  • :material-folder-table: Dataset structure Scenarios, weather/time variants, agents, and sensor streams.

  • :material-code-json: Label format OpenLABEL schema, coordinate systems, and the transform chain.

  • :material-tools: Tools 3D visualizer, label-generation automation, and more.

  • :material-chart-box: Benchmark CoopDet3D cooperative-perception results.

Why RealSim-CP?

Existing cooperative-perception datasets share two gaps that RealSim-CP is built to close:

  1. Japanese traffic is under-represented. Left-hand traffic, distinctive vehicle types (kei cars, Japanese buses/trucks), and region-specific infrastructure are rarely captured. Models trained elsewhere degrade when moved to a new region, so region-specific data matters.
  2. Real-world collection is expensive. Instrumenting many cooperative agents and labeling multimodal data at scale is costly and slow. High-fidelity simulation produces large, diverse, perfectly-labeled data at far lower cost.

See the Overview for the full motivation, methodology, and contributions.

At a glance
Domain Cooperative perception · V2V / V2I / V2X · autonomous driving
Modalities RGB camera images + LiDAR point clouds (.pcd)
Agents Connected vehicles + roadside units (RSUs)
Region Tokyo, Japan (Aomi, Odaiba/Daiba, Shutoko Expressway)
Conditions Clear day · rainy day · clear night
Annotations 3D cuboids, 12 object classes, ASAM OpenLABEL 1.0.0
Rate 10 Hz synchronized capture
Generator DIVP physics-based simulator
Dataset license CC BY 4.0
Code license MIT

Citation

If you use RealSim-CP, please cite the paper — see the Citation section of the README or CITATION.cff.