Sensors

Each cooperative agent carries a calibrated sensor rig. The data is generated by the DIVP physics-based simulator, which models the optics and signal physics of real sensor hardware rather than producing idealized outputs.

Sensors per agent

Agent	Cameras	LiDAR
Vehicle	4 (arranged around the body)	1 (roof)
RSU	1	1 (sometimes camera-only)

Cameras

• Modeled hardware: Sony IMX490 automotive sensor, 120° horizontal field of view. DIVP models the spectral response on the CMOS sensor, solar illumination via sky models, and lens distortion, rendering spatially with Monte Carlo ray tracing.
• Image resolution (Daiba release): 1158 × 750 px, saved as PNG.
• Calibration: a custom intrinsic model — fx, fy, cx, cy plus six distortion coefficients k1…k6 — and an external distortion table (lens_distortion_i49.csv). The intrinsics for each camera are declared in the label file's streams block.

Projection axis convention

RealSim camera projection uses +X forward, +Y image-left, and +Z up, so a projection helper must flip the horizontal axis when drawing onto image pixels. Some rendered images also align better with a pinhole model than with the full k1…k6 distortion — the visualizer exposes a distortion toggle for exactly this reason.

Example camera stream entry from a label file:

"vehicle_10000/camera_9": {
  "type": "camera",
  "stream_properties": {
    "intrinsics_custom": {
      "distortion_file": "lens_distortion_i49.csv",
      "camera_parameters": { "fx": 599.79, "fy": 599.68, "cx": 579.0, "cy": 375.0,
                             "k1": 2.69,   "k2": 0.60,   "k3": 0.0081,
                             "k4": 3.15,   "k5": 1.51,   "k6": 0.093 }
    },
    "width": 1158, "height": 750
  }
}

LiDAR

• Modeled hardware: Velodyne VLS-128. DIVP traces rays to obtain signal propagation and solves the corresponding physical equations for received intensity, reproducing effects such as backlighting.
• Files: one .pcd per frame, named point_cloud_<i>.pcd, under point_clouds/<agent>/<lidar>/velodyne_vlp_128/.
• Format: ASCII PCD with FIELDS x y z (32-bit floats).

Stream identifier

Inside the dataset the LiDAR stream directory and identifier use the string velodyne_vlp_128. The simulated sensor characteristics follow the Velodyne VLS-128 model described in the paper.

Sensor placement

• RSUs (Aomi & Odaiba): LiDAR mounted at ~2 m height, horizontal; cameras tilted toward the roadway. On the Shutoko Expressway, sensors are placed on virtual objects near signboards (one facing the road, one facing the opposite direction). RSUs are a mix of camera-only and camera + LiDAR.
• Vehicles: 4 cameras (front / rear / left / right) and 1 roof LiDAR.

Each sensor's mounting (extrinsic) pose is given relative to its parent agent in the label file's coordinate_systems; the agent's pose in the scene is given per frame in frame_properties.transforms. To place a sensor in the world you compose the two — see Label format → Coordinate systems.

Capture rate

All cameras and LiDARs are recorded synchronously at 10 frames per second.