Dataset

EagleVision is a unified LiDAR-based perception benchmark for high-speed autonomous racing, standardized into a common annotation and coordinate convention.

Domains

Three standardized datasets

IAC (Indy Autonomous Challenge): ROS bag recordings with GPS-based state; 3D boxes are manually labeled.
A2RL Simulator: Official simulator data with ground-truth 3D boxes exported to the unified format.
A2RL Real-World: Competition racing data (ROS bags) manually labeled under real sensor noise and occlusion.

What’s annotated

Detection + prediction

3D Detection: single class Car with 3D bounding boxes (PSR).
Trajectory Prediction: per-frame ego/vehicle pose entries used to build observation/prediction windows.

Dataset statistics

Domain	LiDAR	Hz	Annotated frames	Avg objects/frame	Points/scan (approx.)

Annotation format

3D Detection (PSR JSON)

Position, Scale, Rotation (yaw)

Boxes are defined in the ego-vehicle LiDAR frame, without motion compensation.

Trajectory Prediction (Pose JSON)

Frame-wise position + quaternion

Each entry stores frame id, timestamp, 3D position, and unit quaternion orientation.

Coordinate convention (minimal)

All labels are in the ego-vehicle LiDAR coordinate frame.
3D boxes are parameterized by center (x,y,z), size (l,w,h), and yaw.
Only one semantic class is used: Car.