Driving-scene data that holds together in 2D, 3D and time.

Perception models learn from camera frames and LiDAR sweeps that have to agree with each other. We produce and check the boxes, cuboids, lanes and tracks that keep them consistent across sensors and across a sequence.

Data problem Inputs and tasks Methods Validation Example

The data problem.

A driving dataset is synchronised sensors over time, and a label is only correct if it is correct in all of them.

Cross-sensor agreement. A cuboid in the point cloud must project onto the same car in every camera.
Temporal consistency. IDs, sizes and headings should not jump when the object did not.
Sparse returns. Distant or occluded road users may be a handful of LiDAR points.
Ambiguous road structure. Worn markings, merges and roadworks make lanes a judgment, not a trace.
The long tail. Rare road users, night, rain and glare.

What we label.

Inputs

Front and surround camera video
LiDAR sweeps as point clouds
Sensor calibration and ego pose, where provided
Drone or overhead footage of road scenes

Tasks

2D boxes on vehicles, pedestrians and cyclists
3D cuboids with position, size and heading
Lane lines and road edges as polylines
Drivable area and road-surface segmentation
Multi-object tracking across frames and sweeps

How it is checked.

Approaches we use or evaluate: ground removal, 2D and 3D detectors that propose labels, keyframe interpolation, and cuboids projected into the cameras to flag mismatches. A reviewer then checks:

Cuboids: tight to the points, correct heading, consistent dimensions over time
Tracks: no ID swaps at crossings, merges or occlusions
Road: lane topology and drivable area, set by the guidelines

LiDAR · 3D cuboids · objects 5 · drivable area

A LiDAR sweep at an intersection.

A city point cloud: cuboids on parked and moving cars, their distances, and the drivable area.

Point AI’s role.

You define what the model needs. We run the operation that produces it.

Calibration on hard, representative sequences to agree classes and tolerances
Managed operations by an employed, trained team that stays on the project
Delivery in Point AI Studio: review in the browser, export or API

Have a challenging data problem?

Tell us what you’re building and what data you have. We’ll suggest a practical way to process, check and deliver it.

Discuss your project