AP-01 · SPECIFICATION
Defined together
Label definitions, data selection, coordinate conventions, and acceptance criteria are written down before labeling starts, with worked examples for the hard calls.
Aperture
Powering Physical AI
( 01 ) APPROACH
We label to your specification, then prove it on your data.
Aperture creates and reviews annotations for autonomous vehicle perception. We adapt the labeling process and delivery format to your data and engineering requirements, so the labels slot into your training and evaluation workflow.
We can work from your existing guidelines or develop a specification with you from representative recordings. Scope, preparation, quality targets, and delivery are agreed with your team.
AP-02 · PILOT
Proven on a pilot
A small, representative set of your scenes runs through annotation, review, and export first, so your team can inspect complete sequences against your sensor data.
AP-03 · DELIVERY
Versioned batches
Reviewed annotations arrive in agreed batches with source references, labeling status, and a record of the quality checks behind each release.
( 02 ) WHAT WE LABEL
Objects in 3D, tracked across time.
For 3D detection and tracking, each label carries an object's position, dimensions, orientation, class, and identity over time, with attributes where your task needs them.
LB-01 · CUBOID
3D boxes
Center, length, width, height, and heading in the vehicle frame, to the box conventions in your specification.
LB-02 · TRACK
Track IDs
The same object keeps one identifier across frames, including through brief occlusion.
LB-03 · ATTRIBUTE
Classes and attributes
Your taxonomy, with attributes such as motion state and visibility applied by agreed evidence rules.
LB-04 · FUSION
Camera and LiDAR together
Images and point clouds are read side by side, aligned with your calibration and vehicle pose.
LB-05 · EDGE CASE
Sparse and occluded objects
Rain, distance, and occlusion are handled by your rules, and open questions come back to you with examples.
LB-06 · REVIEW
Reviewed before release
Every batch is checked against your criteria, and the report says what was checked and what is still open.
Figure data: Boreas dataset, University of Toronto, CC BY 4.0. Points downsampled and 3D boxes added by Aperture.