Detection & localization
Locate objects in image frames and 3D scenes.
- 2D bounding boxes
- Localize objects in image and video frames.
- 3D cuboids
- Capture position, dimensions, and heading in point-cloud scenes.
Training outputs for Physical AI
Ground truth shaped for the model.
Turn synchronized camera, LiDAR, and sequence data into the geometry perception systems need: 2D boxes and masks, 3D cuboids and point-cloud labels, polylines, classifications, and tracked objects.
Bring one representative scene. We will map the output, review path, and delivery format before production.
0142
Synchronized scene input
Multi-camera videoLiDAR and point clouds
0087 / 0120
Output contract
Detection & localization
01
Segmentation
02
Scene geometry
03
Tracking & classification
04
MCAP and ROS data
LiDAR and point clouds
Multi-camera video
Telemetry and trajectories
Training outputs for Physical AI
Choose the output your training loop needs.
Select a shipped primitive, then configure its classes, attributes, annotation scope, and review path for the model task.
Segmentation
Create class or object regions in camera and point-cloud data.
- Image masks
- Create pixel-level class or object boundaries in camera data.
- Point cloud segmentation
- Assign semantic classes to points in 3D scenes.
Scene geometry
Trace task-defined boundaries and sequence geometry.
- Polygons & polylines
- Trace lanes, curbs, regions, and other task-defined geometry.
- Sequence polylines
- Carry frame-aware 3D polyline geometry through supported sequences.
Tracking & classification
Preserve identity and attributes across supported sequences.
- Tracked objects
- Maintain object identity across supported sequence frames.
- Classifications
- Add frame-, object-, or track-level labels and attributes.
Define the task
Start with the model decision.
The training objective determines the ground-truth representation, not the other way around. Define the output against your sensors, edge cases, and delivery contract before production work scales.
Keep camera, LiDAR, transforms, and sequence context visible where the workflow supports them.
Set class rules, geometry, attributes, and frame or sequence scope before production.
Route outputs through human review and acceptance states before release; add task-specific consensus analytics where supported.
Connected ground truth
Built for the loop after labeling.
Ground truth becomes valuable when it can move through review and into a versioned training-data workflow.
Preserve scene context
Keep camera, LiDAR, telemetry, and dataset context connected for the labeling task.
Gate on quality
Use human review, attribution, and acceptance states; add task-specific consensus analytics where supported.
Release training data
Move approved outputs into versioned dataset workflows for downstream export and model work.
Scope the work
Bring us one hard scene.
We will map the training objective to the ground-truth geometry, workflow, and delivery contract.