Bounding boxes, polygons, segmentation, keypoints and object tracking, labeled to your written guidelines and checked against a gold set before delivery.
Pick one or combine several in the same project. We follow your class list, attributes and edge-case rules exactly.
Rectangles around each object, with the class and any attributes you need.
Used for: Object detection, counting, retail shelves
Point-by-point outlines that follow the real shape of an object.
Used for: Irregular objects, damage, rooftops, produce
Every pixel assigned to a class such as road, sky, person or vehicle.
Used for: Scene understanding, drivable area, land use
Pixel masks that also separate one object from the next of the same class.
Used for: Crowded scenes, cell counting, picking robots
Landmarks such as body joints, facial points or product corners, in a fixed order.
Used for: Pose estimation, gesture, sports, ergonomics
Lines along lanes, curbs, wires, pipes and other thin structures.
Used for: ADAS, road mapping, infrastructure
Image-level labels and attributes, single or multi-label.
Used for: Content moderation, search, quality checks
Boxes with depth on 2D images, showing an object’s orientation in space.
Used for: Vehicles, warehouse robots, AR
Video adds time. The hard part is keeping labels consistent from frame to frame, so every project gets written rules for occlusion, re-entry and ID changes.
The same ID follows each object across every frame, through occlusion and re-entry.
Annotators label keyframes and correct the in-between frames, so long clips stay consistent.
Start and end times for actions such as a fall, a lane change or a pick from a shelf.
Labels for whole segments of a clip: scene type, weather, activity, or safety status.
More precise labels take longer and cost more. Use the cheapest type your model can actually learn from.
| Type | Shape precision | Effort per object | Best when |
|---|---|---|---|
| Classification | None | Lowest | You only need to know what is in the image |
| Bounding box | Low | Low | You need where an object is, not its exact outline |
| Keypoints | Point-level | Medium | Pose, landmarks or alignment matter |
| Polygon | High | Medium | Shapes are irregular and boxes include too much background |
| Semantic / instance segmentation | Pixel-level | Highest | Every pixel counts, such as drivable area or medical regions |
You judge us on real output before anything scales.
A sample of images or clips, your class list, and what the model needs to do.
We write up classes, attributes and edge cases with examples, and you sign them off.
A small batch, reviewed by a second annotator and scored against your gold set.
You review the pilot. We update the guidelines and fix what you flag.
Full delivery in batches, each one audited with the same QA report.
Acceptance criteria are agreed in writing before the pilot, so "done" means the same thing to both teams.
We can work in your annotation tool, or set up one of ours and deliver in the format your training pipeline reads.
Your images are your product. We treat them that way.
The annotation types above, applied to the problem your model is solving.
Vehicles, pedestrians, lanes, signs and drivable area, with tracking across frames.
Products on shelves, packaging, attributes and visual search.
Defects, scratches, missing parts and assembly checks.
Crops, weeds, fruit counting and plant health from field or drone images.
Objects to pick, grasp points, 3D cuboids and human pose for safety.
Rooftops, roads, vehicles and land use from aerial and satellite images.
If yours isn't here, ask in the form below.
Per object, per image or per hour, depending on the annotation type and how complex your images are. We quote after seeing a sample, and confirm the rate after the pilot, when both sides know the real effort per item.
Usually large enough to cover your main classes and known edge cases. We agree the size, deadline and acceptance criteria with you before starting.
Yes. We can work inside your platform with accounts you control, or set up CVAT or Label Studio and deliver in your format.
They are written into the guidelines with picture examples before labeling starts. New cases found during work are logged, and we ask you for a rule instead of guessing.
It depends on volume, annotation type and image complexity. After the pilot we know the real time per item and commit to a delivery schedule in writing.
Yes. See text and NLP, audio and speech, LLM evaluation and data collection.
Tell us about your data and the model you're building. We'll reply with questions and a practical pilot plan.
Prefer email? Write to hello@norbeon.com.
Run an annotation team? Register as a vendor.
Analytics cookies
We use Google Analytics and Microsoft Clarity to see how visitors use this site. No advertising cookies. You can change this anytime from Cookie settings in the footer. Privacy Policy