Annotation services

Polygon annotation & instance segmentation services

Polygon annotation traces the outline of each object with a sequence of points, so the label follows the object's true shape instead of a rectangle. Each outline is a separate instance with its own class, which makes it training data for instance segmentation models such as YOLO-seg and Mask R-CNN.

Trains
Instance segmentation
From
$0.14 / label
Review
Every job
Formats
COCO JSON

Is it the right label?

When to use polygons & instance segmentation

Instance segmentation — the model outlines every object separately, so you can measure its area, shape and position.

  • You need the object's actual shape or area: defect size, crop coverage, organ boundaries, wound measurement.

  • Objects are irregular, so a box would be mostly background: cracks, cables, people in motion, produce.

  • Objects overlap and must stay separate — two touching cars are two instances, not one blob.

When something else fits better

If you only need to count or locate objects, polygons cost more than they return; bounding boxes are faster and cheaper. If you need every pixel in the scene classified (road, sky, building), use semantic segmentation.

Semantic segmentation

How we label it

Our process

  1. 01

    Set the edge rule

    Vertex density, how to handle holes, hair and fuzzy edges, and whether occluded parts are outlined. Agreed in the guideline up front.

  2. 02

    AI-assisted outlining

    Interactive segmentation proposes a mask from a click; annotators refine the edge point by point until it follows the object.

  3. 03

    One instance per object

    Overlapping objects get separate polygons with a consistent depth order, so masks never double-count pixels.

  4. 04

    Edge review

    Reviewers zoom in on boundaries — where segmentation quality actually lives — and return anything that cuts corners.

Quality

What our QA checks

Every job is reviewed by a person other than the annotator, on top of automated checks. Work below the agreed threshold goes back for rework.

Mask IoU against reviewer-corrected outlines

Boundary checks at high zoom on thin and irregular objects

Self-intersecting and degenerate polygons caught automatically

Overlap and z-order consistency between touching instances

Pricing & delivery

Rates and formats

AnnotationRate
Instance segmentation$0.14 / label
Segmentation$0.21 / label

QA review is included. Bulk discounts apply to large projects; the final quote is confirmed after we review a sample of your data. Full pricing

Delivered as

COCO JSON (polygons or RLE)YOLO segmentation TXTPNG instance masksCVAT XMLDatumaro

Used for

Defect inspection and measurement · Medical imaging · Agriculture and plant phenotyping · Robotic grasping · Background removal

Polygons & instance segmentation FAQ

Polygons & instance segmentation, answered

What is polygon annotation?

Polygon annotation outlines an object with a series of connected points that follow its edge. It captures the object's real shape and is used to train instance segmentation models.

What is the difference between polygon annotation and instance segmentation?

Instance segmentation is the task; polygons are one way to label it. A polygon stores the outline as points, a mask stores it as pixels. We can deliver either, and convert one to the other on export.

Is polygon annotation better than bounding boxes?

It is more precise, not always better. Polygons are worth it when the shape or area matters or objects are irregular. For counting and locating compact objects, bounding boxes give the model what it needs at lower cost.

How many points should a polygon have?

Enough that the outline stays within a few pixels of the true edge at full resolution — a few points for a box-like object, dozens for a leaf or a crack. We set the tolerance in the guideline rather than a fixed point count.

Can you train YOLO segmentation models on your polygons?

Yes. We export YOLO segmentation TXT (normalised polygon points per instance) as well as COCO JSON, which most segmentation frameworks read directly.

Need polygons & instance segmentation on your data?

Send us a few sample images or clips. We'll agree the guideline with you and send a plan and a per-label quote.