Data labeling
Types of data annotation
Each annotation type carries a different cost and a different capability. Choosing wrong makes the dataset expensive without improving the model — or caps the model before it exists.
Describe what the model needs to learn. We'll point to the right annotation type.
The annotation types
From the cheapest label to the most expensive contour, and what each one enables.
-
Classification
One label per image or crop. It answers “what is this”, not “where is it”. The cheapest format, and enough when position doesn't matter.
-
Bounding boxes
A rectangle around each object. Gives approximate location at low cost, and is the format detectors like YOLO and Faster R-CNN expect.
-
Polygons
A contour fitted to the object. Worth it when a rectangle would capture too much background — irregular parts, bulk products, objects on a diagonal.
-
Semantic segmentation
Every pixel gets a class, without separating instances. Suited to surfaces and regions: stains, corrosion, coverage area.
-
Instance segmentation
Each object gets its own mask. Needed when objects of the same class touch and have to be counted or measured separately.
-
Keypoints
Reference points on the object. Used for orientation, pose, articulation and measuring distances between parts.
-
Video tracking
The same object keeps its identity across frames. That's what allows counting without duplicates and trajectory analysis.
-
OCR and text reading
Extraction of codes, batches, expiry dates or labels from the image, with verification of what was read.
-
Object relationships
Recording the link between elements: which label belongs to which package, which part sits in which fixture position.
How to choose
Four criteria that stop you paying for precision the model never uses.
-
Start from the model's decision
If the system only has to pass or fail, classification is enough. If it has to point at where, add localization. If it has to measure, add masks.
-
Cost rises with the contour
A mask can cost several times what a box costs. That premium is only justified when the contour changes the outcome.
-
Formats can coexist
One dataset can use boxes for most of the material and masks for a subset, when only some cases need detail.
-
Deciding early avoids relabeling
Switching format later usually means redoing the work. This is the discussion worth having before the first batch.
Export formats
Delivery adapts to your pipeline, not the other way around.
-
COCO
JSON with images, categories and annotations. Covers boxes, polygons and keypoints.
-
YOLO
One text file per image, with normalized coordinates. Direct input for the YOLO family.
-
Pascal VOC
XML per image, still used by older pipelines and tools.
-
Masks and custom schemas
PNG per class, or whatever schema your team already uses internally.
Frequently asked questions about annotation types
What ML teams ask when settling the dataset format.
Which annotation type is most common?
Bounding boxes, because they balance cost and usefulness. But the most common isn't always the right one — the choice depends on what the model has to decide.
Do you handle video annotation?
Yes, including identity tracking across frames, which is what allows counting without duplicates and trajectory analysis.
Can we mix types in one project?
Yes. It's common to annotate most of the material with boxes and reserve masks for the subset that needs a precise contour.
Do you use our tool or yours?
We work with whatever the project requires. If you already have a platform and licenses, we operate in it; if not, we use ours and deliver in your format.
Talk to our team
Let's understand your challenge
Prefer to book the diagnosis conversation? Use the scheduling link. Or fill in the form — our technical team replies.