Image annotation explained: boxes, polygons, keypoints and labels
· a.pannella87
Last updated: 27 October 2026 · OpenJob editorial team
Image annotation uses four main task types: bounding boxes around objects, polygons or segmentation that trace outlines, keypoints placed on specific parts such as joints, and classification with one label per image. Each rewards a different habit, and in every case the project guide decides over your own judgement.
Key points
- Boxes should usually touch the object’s outermost visible pixels.
- Keypoints follow the subject’s left and right, not the viewer’s.
- Flag ambiguous images instead of forcing a label.
- A large screen and a mouse make polygon work faster and more precise.
Image annotation is often described as “drawing boxes”, which undersells it. Most projects use one of four task types, and each rewards a slightly different habit.
Bounding boxes
A rectangle around each object of a given class. The rule that trips people up is tightness: guides usually want the box to touch the object’s outermost visible pixels, not leave a margin. Partly hidden objects have their own rule, often “box only the visible part”.
Polygons and segmentation
Here you trace the outline. It is slower and tests patience; the quality check is whether edges follow the object rather than cutting corners.
Keypoints
Points placed on specific parts, such as joints on a human figure. Consistency is everything: the “left wrist” must always be the person’s left, not the viewer’s.
Classification
One label for the whole image. Quick per item, but ambiguous images pile up. Good annotators flag them instead of forcing a choice.
Why the guide wins over your judgement
Annotation data trains models by repetition, so a label applied differently by two people adds noise. If the guide says a bicycle leaning on a wall is “parked”, label it parked even if you would call it abandoned. Raise disagreements through the project’s question channel, not in the labels.
Equipment
A large screen and a mouse make polygon work far less tiring than a laptop trackpad. Projects rarely require them, but reviewers notice the difference in edge quality.
Visual-data roles on OpenJob currently lean towards video, such as the Video Annotation Specialist. The same habits apply frame by frame.
Open roles that do not ask for previous experience
Refreshed automatically from our verified catalogue. Check eligible countries and requirements on each listing.
Corporate Attorney (BigLaw Firms)
Remote contract work for corporate attorneys from large law firms on an AI training project: analyze and draft complex agreements and legal instruments in corporate law, M&A, private equity and cap…
Finance Specialist
Remote contract for finance professionals on an executive-presentation project: rebuild and improve PowerPoint slides that present complex financial data and insights for investors and executives…
Bilingual Japanese Psychologist (PhD)
Remote contract for bilingual psychologists: assess psychological case studies and documentation and develop culturally sensitive scenarios and materials in Japanese and English for AI training.
Simulation Expert — Real-to-Sim 3D Environments & Assets
Remote contract work for simulation experts on an AI training project: build 3D assets and environments in Isaac Sim, Gazebo, MuJoCo or similar, and tune physical parameters so they match real-worl…
Japanese Language Expert
What if your mastery of Japanese could directly influence how AI understands, speaks, and connects with millions of people?
Data collection roles · All open roles
Frequently asked questions
What is the difference between a bounding box and a polygon?
A bounding box is a rectangle around an object; a polygon traces its actual outline. Polygons and segmentation are slower and are checked on whether edges follow the object closely. Boxes are checked mainly on tightness and on the rule for partly hidden objects.
Why must I follow the guide even when I disagree?
Annotation data trains models by repetition, so the same item labelled differently by two people adds noise. Consistency across the team matters more than the best individual choice. Raise disagreements in the project’s question channel, not in the labels themselves.
Is image annotation work available now?
Visual-data listings change often; at the moment many lean towards video, where the same habits apply frame by frame. Check the current roles in the data collection and annotation categories, and read each listing’s equipment and country requirements.