If you've browsed image annotation job postings, you've seen these three terms constantly: bounding boxes, polygons, and semantic segmentation. They're often used as a checklist of required skills, but rarely explained clearly. Here's exactly what each one involves.
Bounding Boxes — The Simplest Technique
A bounding box is a rectangle drawn around an object in an image to mark where it is. If you're labeling a photo containing a car, you draw a rectangle that fully contains the car — the box doesn't trace the car's actual shape, just its rough boundary.
Computer vision annotation is one of the most learnable AI training skills with the steepest quality-to-pay curve. The difference between a mediocre bounding box and a precise one is visible, measurable, and directly affects your quality score — which determines everything else.
This is the fastest, simplest annotation technique, which is why it's typically the entry point for image annotation work — including the kind described in our beginner's guide to data annotation. Bounding boxes are commonly used to train object detection models — systems that need to identify "there's a car here" without necessarily needing the car's exact outline.
Polygons — More Precision for Irregular Shapes
A polygon traces the actual outline of an object using multiple connected points, rather than a simple rectangle. If you're labeling a person standing with one arm raised, a bounding box would include a lot of empty background space around the irregular shape; a polygon traces the person's actual silhouette more precisely.
Polygons take longer to create than bounding boxes since they require more individual points and more careful tracing, which is reflected in pay — polygon annotation work generally pays somewhat more than basic bounding box work for this reason.
Semantic Segmentation — The Most Detailed Technique
Semantic segmentation goes a step further than polygons: instead of outlining individual objects, you label every single pixel in an image according to category. A street scene might have every pixel categorized as "road," "sidewalk," "car," "pedestrian," or "sky" — leaving no pixel unlabeled.
This is the most time-intensive and detailed of the three techniques, used for applications requiring extremely precise understanding of an image's full content — autonomous vehicle training data is a common example, since a self-driving system genuinely needs to understand every part of its visual field, not just identify a few key objects.
Comparing the Three
| Technique | Precision | Time Required | Common Use Case |
|---|---|---|---|
| Bounding Box | Low — rectangle only | Fast | Basic object detection |
| Polygon | Moderate — traces outline | Moderate | Irregular shapes, precise boundaries |
| Semantic Segmentation | High — every pixel labeled | Slow | Autonomous vehicles, medical imaging |
Why This Terminology Matters for Your Resume
As we cover in our piece on putting AI training work on your resume, using these specific technical terms accurately — rather than a vague "I labeled images" description — is exactly what applicant tracking systems and recruiters in this space search for. If you've done semantic segmentation work, say so specifically; it signals a more advanced skill than generic bounding box labeling.
Where to Find This Type of Work
Several platforms we review include image annotation tasks using these techniques as part of their broader task offerings, alongside text and response evaluation work covered throughout our top platform rankings. The specific technique required varies by project, so it's worth checking project descriptions for which of these three terms applies before assuming all "image annotation" tasks are equivalent in either difficulty or pay.
See our full data annotation jobs guide for the complete pay landscape across all annotation categories.
Ready to Start?
See our top-ranked platforms with image annotation and broader AI training tasks.