How much does data annotation cost? Pricing explained
By the Annoting team · Updated · 6 min read
Data annotation is usually priced per label (per box, polygon or keypoint set), per image, or per annotator hour. The biggest cost drivers are the annotation type — classification is cheapest, pixel-level segmentation most expensive — and how many objects each image contains. At Annoting, rates are per label, QA review is included, and large projects get volume discounts.
The three pricing models
Providers price annotation in one of three ways, and each shifts risk differently:
- Per label — you pay for each box, polygon or keypoint set. Predictable and tied to output; dense images cost more because they contain more labels.
- Per image — a flat price per image regardless of contents. Simple, but the provider prices in the risk of dense images.
- Per hour — you pay for annotator time. Flexible for exploratory work, but the total is hard to predict and speed is not your incentive to control.
What drives the price
For the same number of images, the bill can vary by more than tenfold. The main factors:
- Annotation type — a class tag takes seconds; a careful polygon or full-frame mask takes minutes.
- Objects per image — a shelf photo with sixty products is sixty labels, not one.
- Guideline complexity — many classes, attributes or fine distinctions slow annotators down.
- Quality requirements — review depth, consensus labeling and gold-standard checks add human time.
- Turnaround — rush delivery means more annotators in parallel and usually a premium.
- Domain expertise — medical or specialist imagery may need trained annotators.
Annoting's current per-label rates
These are our live rates by annotation type, read from the same rate card our quotes use. QA review is included. The final quote is confirmed after we review a sample, because object density decides the label count.
| Annotation type | Rate per label |
|---|---|
| Classification | $0.017 |
| Multi-label classification | $0.028 |
| Object detection | $0.057 |
| Multimodal | $0.068 |
| Keypoints & pose | $0.091 |
| Instance segmentation | $0.14 |
| Semantic segmentation | $0.18 |
| Segmentation | $0.21 |
| Project size | Volume discount |
|---|---|
| 25,000+ labels | 6% |
| 100,000+ labels | 12% |
| 500,000+ labels | 20% |
How to estimate your project
Multiply images by the average number of objects per image to get labels, then by the per-label rate. Label 50–100 representative images yourself or ask for a pilot to get a real objects-per-image figure; guessing this number is where most budgets go wrong.
Example: 10,000 images with an average of 6 objects is 60,000 bounding boxes. Volume discounts apply once a project passes the thresholds in the table below.
How to keep annotation costs down
- Use the simplest annotation type that gives the model what it needs — boxes, not polygons, if you only count.
- Remove near-duplicate frames before labeling, especially from video.
- Lock the class list early; mid-project taxonomy changes force relabeling.
- Invest in the guideline — rework from ambiguous rules costs more than writing them.
- Start with a pilot batch to measure real density and quality before committing the full dataset.
Annoting services mentioned in this guide