"How much does annotation cost?" has no single answer, because a tag on an image and a 3D cuboid on a LiDAR frame are completely different amounts of work. What you can know is how providers price, what pushes the price up, and how to estimate a budget before you ask for quotes.
The four common pricing models
| Model | How it works | Best for | Risk |
|---|---|---|---|
| Per item or per label | A price per image, per object, per minute of audio or per frame | Clear, stable tasks with predictable effort | Hard to price fairly when items vary a lot |
| Per hour | You pay for annotator time | Exploratory or changing tasks | Cost depends on the team's speed; you need visibility |
| Dedicated team (FTE) | A monthly fee per full-time annotator or a whole pod | Long, continuous programmes | You pay for capacity even in quiet weeks |
| Fixed project price | One price for an agreed scope, volume and quality bar | Well-defined datasets | Scope changes need a new quote |
What drives the price
- Data type. Text and image tags are quick; video and LiDAR take much longer per item.
- Label type. Roughly from cheapest to dearest: classification, bounding boxes, polylines and keypoints, polygons, segmentation masks, 3D cuboids and point cloud segmentation.
- Objects per item. A street scene with 40 objects costs far more than a product photo with one.
- Guideline complexity. Many classes, attributes and edge-case rules slow people down.
- Domain expertise. Medical, legal or engineering data may need trained specialists.
- QA level. Reviewing 100% of labels costs more than reviewing a random sample.
- Turnaround. Rush deadlines need a bigger team and more coordination.
- Security. Locked-down environments, on-premises work or restricted data handling add overhead.
A worked estimate (illustrative)
Price is mostly time. Estimate the hours, then apply the rate a provider quotes you. For example, using assumed numbers only:
| Step | Assumption | Result |
|---|---|---|
| Images | 10,000 | |
| Objects per image | 8 boxes on average | 80,000 boxes |
| Time per box | 6 seconds with AI pre-labelling | about 133 labelling hours |
| QA overhead | +25% for review and rework | about 167 total hours |
Multiply the total hours by the hourly rate, or divide by the number of items to compare against a per-image quote. Then replace the assumptions with real numbers from a pilot: time per object is the figure that changes your budget most.
How to lower cost without losing quality
- Write clear guidelines with examples. Confusion is the most expensive thing in annotation.
- Use the simplest label that works. Do not pay for masks if boxes train a good enough model.
- Use AI pre-labelling and have people correct it.
- Sample your QA once quality is stable; keep full review for rare or safety-critical classes.
- Prioritise data. Label the hard, diverse examples first rather than thousands of near-duplicates.
- Start with a pilot so the quote is based on your real data, not a guess.
Questions to ask when you compare quotes
- Is QA included, and what share of work is reviewed?
- Is rework of failed batches free?
- How are guideline changes priced mid-project?
- Which accuracy measure is promised, against what reference?
- Are tool licences, storage and export formats included?
How Annova prices annotation
We start with a free pilot (up to 100 images, 10 minutes of video or 1 hour of audio), measure the real effort on your data, then send a fixed quote based on data type, volume, label complexity, QA level and turnaround. See pricing or request a pilot.
