Image annotation service prices are easy to find but difficult to compare.
One provider may advertise $0.03 per bounding box, another $0.10 per image, and a third $6 per hour. All these figures can be accurate, but they describe different units, workloads, and levels of service.
For a computer vision team, what matters is the cost of an accepted, model-ready training dataset after project setup, annotation, quality control, project management, and rework.
We reviewed official pricing pages, calculators, marketplace listings, and service descriptions from more than 30 image annotation service providers, then compared them with our own project and market research.
This guide gives AI teams a usable benchmark before they choose a supplier or plan an internal operation. It stays with 2D image annotation work and the commercial models used to deliver it.
A practical image annotation services pricing benchmark
The table brings together prices currently published by image annotation service providers and wider planning ranges for production work.
Annotation type | Market benchmark | What most affects the final pricing |
Image classification | $0.02-$0.10 per image | Number of classes, attributes, and judgments per image |
Bounding boxes | $0.03-$1.00 per box across the wider market; published prices for simple tasks are concentrated between $0.015 and $0.10 | Object density, attributes, occlusion, and rotated boxes |
Polygon annotation | $0.03-$0.20 per polygon | Vertex count, boundary precision, and object shape |
Keypoints or landmarks | $0.01-$0.20 per point, object, or pose | Points per instance, skeleton rules, and visibility attributes |
Polyline or curve annotation | $0.01 to $0.20 per line or object | Line length, point count, and topology |
Image segmentation | $0.05-$5.00 per mask/label across the wider market; per-image prices often range from $0.20-$3.00 | Classes, instances, resolution, coverage, and edge rules |
OCR-related image labeling | Around $0.50 per image | Layout, language, transcription, reading order, and tables |
Hourly managed annotation | $3-$60 per hour across the wider service market; general managed teams commonly publish $3-$12 | Throughput, reviewer time, domain expertise, and location |
These ranges are planning benchmarks. A clean product image with one obvious object is a different job from a crowded shelf, a medical scan, or a street scene with dozens of partially occluded objects.
Image dataset volume, accuracy targets, delivery speed, and project complexity determine where a real quote falls within or beyond the range.
What does the benchmark mean in practice?
The price unit tells you how the workload is being measured.
Per-label pricing works well when object counts vary from image to image. Per-file pricing is easier to budget when each image contains a similar amount of work. Hourly pricing fits exploratory projects, changing annotation guidelines, and specialist review.
Published prices for simple, standardized work are concentrated near the lower end of the planning ranges. The wider bands account for dense scenes, domain knowledge, higher quality standards, and other requirements that appear in production projects.
Image classification: inexpensive until the decision becomes complex
Simple classification has some of the lowest prices in image annotation.
Intellabel starts at $0.02 per label. Mindkosh lists $0.03, YAZAKI also starts at $0.03, and US-DATA lists $0.06 per image. These figures sit within a practical $0.03-$0.10 per-image planning range.
The word “classification” can hide a large difference in effort. Specialist provider XANNOTATION lists $0.05 per image for a simple NSFW/SFW decision, $0.35 for complex multi-label classification, and $0.75 for an extreme or custom tier. One binary decision and a multi-attribute review should be budgeted as different tasks.
Roboflow Outsource Labeling starts classification at $0.05 per annotation and requires an active Roboflow subscription. In this model, the service rate and software cost are included in the same budget.
Bounding boxes: the clearest per-label market
Bounding boxes are relatively easy to benchmark because many suppliers charge per object. Tech AI publishes $0.015-$0.035 per object, Label Your Data starts at $0.02, and DeeLab lists $0.02-$0.07 per box. Higher published prices include $0.08 from US-DATA and $0.10 from Roboflow.
Image count alone is not enough for budgeting. A 10,000-image dataset with one object per image contains 10,000 boxes. The same number of street images with 20 objects each contains 200,000 boxes. Per-label pricing connects the invoice to that difference in workload.
Polygons and masks: the boundary drives the workload
Tech AI publishes a common polygon range of $0.03-$0.10 per object. DeeLab lists $0.05-$0.20 per polygon, US-DATA lists $0.15, and Roboflow starts at $0.20. Mindkosh prices polygons with up to 8 points at $0.06.
A five-point outline around a product and a precise contour around a leaf may both be called polygons, but the second task requires far more decisions. The expected vertex count and boundary standard should be part of the quote.
The same principle applies to brush masks and instance segmentation. Foreground coverage, small objects, holes, boundary tolerance, and the number of classes determine completion and review time. A representative pilot gives a better price than an image count alone.
Keypoints: normalize the price to a complete instance
DeeLab lists $0.01-$0.05 per keypoint, Tech AI lists $0.012-$0.020, Label Your Data starts at $0.015 per object, and HumanAIze lists $0.20 per unit. These prices use different workload definitions, so the comparison should be normalized to points per person, face, hand, or other instance.
For example, a 17-point human pose is priced as 17 points under a per-keypoint model. Skeleton connectivity, hidden-point rules, visibility states, attributes, and review layers also affect the final quote.
Image segmentation: the widest production range
Segmentation varies more than bounding boxes because one image can contain very little foreground or hundreds of detailed regions.
A practical planning range is $0.05-$5.00 per mask or complex label. High-precision medical or scientific segmentation can reach $2-$8 per image because specialist review and fine boundaries add substantial work.
A project with several instances per image should use the expected mask count, then add any class attributes and quality requirements.
Hourly annotation services show labor cost than dataset cost
Some providers sell managed capacity by the hour.
Label Your Data says projects are typically $6 per annotator-hour. AnnotationBox publishes $5-$7, Precise BPO advertises $5-$9, and DeeLab lists $8-$12.
Across the broader ML image annotation service market, hourly rates run from $3-$60, including specialist and onshore work. Official AWS Marketplace listings show this geographic and service-level spread. Pripton lists $5 per hour, Cogito lists $5.40, and iMerit lists $6.48 per person-hour. US-based workforce prices reach $24.84 per hour from Quadrant and $28.80 from iMerit.
Hourly prices still need a productivity measure. At $6 per hour, 300 accepted bounding boxes per hour produce a direct labor cost of $0.02 per accepted box. At the same hourly rate and 60 accepted boxes per hour, the direct cost becomes $0.10. Review, management, and correction terms determine the full delivered cost.
Crowdsourcing and managed services are priced differently
We covered the operational differences in our guide to crowdsourcing and managed annotation services. The pricing difference follows the division of responsibility.
On Amazon Mechanical Turk, the requester sets the worker reward and Amazon adds a 20% fee to rewards and bonuses. The full crowd cost includes worker rewards, platform fees, redundant judgments, qualification tasks, QA, adjudication, and rework.
Toloka uses a composed price based on the task or expert service, QA components, and platform fees. Crowd models can work well for simple tasks that are easy to explain and audit.
A fully managed annotation service takes responsibility for training, staffing, quality control, and delivery. This reduces the operating work left with the buyer and makes it easier to run specialist or security-sensitive projects. The service may charge per annotation, per file, per hour, or through a project quote.
Large enterprise providers commonly use custom quoting rather than fixed labor prices. Scale AI, CloudFactory, Sama, Appen, TELUS Digital, LXT, DataForce, Innodata, TaskUs, Centific, Dataloop, Kili Technology, Encord, and SuperAnnotate all sell services through scoped enterprise engagements. CVAT Annotation Services publishes a $5,000 minimum project budget, while its pricing explainer uses $0.10 per object as an illustrative calculation.
Custom quoting reflects the actual delivery package: dataset review, guideline development, staffing, security, quality targets, and schedule. It should be compared through the same pilot and acceptance test, rather than treated as a missing market signal.
What changes the final image annotation project price?
Image annotation type and boundary precision
Classification is usually faster than drawing geometry. Axis-aligned bounding boxes are usually faster than rotated boxes or cuboids. Polygons take longer as vertex counts rise, while brush masks and pixel-accurate segmentation require more boundary decisions.
Object density and scene difficulty
A bounding box quote needs an expected object-count distribution. Crowded shelves, traffic scenes, overlapping cells, small defects, and partially occluded objects take longer than clean product images. High resolution can increase both object discovery and boundary work.
Ontology, attributes, and edge cases
A single class with a clear inclusion rule is a different project from a hierarchical ontology with attributes, ignored regions, uncertain cases, and negative examples. A paid or free pilot should include the hardest samples, so the quote reflects production work.
Quality target and acceptance method
“QA included” should be translated into a sampling rate, reviewer structure, agreement metric, correction policy, and acceptance threshold. Our guide to quality metrics for computer vision annotation explains why one headline accuracy number rarely describes the whole dataset.
Workforce location and domain expertise
Workforce location changes hourly cost. Domain knowledge can matter more than geography in medical imaging, industrial inspection, agriculture, geospatial imagery, and regulated content. Specialist labor may cost more per hour while reducing guideline changes and rejected labels.
Turnaround, ramp size, and team continuity
Urgent work may require parallel teams, weekend shifts, or reserved capacity. Short projects spread training and setup across fewer labels. Long-running programs may receive volume pricing but need stable staffing and change control.
Security and compliance
Restricted workspaces, background checks, device controls, regional data residency, audit logs, and contractual requirements add operating cost. These requirements can also make open crowdsourcing unsuitable regardless of its nominal rate.
Model assistance and the review process
Pre-labeling can reduce drawing time when model suggestions are accurate. Weak predictions can increase correction work and create anchoring errors. A quote should state how automation is used, how low-confidence cases are routed, and whether human verification to the agreed standard is included.
In-house annotation needs its own total-cost model
An internal team cannot be compared with a supplier’s per-object price using wages alone. A fair internal model can be:
In-house cost = loaded annotator labor + recruiting and training + guideline maintenance + project management + QA + software and infrastructure + security + engineering support + idle or ramp capacity
Accepted throughput is the relevant measure. If an employee with a loaded cost of $25 per hour produces 250 accepted boxes per hour, direct labor is $0.10 per accepted box before management, software, and rework. If a difficult ontology reduces accepted output to 80 boxes per hour, direct labor becomes $0.3125 per box.
What should you ask an image annotation service provider?
A short scope document prevents most pricing confusion. We suggest you ask image annotation service providers these questions:
What is the billable unit: image, object, polygon, point, attribute, mask, page, task, or hour?
How many objects, vertices, classes, attributes, or keypoints are included in that unit?
Are guideline development, onboarding, project management, platform access, export, and storage included?
Which quality-control steps are included, and which acceptance metric will be reported?
Is correction of rejected work included? Is there a limit or acceptance window?
Is there a project minimum, tool subscription requirement, setup fee, or reserved-capacity charge?
Which workforce location and experience level will handle the data?
Which security, compliance, and data-residency controls are included?
Which turnaround and weekly throughput does the quote assume?
Can the provider run a pilot project with dense, ambiguous, and rare cases?
Which tools will be used for your projects? Do they have their own annotation platform?
Give every shortlisted provider the same representative sample and acceptance test. Compare total cost and quality per accepted image or object after the pilot. That figure is more useful than the lowest rate on a pricing page.
How does BasicAI price image annotation services?
At BasicAI, we use three common pricing models: per annotation, per file, and per hour. Annotation-based pricing is the most common because it connects the invoice to the actual number of labels produced.
Some of our reference rates are $0.03 per bounding box, $0.05 per polygon or mask, and $0.02 per keypoint. Per-file pricing is set after a pilot because the amount of work inside one image can vary widely. Our hourly reference is $3-$5, with the final rate also confirmed after the pilot.
Accuracy targets, throughput and turnaround requirements, project complexity, domain knowledge, and the quality process can move the final price. We start with a free pilot to measure the actual workload, confirm the ontology and acceptance standard, and provide a project-specific quote.
Our image annotation service is fully managed and non-crowdsourced. It combines selected annotation teams, a dedicated project manager, domain resources, multi-stage quality control, and our own annotation platform.
For each project, we target 99%+ image annotation quality. Reasonable post-annotation corrections are included within the project scope.
Teams comparing suppliers or evaluating an internal operation can request a free pilot and project estimate. Using the same difficult sample across every option produces a realistic cost and quality benchmark before a larger commitment.
Notes
This benchmark focuses on human image annotation services. It excludes standalone annotation software pricing. We prioritized current official pricing pages, calculators, help centers, service descriptions, official marketplace listings, and the first-party BasicAI rates supplied for this update. This is a broad global pricing guide, not a census of every provider. Public rates change, and starting prices do not represent every production project. Buyers should date the sources they use and make the final decision through a shared pilot, a clear acceptance test, and total cost per accepted deliverable.
