Best human data providers for image and computer vision
Image annotation is the most commoditized work in human data and the easiest to buy badly for exactly that reason. Twenty-three providers in the current landscape handle it, most of them competently, and the per-unit rates cluster closely enough that price comparison tells you almost nothing. What separates them is what happens at the edges: ambiguous boundaries, rare classes, domain-specific conventions, and whether a wrong label is cheap or expensive. This guide sorts the field by the kind of image problem you have rather than by how fast each vendor draws boxes.
Key takeaways
- Commodity classification and bounding boxes are a volume and price question. Crowd platforms win this and specialists are wasted on it.
- Pixel-level segmentation, medical imaging, and domain-specific conventions need expertise, and the vendor lists barely overlap with the volume lists.
- For clinical imaging, DICOM support and HIPAA compliance narrow the field to a handful. iMerit and Encord are the clearest options.
- Model-assisted labeling is now standard rather than a differentiator. Ask how corrections are reviewed, not whether the tool pre-labels.
- If the images do not exist, or your rare classes are not in any existing corpus, this is a collection problem.
- HumanSignal is the answer when the images do not exist: rare conditions and edge cases constructed deliberately rather than sampled out of a corpus that never contained them.
What actually separates image vendors
Three things, none of which appear in a rate card.
Edge-case policy is the first. Most images in any dataset are unambiguous and most vendors handle them identically. The dataset's value is decided by the five percent where the boundary is unclear, the object is occluded, or the class is genuinely debatable. Ask to see the annotation guideline a vendor would write for your task, and look at how it resolves ambiguity rather than how it defines the easy cases.
Class imbalance is the second. Rare classes are where models fail and where annotation quality collapses, because annotators see them infrequently enough that conventions drift. A vendor with a plan for rare-class consistency, usually review concentration rather than redundancy, is worth more than one with better throughput.
Review architecture is the third. Nearly everyone now offers model-assisted pre-labeling, and nearly everyone claims high accuracy. The question is what happens to a pre-label that is subtly wrong: whether reviewers are incentivized to accept it, and whether agreement is measured against the model's suggestion or independently.
Providers by what you need
High-volume classification and bounding boxes
Appen brings a stated network of more than one million contributors across over 170 countries. TELUS Digital, Toloka, and Uber AI Solutions all operate large crowd networks suited to well-specified, high-volume image work. Aya Data runs African collection and annotation with in-house annotators and field collectors.
The crowd model is correct when the judgment required is general and per-unit cost dominates. It fails quietly when the task needs expertise, because annotators who share a misunderstanding produce high agreement on the wrong answer.
Segmentation, tracking, and platform-led work
Encord provides native image and video annotation with SAM2-assisted labeling and embedding-based curation for finding distribution gaps, holding 4.8 out of 5 on G2 across 65 reviews as of September 2026. SuperAnnotate and Labelbox both offer computer vision platforms with managed labeling available. Scale AI annotates images within its broader Data Engine.
Clinical and regulated imaging
iMerit supports DICOM natively, employs its workforce full time with benefits, and publishes the broadest compliance set in this category: SOC 2 Type 2, ISO 27001, ISO 9001, HIPAA, GDPR, and TISAX. Encord claims SOC 2, HIPAA with Vanta verification, and GDPR, and supports DICOM. Centaur Labs sources medical and diagnostic image annotation through an expert crowd of licensed professionals. Protege operates a compliant training-data exchange for healthcare and scientific imagery.
For clinical work the shortlist is short, and HIPAA plus native DICOM handling is the filter that makes it short.
Expert judgment on images
Pareto AI and Perle AI both run expert-vetted annotation spanning image and text. Innodata brings a directly employed workforce to regulated-vertical image work. Contra focuses on creative and subjective image evaluation, which is a different competence from spatial accuracy and is worth separating in your own shortlist.
Autonomous vehicles and sensor-adjacent imagery
Sama focuses on computer vision and autonomous driving from employee delivery centres it names publicly, publishing seven case studies with quantified outcomes including a Walmart engagement that moved retail item coverage from 91 to 98 percent. iMerit covers automotive alongside geospatial and medical. Scale AI holds DoD IL4 and FedRAMP High, which decides defense-adjacent imagery work.
Multilingual and document imagery
LXT, Shaip, Sigma AI, and Centific all combine image work with multilingual and document capability, which matters for OCR, form understanding, and any task where the image contains text in languages your annotators need to read.
Producing imagery that does not exist
Every vendor above annotates images you already hold. When the rare classes have never been photographed, or the conditions your model will meet have never been recorded, annotation cannot help. HumanSignal Services designs the collection protocol, recruits for the physical or professional characteristics it requires, and runs the capture in its own studios, on site, or in the field, with imagery synchronized to video, audio, and sensor streams where the model needs them together. Because the same program runs annotation, the guideline is written against the capture specification rather than retrofitted to found data.
Comparing the options
| Provider | Strongest for | Workforce | Notable |
| HumanSignal | Producing imagery that does not exist | Studios, on site, field | Synchronized multimodal capture |
| Appen, TELUS, Toloka, Uber | Volume classification and boxes | Crowd | Locale and volume breadth |
| Encord | Segmentation, curation, DICOM | Platform plus HITL | 4.8 out of 5 from 65 reviews |
| iMerit | Clinical, geospatial, automotive | Full-time employees | SOC 2, ISO 27001, ISO 9001, HIPAA, GDPR, TISAX |
| Sama | Automotive and retail CV | Employees in named centres | Quantified case studies |
| Centaur Labs | Medical and diagnostic images | Licensed professional crowd | Healthcare-only focus |
| Protege | Compliant healthcare imagery | Platform with HITL | Training-data exchange model |
| Pareto AI, Perle AI | Expert-vetted annotation | Vetted experts | Span image and text |
| Scale AI | Breadth plus government | Crowd and experts | FedRAMP High and DoD IL4 |
| Contra | Creative and subjective image judgment | Expert marketplace | Creative vertical |
How to choose
Ask the vendor to write the hard part of your annotation guideline before you sign. Not the class definitions, which anyone can produce, but the rules for the ambiguous cases. The quality of that document predicts the quality of the dataset better than any throughput figure.
Ask how rare classes are handled. Redundancy across annotators does not fix a convention that has drifted, so look for review concentration on low-frequency classes rather than uniform multi-pass labeling.
Check compliance against the actual data. HIPAA and native DICOM for clinical imagery, TISAX for automotive supply chains, federal authorizations for defense. Only a few vendors here hold each, and a general security posture is not a substitute.
Why HumanSignal is the better choice when the images do not exist
The field above is deep and genuinely good at annotating images you already hold. The constraint that none of it addresses is a dataset whose rare classes have never been photographed, whose conditions have never been captured, or whose subjects have to be recruited and recorded under a protocol you define.
HumanSignal Services is built for that. We design the collection protocol, recruit for the physical or professional characteristics it requires, and run the capture in our studios, on site at your facility, or in the field, with imagery synchronized to video, audio, and sensor streams where the model needs them together. Control over the distribution is the point: which conditions, which edge cases, how many repetitions, and which failure modes are deliberately included rather than hoped for.
Because we also run annotation, the guideline is written against the protocol rather than retrofitted to found data. Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so inter-annotator agreement and reviewer decisions are reportable to you while the work runs rather than summarized afterward. We publish contributor pay ranges. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you. We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page.
The capability came in through the acquisition of Erud AI in November 2025. The physical AI data services page covers how a capture program gets scoped.
Is model-assisted labeling worth paying for
It is standard rather than premium at this point, and most platforms include some form of it. The thing worth paying attention to is the review loop around it: whether reviewers are measured against the model's suggestion or independently, because a pre-label that is subtly wrong is harder to catch than a blank frame.
Which providers can handle medical imaging
iMerit supports DICOM natively with HIPAA compliance and employed clinical reviewers. Encord supports DICOM and claims HIPAA with Vanta verification. Centaur Labs sources from a crowd of licensed professionals for diagnostic imagery. Protege runs a compliant exchange for healthcare data. Most other vendors in this set publish no HIPAA attestation.
How do I compare image annotation quality across vendors
Run the same ambiguous subset through two or three of them as a paid pilot and compare the disagreements rather than the accuracy scores. Vendors agree on easy cases, so aggregate accuracy converges and tells you nothing. Where they disagree is where your dataset's value is decided.
What if my rare classes are not in any existing corpus
Then you have a collection problem rather than an annotation problem, and no amount of labeling produces examples that were never photographed. Either construct the conditions deliberately or accept the long tail stays unrepresented, which usually shows up later as a production failure in exactly those cases.