There are several places to put a dot on a fish. The eye seems straightforward. The start of the tail invites discussion. The center of the body requires a rule. For researchers at King Abdullah University of Science and Technology, these small judgments were part of a larger problem: how to produce consistent examples for a computer vision model studying fish. Somebody had to turn anatomy into instructions that different people could follow.
- Label Your Data supplies annotation teams and reviewers for AI datasets.
- Its distinctive pitch: work inside the customer’s tools, test a sample, then scale.
- A self-serve platform handles simpler computer vision orders; complex projects get a managed pilot.
Seven points, one fish
According to Label Your Data’s account, KAUST had used Amazon Mechanical Turk but wanted more consistent annotations. Labels needed extra checking, and quality problems affected research schedules. The replacement workflow used CVAT, written guidelines, reference images, and a short demonstration video. Twenty annotators marked seven landmarks per fish. Five senior reviewers checked their work.
The company reports meeting the project’s 98%+ accuracy requirement in under a month. That figure belongs to this assignment, with its particular rules and review process. The more transferable detail is the separation of production from inspection. A dot became useful when someone else could check why it belonged there.
A services business with a software door
Founded in 2020, Label Your Data grew out of SupportYourApp’s outsourcing business. Its official team page identifies Daria Leshchenko as founder and partner/advisor, and Karyna Naminas as CEO. The background matters: arranging people, instructions, schedules, and customer feedback is much of the product.
Customers buy labeled images, video, text, audio, and spatial data, along with quality assurance and model validation. The company reports more than 200 customers. Its work sits between a team’s raw dataset and the model that needs usable examples. Research groups and engineering teams hire it when preparing those examples consumes time they need elsewhere.
The quarry that looks like a landfill
A project for researchers at Yale University and the Paris School of Economics makes the judgment problem clearer. They were developing models to identify landfills in satellite imagery. From above, a quarry, a mine, or an animal feeding operation can resemble a waste site. Cloud cover can remove the evidence altogether.
Label Your Data describes two annotators and a QA specialist checking coordinates and assigning confidence. Unclear imagery received lower confidence instead of forced certainty. The work covered 10,400 coordinates across 16 locations. Its value was partly in preserving the difference between “this is a landfill” and “this photograph does not tell us enough.”
The company reports improved model scores after this work. The practical lesson is narrower: an annotation scheme needs somewhere to put uncertainty. Otherwise, an honest doubt can become a confidently wrong training example.
“The biggest benefit has been peace of mind around annotation quality.”Faizan Khan · KAUST researcher, in the company’s case study
Bring the team to the tools
AirSeed, an Australian reforestation business, needed maps that could help models identify planting zones for drones. Its engineers lacked capacity for laborious GIS labeling. Label Your Data worked inside QGIS, with five annotators and a reviewer, over five months. Weekly conversations settled difficult shapes. The company reports delivering more than 30 annotated geofiles.
NODAR’s problem was different: public datasets did not cover the scenes and sensor-specific details its depth-perception models required. The annotation team used CVAT alongside NODAR’s own tools, refined instructions after pilot tasks, and logged edge cases. The published account describes roughly 60,000 polygon masks across datasets.
Together, these projects explain the positioning. Label Your Data sells a managed operation that can enter an existing workflow. Buyers can also consider providers such as iMerit or Sama, or run their own workforce with CVAT or Label Studio. The useful comparison is who will train the annotators, settle disagreements, and own delivery. Keeping that responsibility explicit makes vendor comparisons more useful than a list of annotation shapes. Two suppliers may both offer polygons while leaving very different amounts of coordination to the customer.
- 01 Sample
- 02 Agree rules
- 03 Annotate
- 04 Review
- 05 Deliver
The price of a decision
The public pricing calculator gives an indication of the economics. At the time of review, its inputs estimated bounding boxes at $0.02 each and polygons at $0.045. On that arithmetic, 10,000 boxes would be $200; 10,000 polygons, $450. These are calculator examples. A representative sample is needed to establish a project’s actual price.
The company offers hourly, per-task, per-object, and project arrangements. A no-cost pilot tests complexity and delivery before production volume. That gives buyers something concrete to inspect: how much checking their own engineers must still do.
Start with the awkward examples
The self-serve platform, launched on Product Hunt in 2024, puts an ordering interface around the service. Customers upload computer vision data, specify instructions, and retrieve results. Features include an instruction template, API access, team permissions, and a free task for ten images or frames. The help center recommends this route for simple datasets, and a managed pilot for sensitive work requiring agreements and closer discussion.
For a buyer, the method is easy to borrow. Put ambiguous examples into the trial. Define what an acceptable label looks like. Check the returned format in the actual training pipeline. Record disagreements before ordering another batch. A trial should also reveal whether the buyer and reviewer interpret the same instruction in the same way. If they disagree, the small batch has already done useful work: it has exposed a problem while it is still cheap to correct.
This approach needs a customer who can answer domain questions and provide representative data. Reviewers cannot recover a detail hidden behind a cloud. A larger labeling team cannot resolve a definition the buyer keeps changing. The fish needs an agreed tail before anybody starts multiplying the dots.