How We Work
The method behind every Abelling project: four stages, agreed with you up front, with review built in rather than added at the end.
Scope and taxonomy
Before production starts we agree the classes, attributes and edge cases with you, and write them into guidelines that every annotator and reviewer works from.
Where a rule is ambiguous, such as a partly occluded person, a hybrid material or a label that could belong to two classes, we raise it and record the decision, so the same case is handled the same way across the whole dataset.
Pilot and calibration
Where a project calls for it, a first batch is used to calibrate annotators against your guidelines and to settle open questions about edge cases before volume increases. What we learn from that batch goes back into the guidelines, so production starts from a shared, tested understanding of the task.
Production and review
Trained annotators work through the volume while reviewers check their output. On many projects this is two-tier review, so labels are checked before they leave the production team. Where a project needs it, dedicated QA and senior sign-off are added on top.
Acceptance and export
Completed work is checked against the acceptance criteria agreed at the start, then exported in the agreed format and handed back by the transfer method agreed with you. Long-running programmes repeat this cycle batch by batch.
Tools and Delivery
We work in established annotation platforms such as Encord, Roboflow, CVAT, Supervisely and Darwin (V7), in your own tooling, and with MCAP data as used with Foxglove. See Tools and Delivery for detail.
Request a Project Quote
Now that you know the method, tell us about your project. We will map it onto these four stages and propose a scope.