The box is the easy part
Bounding box, polygon, polyline, keypoint: the tools take an afternoon to learn. Judgement takes longer. Where exactly does the object end? Is a half-hidden wheel part of the car? Every dataset has edge cases, and every edge case is a decision.
Two years of freelance annotation taught me that the drawing is the smallest part of the job. The decisions around it — what counts, what does not, what to do when the rule is unclear — are the work.
Consistency is the product
An ML team can work with any consistent rule, but not with five different ones. If the same object is labelled three ways across a dataset, the model learns noise. That is why edge cases need written rules, not memory: a short guideline, agreed before labelling starts, beats a thousand careful guesses.
Consistency also means honesty about limits. Some images are ambiguous, and labelling them confidently anyway does not help anyone. Flagging them is part of the job.
Check your own work
Before delivery, I review my own labels with fresh eyes and against the written rules — a second pass that catches drift, the slow slide where the rule you apply on image 900 is not quite the rule you applied on image 10. Clear acceptance criteria, agreed upfront, make that review possible.
References and further reading
These sources informed the technical framing of this note. Consult the original documentation for full definitions and context.