In short: the quote for a labeling project depends on six things: the data, the labels, the volume, the quality target, the deadline and the tools. No vendor can name a fair figure before seeing sample data and the labeling rules. This guide shows what moves the quote and what to send to get a firm one.
The six things that set the quote
- The data and the kind of label. A box around an object takes seconds. A pixel-exact outline takes far longer. Video adds time, and 3D point clouds add another dimension. Text and audio depend on length and on the language.
- How much there is in each item. One product on a white background is one label. A street scene can hold dozens of objects, each needing its own.
- The volume, and how steady it is. A large, steady flow lets a team be trained once and kept on the task. Small or irregular batches mean training again.
- The quality target and the review. One pass with spot checks is the lightest process. Two annotators on every item, with an expert settling the differences, is the heaviest. See how to measure data annotation quality.
- The deadline. A tight date needs more people on the task at once.
- Tools, format and security. Whose labeling tool is used, the format the labels come back in, and where the data may be held and by whom.
Three ways vendors charge
- By unit of work: for each image, label, document or minute of audio. It is predictable when the items are alike. Ask what counts as one unit and whether rework is included.
- By time: for the hours a team works. It fits tasks that vary a lot or are still being defined. Ask how output is reported.
- By project: one figure for a defined dataset and scope. It fits a task with clear edges. Ask what happens when the scope changes.
A project can start by time during a trial, and move to a unit basis once the pace is known.
What to send to get a firm quote
- 50 to 100 sample items that include the hard cases, not only the clean ones
- the labeling guideline, even as a draft
- the list of classes
- the total volume and the timeline
- the format you want back
- the accuracy you need and how you will measure it
- any limits on where the data may go
The more of this a vendor has, the less padding the quote needs.
Why a small batch makes the quote honest
Until someone has labeled your data, every figure is an estimate. A small batch shows the real time each item takes, the questions the guideline leaves open, and the accuracy that can be reached. See how to run a data annotation pilot.
Comparing quotes
- Compare the same thing: the same sample, the same guideline, the same quality target.
- Ask what is included: review, rework, project management, the tool.
- Count your own team's hours too. The lowest quote can be the dearest once your people are fixing labels. If you are still weighing your own team against a vendor, read in-house or outsourced data annotation.
How Pripton quotes
The cost depends on the type of data, the volume and how detailed the labels are, so we quote each project. Send a short description and a few samples through the contact page, and the team will reply with a quote. Our services are also listed on AWS Marketplace, where each listing states its own terms: see the platform page.
