Short answer. There is no single annotation rate, because the output artefact drives the time far more than the modality does. COCO's own measurements put a bounding box at 7 seconds of drawing but 50.2 seconds end to end per instance — 86% of which is finding and categorising the object, not drawing it. A polygon mask is about 122 seconds end to end, roughly 2.4× a box. Annotating COCO train2017 works out at 493 days for boxes and 1,204 days for masks on identical images. None of those numbers can be divided by headcount to produce a delivery date, because review, rework and ramp-up sit between throughput and delivery.
Key takeaways
- The same image can take 7 seconds or 2 minutes to annotate depending on the output artefact, which is the largest single driver of a delivery timeline.
- COCO research puts bounding box drawing at roughly 7 seconds per instance, with full end-to-end instance annotation at 50.2 seconds: 28.8 for category labelling, 14.4 for instance spotting, 7.0 for the box.
- Object spotting and categorisation is 86% of instance time, so pre-labelling that finds objects saves far more than pre-labelling that refines geometry.
- Polygon masks take roughly 79.2 seconds per instance, or 122.4 seconds end to end, about 2.4 times a box; point-based annotation with 10 points takes about 9 seconds, more than eight times faster than a mask.
- Nominal capacity is not delivered capacity: QA can consume 15 to 30% of hours, rework consumes time twice, new cohorts ramp for weeks, and a global workforce is not fungible across languages and tasks.
How long does a bounding box or a polygon mask actually take to annotate?
A bounding box takes about 7 seconds to draw once the object has been located, but 50.2 seconds end to end once finding and categorising it are included; a polygon mask takes roughly 2.4 times as long as a box in the same pipeline. Annotation throughput is the rate at which finished, reviewed instances are produced per unit of time, and it changes by an order of magnitude depending on the artefact requested, not the image content.
The most useful published timing data comes from the COCO dataset annotation work, because researchers documented per-instance times precisely enough to use as a benchmark. Bounding boxes run roughly 7 seconds per instance using extreme-click techniques — the drawing time in isolation. The full pipeline number is more instructive: annotating one instance end to end took 50.2 seconds, split as 28.8 seconds for category labelling, 14.4 seconds for instance spotting, and 7.0 seconds for the box itself. The box is 14% of the total time; finding and classifying the object is the other 86%. That ratio matters for scoping, because pre-labelling that finds and classifies objects saves far more time than pre-labelling that draws boxes around objects already identified.
Polygon masks run roughly 79.2 seconds per instance for the drawing alone, or 122.4 seconds end to end once category labelling and instance spotting are added — about 2.4 times the total cost of a box in the same pipeline, and about 11 times the drawing cost. Point-based annotation, using 10 points against an existing bounding box, takes about 9 seconds per instance, more than eight times faster than polygon masks.
The aggregate view makes the difference concrete. Annotating COCO train2017 — 118,287 images containing 849,949 instances — was calculated at 493 days for bounding boxes, 1,204 days for masks, and 582 days for point-based annotation. Same images, same objects, more than double the calendar time between the fastest and slowest artefact.
How does annotation time and cost scale across other modalities?
Timing research is thinner outside computer vision, but the relative ordering of task complexity is consistent across the industry and pricing data tracks it closely, since price is mostly a proxy for time. For image data, complexity rises from classification through bounding boxes and polygon annotation to semantic and instance segmentation. For text, document classification is fastest, followed by named entity recognition, sentiment analysis, relation extraction, and coreference resolution. For audio, transcription is the baseline, with diarization, emotion tagging and phoneme annotation adding cost in that order. For video, frame-level tasks multiply the equivalent image cost by the number of frames, and tracking adds the overhead of maintaining consistency across a sequence. For 3D data, cuboid and point cloud annotation requires geometric expertise and is consistently the highest-cost category per item — a starting point worth checking against a full pricing breakdown by modality.
Pricing corroborates the ordering: basic labels such as bounding boxes run roughly $0.03 to $1.00 each, while complex labels such as precise semantic masks hold at $0.05 to $5.00. A 3D cuboid starts from around $0.121 against $0.036 for a 2D bounding box, reflecting the extra work of establishing depth, extent and heading, and segmentation, polygon and 3D point cloud work has been reported at 10 to 50 times the cost of basic bounding boxes. A trained transcriptionist takes three to six hours to produce one finished hour of transcript, and an experienced annotator working at production pace on complex urban LiDAR scenes processes roughly 8 to 15 frames per hour. Preference pairs are the hardest task type to benchmark honestly: a single-dimension judgement on short responses can be fast, while a multi-dimension rubric applied to long-form responses runs closer to five to ten minutes per pair, and anyone quoting a firm number without knowing the rubric is guessing — see how a preference rubric affects rater agreement for what drives that variance.
Why can't you multiply workforce size by an hourly rate to get a delivery date?
You cannot, because at least four factors separate nominal capacity from delivered capacity: QA overhead, rework, ramp-up, and language- or task-specific fungibility. A number produced by multiplying headcount by an hourly rate looks precise but has almost no relationship to what ships.
QA overhead is real capacity, not waste — a programme running peer review plus sampling can spend 15 to 30% of available hours on review rather than production, and that time is what makes the output usable. Rework is the return and re-annotation of items rejected at review, and it consumes time twice: a programme with a 10% rework rate is delivering at roughly 90% of its apparent capacity before QA overhead is even applied. Ramp is the reduced-speed, elevated-QA period a new annotator cohort works through before reaching experienced production rates, typically its first few weeks on a task. And not everyone in a workforce is qualified for every task: a network is a pool from which the subset qualified for a specific language, modality and specification is drawn, so headline workforce size is a poor proxy for delivery speed on any one programme — what determines the timeline is depth in that specific language and task, not breadth across the network. A rough planning rule that holds up reasonably well is to take the nominal per-annotator rate, apply 70 to 80% for QA and rework overhead, and apply a further ramp discount for any cohort in its first month.
What actually speeds up annotation throughput?
The levers that genuinely move a timeline are pre-labelling that finds objects rather than just refining their geometry, task chaining, matching fidelity to what the model needs, and mature guidelines. Given the COCO finding that spotting and categorising an object is 86% of instance time, a model that locates and classifies candidates saves far more than one that only refines geometry on objects already found — a distinction worth applying before scaling any model-assisted labelling and active learning workflow. Task chaining — starting segmentation from completed cuboid runs, or fine annotation from coarse passes — reduces repeated work, and platform capabilities matter here more than annotator speed.
Reducing task complexity where the model does not need the precision is the most underused lever: specifying polygon masks when boxes would perform adequately costs 2.4 times the annotation time for accuracy the model cannot exploit, so matching fidelity to actual model requirements is a scoping conversation worth having before production starts. Guideline maturity matters just as much — an ambiguous specification produces rework, the largest recoverable inefficiency in most programmes, and calibration time spent up front converts directly into later throughput; see how to write annotation guidelines annotators actually follow for what that calibration looks like in practice. Counterintuitively, cutting QA to hit a deadline usually slows delivery, because the errors it would have caught surface at client review instead and return as a larger rework batch.
What should you ask before trusting a throughput estimate?
A useful estimate depends on the output artefact, the instance density per item, the accuracy target, the QA regime, and how many contributors actually qualify for the task and language — not on volume alone. Ask what exactly the annotation task is at the level of the output artefact ("one bounding box per vehicle, pedestrian and cyclist, with occlusion flags and a minimum 20-pixel threshold," not "annotate the images"), because per-instance rates only convert to per-image rates once instance density is known — an urban driving frame with 40 objects and a rural frame with 3 are not the same job. Ask what accuracy level is required and how it is measured: an IoU threshold of 0.92, the reported industry minimum for autonomous driving and medical AI, needs meaningfully more care per object than a looser tolerance. Ask what the QA regime is, since sampling rate, peer review and rework expectations all come out of capacity, and how many contributors qualify for the specific task and language, not the network size. Ask what the ramp looks like if a new cohort is required, since the first month delivers below nominal rate. A supplier who asks these questions before quoting a number has run the programme before; one who quotes from volume alone has not — the same discipline that underpins layered quality control before AI data is delivered and the gold-set and consensus methods used to QA annotated data.
Lifewood applies 95%+ inter-annotator agreement thresholds measured against a customer-approved gold set and two independent review passes with timestamped approval records before delivery, which is the QA overhead a realistic throughput estimate has to account for rather than treat as slack to cut — an approach detailed in how a new annotation centre is opened and trained and reflected across providers compared in large-scale AI data annotation. Lifewood's contributor network logged 414,120 training hours in Bangladesh in 2025, illustrating the scale of ramp and calibration investment that sits behind a qualified, task-specific cohort rather than a raw headcount figure. Buyers scoping a mixed programme can also compare unit-rate assumptions against a broader service overview of AI data annotation and validation or a dedicated AI data validation process before committing to a delivery date.