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. And 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.
The question comes up in almost every scoping conversation, usually phrased something like: we have 400,000 images, how fast can you turn them around?
It is a completely reasonable question and it does not have a single-number answer, because the honest response depends on what "annotate" means for that dataset. The same image can take seven seconds or two minutes depending on whether you want a box around the car or a pixel-perfect mask of it. That is a seventeen-fold difference in the same modality, on the same file, and it is the single biggest driver of a delivery timeline.
So this piece is an attempt to put real numbers on it, task by task, with sources. And then to explain the more important thing, which is why you cannot take those numbers, multiply by a workforce size, and get a delivery date.
The per-task numbers, with actual research behind them
The most useful published timing data comes from the COCO dataset annotation work, because the researchers documented their per-instance times precisely enough to be usable as a benchmark.
Bounding boxes: roughly 7 seconds per instance using the extreme clicks technique, once the object has been located and categorised. That is the drawing time in isolation.
The full pipeline number is more instructive. In the COCO protocol, annotating one instance end to end took 50.2 seconds:
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 surprises people and it matters enormously for scoping, because it means pre-labelling that finds and classifies objects saves far more time than pre-labelling that draws boxes around objects you have already identified.
Polygon masks: roughly 79.2 seconds per instance. The COCO figure is 22 hours per 1,000 instances for polygon-based mask annotation. With category labelling and instance spotting added, a full mask instance is 122.4 seconds.
So a mask is roughly 2.4 times the total cost of a box in the same pipeline, and about 11 times the drawing cost.
Point-based annotation: about 9 seconds per instance for 10 points given an existing bounding box, which the researchers noted is more than eight times faster than polygon-based mask annotation.
The aggregate view makes the difference concrete. To annotate COCO train2017, 118,287 images containing 849,949 instances, the researchers calculated 493 days for bounding boxes, 1,204 days for masks, and 582 days for point-based annotation. Same images, same objects, two and a half times the calendar.
The complexity ladder across modalities
Timing research is thinner outside computer vision, but the relative ordering is consistent across the industry and is worth knowing when scoping a mixed programme.
For image data, from least to most time-intensive: classification, bounding boxes, polygon annotation, then semantic and instance segmentation at the most complex end.
For text, document classification is the 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 annotation cost by the number of frames, and tracking annotation adds the overhead of maintaining consistency across sequences on top of that.
For 3D data, cuboid and point cloud annotation requires significant geometric expertise and is consistently the highest-cost annotation category per item.
Pricing data corroborates the ordering, since price is mostly a proxy for time. 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. In 3D specifically, a cuboid starts from around $0.121 against $0.036 for a 2D bounding box, reflecting the extra work of establishing depth, extent and heading. Segmentation, polygon and 3D point cloud work has been reported at 10 to 50 times the cost of basic bounding boxes.
Transcription and speech. A trained transcriptionist takes three to six hours to produce one finished hour of transcript, which is the figure that governs any speech programme timeline. Diarization and timestamping add to that.
LiDAR frames. An experienced annotator working at production pace on complex urban scenes processes roughly 8 to 15 frames per hour, substantially fewer in difficult conditions.
Preference pairs. This is the least standardised and the hardest to benchmark honestly. A single-dimension preference judgement on short responses can be fast. A multi-dimension rubric applied to long-form responses, where the rater must read both outputs carefully and score several axes separately, is closer to five to ten minutes per pair. The variance here is larger than in any other task type, and anyone quoting a firm number without knowing the rubric is guessing.
Why you cannot multiply by headcount
Here is the part that actually determines delivery dates, and where most naive estimates go wrong.
Take a workforce number, multiply by the per-hour rate, multiply by working hours. It produces a beautiful figure that has almost no relationship to what ships.
Four things eat the difference.
QA overhead is real capacity. If a programme runs peer review plus 15% QA sampling, a meaningful share of the workforce's hours are spent reviewing rather than producing. That is not waste, it is the mechanism that makes the output usable, but it has to come out of the throughput calculation. Depending on programme phase and sampling rate, review can consume 15 to 30% of total available hours.
Rework is not free. Rejected items come back and consume time twice. A programme with a 10% rework rate is delivering at roughly 90% of its apparent capacity even before QA overhead.
Ramp is not instant. A new cohort does not produce at experienced rates. New annotators work at reduced speed with elevated QA for their first weeks, so a project that staffs up on day one delivers substantially below nominal capacity for the first month.
Not everyone is on every task. A workforce spanning many countries and languages is not fungible. The 56,000-plus contributors in Lifewood's network are not 56,000 people available for any given task; they are a pool from which the subset qualified for a specific language, modality and specification is drawn. For a Yoruba speech programme, the relevant capacity is the number of qualified Yoruba speakers, not the network total.
That last point is the one that matters most for multilingual work, and it is why headline workforce numbers are a poor proxy for delivery speed on a specific programme. What determines the timeline is depth in the specific language and task, not breadth across the network.
A rough planning rule that holds up reasonably well: 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. That gets you closer to a defensible number than a straight multiplication.
What actually speeds things up
Given all that, the levers that genuinely move a timeline are worth knowing.
Pre-labelling that finds objects, not just draws them. Given the COCO finding that spotting and categorising is 86% of instance time, a model that locates and classifies candidate objects saves far more than one that refines geometry on objects you have already found.
Task chaining. Starting segmentation from completed cuboid runs, or starting fine annotation from coarse passes, reduces repeated work. Platform capabilities matter here more than annotator speed.
Reducing task complexity where the model does not need the precision. This is the most underused lever. If a detection model will perform adequately on boxes, specifying polygon masks costs you 2.4 times the annotation time for accuracy the model cannot exploit. Matching annotation fidelity to actual model requirements is a scoping conversation worth having before production starts.
Guideline maturity. An ambiguous specification produces rework, and rework is the largest recoverable inefficiency in most programmes. Calibration time spent up front converts directly into throughput later.
Not compressing QA. Counterintuitively, cutting QA to hit a deadline usually slows delivery, because the errors surface at client review instead and come back as a larger rework batch with a worse relationship attached.
How to ask for a throughput estimate properly
If you are scoping a programme, the questions that produce a useful answer rather than a hopeful one are:
What exactly is the annotation task, at the level of the output artefact? Not "annotate the images" but "one bounding box per vehicle, pedestrian and cyclist, with occlusion flags, minimum 20-pixel size threshold."
How many objects per item, on average? Per-instance rates only convert to per-image rates if you know the instance density. An urban driving frame with 40 objects and a rural frame with 3 are not the same job.
What accuracy level, measured how? An IoU threshold of 0.92, which is the reported industry minimum for autonomous driving and medical AI, requires meaningfully more care per object than a looser tolerance.
What is the QA regime? Sampling rate, peer review, and rework expectations all come out of capacity.
How many contributors qualify for this specific task and language? Not the network size.
What does the ramp look like? If the programme needs a new cohort, the first month delivers below rate.
A supplier who asks you these questions before quoting a timeline has run these programmes before. One who quotes a number from the volume alone has not.
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. 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 per instance, more than eight times faster than polygon masks.
- Annotating COCO train2017 was calculated at 493 days for boxes, 582 days for point-based and 1,204 days for masks.
- Complexity ordering: image classification, boxes, polygons, then segmentation; text classification, NER, sentiment, relation extraction, coreference; audio transcription then diarization, emotion tagging, phoneme annotation; 3D cuboid and point cloud highest of all.
- Basic labels run $0.03 to $1.00 each against $0.05 to $5.00 for complex masks; a 3D cuboid starts around $0.121 against $0.036 for a 2D box.
- A trained transcriptionist takes three to six hours per finished audio hour. LiDAR annotators process roughly 8 to 15 frames per hour on complex urban scenes.
- 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.
- A rough planning rule is 70 to 80% of nominal rate after QA and rework, with a further discount for cohorts in their first month.
- The real speed levers are pre-labelling that spots objects, task chaining, matching annotation fidelity to actual model requirements, and guideline maturity.
Sources and further reading
- "Pointly-Supervised Instance Segmentation" (arXiv), on COCO per-instance annotation times for bounding boxes, polygon masks and point-based annotation, and the COCO train2017 day calculations
- DataVLab, "Data Annotation Pricing 2026", on the complexity ordering across image, text, audio, video and 3D modalities
- BasicAI, "How Much Do Data Annotation Services Cost?", on basic versus complex label price tiers
- ANOSUPO AI, "3D Point Cloud Annotation for Autonomous Driving", on 3D cuboid versus 2D bounding box unit rates
- Lightly, "Best Data Annotation Companies in 2026", on segmentation and 3D point cloud costing 10 to 50 times basic bounding boxes
- Precise BPO, "Bounding Box Annotation Benchmarks", on IoU thresholds for autonomous driving and medical AI and the MIT CSAIL finding on mAP loss from annotation error
- ConvertAudioToText, "Cost of Transcription Per Hour in 2026", on the three to six hour ratio for finished transcript hours
- Lifewood, company timeline, on crowd resource growth from 20,000 in 2021 to 56,000+ and the delivery network
- Note on figures: per-instance timings are drawn from published research on specific datasets and protocols. Actual rates vary with task specification, instance density, tooling, annotator experience and accuracy requirements, and should be validated against a pilot rather than assumed.