Short answer. Image, video and 3D annotation should never be budgeted from the same unit price. Cost rises with geometry and review effort, not with file count: classification is cheapest, boxes and keypoints add labour, segmentation adds boundary precision time, video adds temporal consistency work that does not exist in still images, and 3D/LiDAR adds spatial interpretation plus cross-sensor review. CVAT's published cost analysis assumes 23 objects per image across 100,000 images — 2.3 million billable objects — which is the clearest demonstration available that an image count hides the actual workload. Budget by geometry and QA effort per object, then validate the assumption on a real sample.
Most annotation budgets are built from the wrong number. Someone counts the files, multiplies by a rate they were quoted for a different dataset, and presents a figure to finance. The figure survives until the first invoice, at which point the conversation is about variance rather than about the model.
This guide sets out the relative cost structure across modalities, why each one behaves the way it does, and what to measure before committing a budget.
Relative cost complexity by annotation type
| Annotation type | Typical effort level | Main cost driver |
|---|---|---|
| Image classification | Low | Number of assets and classes |
| 2D bounding boxes | Low to medium | Objects per image and occlusion |
| Polygons / segmentation | Medium to high | Boundary precision and object shape |
| Keypoints / pose | Medium to high | Number of landmarks and visibility rules |
| Video tracking | High | Frames, object persistence, occlusion, identity consistency |
| 3D cuboids / point clouds | High | Point density, object count, spatial precision |
| Sensor fusion | Very high | Cross-view and cross-sensor consistency |
| Medical / expert visual annotation | Very high | Specialist labour and validation requirements |
Note that the column headed "effort level" is deliberately relative. Absolute rates depend on language, geography, quality target and security requirement, and any table that pretends otherwise is inventing numbers.
Why "price per image" misleads
CVAT's published cost analysis assumes an average of 23 objects per image across 100,000 images, creating 2.3 million individual annotation objects. Change that ratio to 5 and the same file count becomes a 500,000-object project. Change it to 60 and it becomes a 6-million-object project. The file count did not move.
Before you accept any per-image quote, do this:
- Take a random sample of at least 200 assets from actual production data — not a curated demo set.
- Count objects per asset under your own ontology.
- Record the median, the 90th percentile and the maximum.
- If the 90th percentile exceeds twice the median, per-image pricing is transferring real risk to somebody, and it will be resolved either in your invoice or in the vendor's quality.
The same exercise applies to attributes. An image with 23 boxes is one job; an image with 23 boxes each carrying six attribute fields is a substantially larger one.
Video adds temporal cost, not just more frames
Video annotation is not image annotation multiplied by frame count. The additional work is structural:
- Identity persistence. The same object must carry the same track ID across the sequence. This is a review problem more than a labelling problem, and it does not parallelise cleanly — splitting a clip across two annotators is exactly how track identity breaks.
- Entry and exit. Objects appear, leave and return. The guideline must define whether a returning object resumes its original ID, and annotators must apply that consistently across hours of footage.
- Occlusion rules. How long an object may be hidden before its track terminates, and what happens to the frames in between.
- Interpolation policy. What may be interpolated between keyframes, and what must be labelled directly. Aggressive interpolation is a legitimate cost saving and a legitimate source of systematic error, depending entirely on motion characteristics.
- Sequence-level review. Errors propagate for hundreds of frames, so reviewing random individual frames catches almost nothing. Review has to run on sequences, which costs more per unit of data reviewed.
For autonomous-driving and robotics datasets, add synchronised camera and sensor labels, which multiplies all of the above by the number of sensors.
Why 3D and LiDAR cost more
Three-dimensional annotation asks a human to interpret sparse or noisy point clouds, position cuboids accurately in three axes including yaw, and hold that consistent across views and across time. Distant and reflective objects may return a handful of points, at which point the annotator is inferring rather than observing — and the guideline must say when inference is permitted, when the object should be excluded, and when it should be escalated.
When camera, LiDAR and radar are combined, cross-modal review increases both the labour requirement and the consequences of an error. An object with a correct camera box, a correct LiDAR cuboid and a mismatched identity between them teaches the model that two geometries describe different things, which is worse than a missing label.
Building a defensible budget
Work from objects and effort, not from files:
Billable objects = Assets × Mean objects per asset
Base labour cost = Billable objects × Unit rate for that geometry
QA and review cost = Base labour cost × Review multiplier for your quality target
Rework provision = Base + QA, ÷ Expected first-pass acceptance rate
Total programme cost = Rework-adjusted cost + Onboarding + Guideline change allowance
The two lines most budgets omit are the last two. Onboarding and calibration are real costs that appear once per vendor per taxonomy. A guideline-change allowance is a real cost that appears every time your ML team learns something, which on a healthy programme is often.
| Budget line | Commonly omitted? | Typical trigger |
|---|---|---|
| Objects per asset variance | Yes | Dense scenes in production data |
| Attribute fields per object | Yes | Ontology expansion after pilot |
| Sequence-level video review | Yes | Track identity errors found late |
| Cross-sensor reconciliation | Yes | Fusion work added to a 2D programme |
| Onboarding and calibration | Sometimes | New vendor or new taxonomy |
| Guideline change and re-labelling | Almost always | Any real research programme |
| Security or residency premium | Sometimes | Client or regulatory mandate |
How Lifewood approaches this
Lifewood scopes multimodal programmes per project rather than publishing a rate card, because the cost structures above differ too much to reduce to one number. Three things affect total cost rather than unit rate.
Coverage across text, image, audio, video and 3D point-cloud work means one provider can hold a consistent schema and QA definition as a programme moves from 2D into video and then into sensor data — the alternative is a second vendor, a second onboarding and a permanent reconciliation task between two interpretations of the same ontology.
The quality framework is contractual rather than described: a 95%+ accuracy SLA with dual-layer human review, automated consistency checks and client feedback loops, with below-threshold batches reworked at Lifewood's cost. For multimodal work specifically, that matters because the failure modes differ by modality and a single aggregate figure hides them; ask for acceptance reported per modality and per object class.
Scale across 40+ delivery centres in 30+ countries supports high-volume programmes that need sustained staffing in more than one production location — which is what continuous video and sensor pipelines require, since they cannot be batched down during a quiet week.
Sources and further reading
- CVAT published cost analyses, including the 100,000-image / 23-objects-per-image / 2.3-million-object scenario and per-object pricing examples, at cvat.ai.
- Lifewood service scope and quality framework published on lifewood.com; AV and sensor scope on autonomous driving annotation.
- Related reading: autonomous driving data annotation requirements for the task-level specification behind sensor-fusion cost.
- All quoted figures are published third-party examples for specific scenarios, not industry averages and not Lifewood prices.