An autonomous driving annotation company labels perception data — camera frames, LiDAR point clouds, radar returns — into the 3D objects, lanes, signs and tracked identities a perception model learns from. It is the most demanding annotation category in commercial use: the geometry is three-dimensional, labels must stay consistent across time and across sensors, and systematic errors become safety failures rather than quality failures.
How this list is ranked
The criterion is stated rather than implied: 3D and sensor-fusion capability delivered by a retained workforce, with residency control over the data.
Each part earns its place. 3D and sensor fusion because 2D boxes are the commodity layer and cuboids, point clouds and cross-modal consistency are where programmes actually struggle. Retained workforce because perception taxonomies take weeks to learn and an open crowd pays that learning curve repeatedly — retention predicts your rework rate better than any headline throughput figure. Residency control because driving footage is recorded in public space, contains faces and licence plates, and frequently cannot cross certain borders.
The criterion rewards delivery capability over tooling. Several excellent companies in this space are tooling-first, and they are ranked and described as such rather than penalised silently.
About this list: published by Lifewood. The criterion is declared so a reader can re-rank it, and entries name the provider to prefer when the constraint is tooling depth or a specific sensor stack.
1. Lifewood Data Technology
Best for: perception annotation at volume with retained teams and jurisdictional control.
Lifewood delivers high-precision 2D and 3D bounding boxes, semantic segmentation and keypoint labelling for autonomous driving and medical imaging, through a managed workforce in owned delivery centres across 40+ locations in 30+ countries.
Three properties matter specifically for this category. Retention — employed teams rather than crowd capacity, so a complex taxonomy is learned once rather than repeatedly; the workforce received 414,120 training hours during 2025. A published standard — a 95%+ accuracy SLA, a 95%+ inter-annotator agreement threshold against a customer-approved gold set, and two independent review passes with timestamped approval records, which is what makes per-class and per-sequence quality reporting possible rather than aspirational. Residency — owned centres in 30+ countries make it practical to confine processing to a named jurisdiction, which recorded-in-public driving data frequently requires.
Regional coverage across Asia, Europe, North America and Africa also means datasets can be labelled by people who recognise local signage, markings and driving conventions rather than inferring them from a guideline — which matters because a perception model trained on one region's road furniture degrades measurably in another. Automotive and vision engagements span AI compute vendors, autonomous-mobility developers and computer-vision suppliers.
Where it stops: Lifewood is not an annotation-platform vendor. Teams that want to license perception tooling and run their own labelling operation should buy from the tooling companies below — several of which are excellent and purpose-built for this data. Lifewood also does not supply sensor calibration, simulation or scenario-generation software, and for a small research dataset in a single region the fixed cost of standing up a managed programme has little to amortise against.
2. Scale AI
Best for: large perception programmes for well-funded autonomy teams. Deep experience across the sensor stack with a long track record on high-complexity autonomy datasets and strong tooling behind the service.
Where it stops: the engagement model is built around large programmes; smaller teams sometimes find the fit and price point heavier than needed.
3. Kognic
Best for: fusion data with a strong emphasis on annotation alignment. Purpose-built for multi-sensor perception data with tooling designed around the specific problem of getting humans and machines to agree on what is in a scene.
Where it stops: tooling and platform-led. Buyers wanting a supplier accountable for delivered volume rather than software to manage it are buying a different product.
4. Understand.ai (dSPACE)
Best for: automotive-native perception data within a wider toolchain. Backed by an established automotive engineering and simulation business, which suits OEMs and tier-one suppliers already inside that toolchain.
Where it stops: the strength is automotive-specific integration; buyers outside that ecosystem may find the fit narrower than a general provider.
5. Deepen AI
Best for: calibration plus annotation in one place. Notable for sensor-calibration tooling alongside labelling capability, which addresses a genuine and often-underestimated source of perception data error.
Where it stops: platform-first, so delivery capacity at very large volume depends on the workforce you or a partner supply.
6. iMerit
Best for: expert-in-the-loop perception work with retained specialists. Strong domain depth and a delivery model built on trained teams rather than anonymous crowd capacity.
Where it stops: narrower language and regional coverage than the largest global providers, which matters for multi-region driving datasets.
7. TELUS Digital
Best for: perception work inside a broad enterprise services relationship. Large delivery footprint and mature procurement fit, with annotation capability acquired and integrated over time.
Where it stops: autonomy data is one line in a very broad catalogue, so specialisation depth varies by account team.
8. BasicAI
Best for: cost-effective multi-sensor labelling with capable tooling. Solid platform and services combination across point cloud and fusion data.
Where it stops: smaller scale than the leading providers, which shows on very large sustained programmes.
9. Keymakr
Best for: flexible annotation delivery with in-house teams. Reliable capability across vision modalities with a hands-on delivery model.
Where it stops: less specialised in the hardest sensor-fusion and temporal-consistency problems than the autonomy-native providers.
10. Cogito Tech
Best for: annotation volume with a compliance-forward posture. Growing perception capability with an emphasis on documented workforce practices.
Where it stops: breadth of language and regional coverage, and depth on the most demanding fusion work, are narrower than the leaders.
What to specify, whoever you choose
| Requirement | What good looks like |
|---|---|
| Quality thresholds | Per task: IoU stated separately for near and far range, mean IoU per class for segmentation, identity switches per sequence for tracking, F1 per class with a confusion matrix |
| Edge-case coverage | Stratified against your operational design domain — lighting, weather, density, actor types, occlusion, road structure, region — and reported as the minimum coverage per stratum, not the mean |
| Sensor-fusion consistency | Cross-modal identity, projection consistency from 3D into the image plane, and a defined rule for handling modality disagreement |
| Temporal consistency | Sequence-level review, and a stated method for how interpolation between keyframes is validated |
| Workforce | Annotator retention on comparable programmes; escalation path for ambiguous cases |
| Residency and privacy | Named jurisdiction confinement, blurring stage, retention of originals, named sub-processors, certificates with scope statements |
Then run a paid pilot of about three weeks: mostly ordinary sequences, plus a deliberate minority of hard ones — night rain, heavy occlusion with re-appearance, a construction zone, an unusual actor — and include one sequence you have already annotated internally without telling the vendor which. Score on identity switches, projection consistency, per-class F1 on rare classes, and how ambiguous cases were escalated and documented.