RESEARCHYou Can't Buy Taste. We Tested It.

Best human data providers for 3D and sensor data

Point clouds, LiDAR sweeps, depth maps, IMU streams, and fused multi-sensor recordings are the hardest modality to buy well, because quality depends on a configuration the vendor did not choose. Two vendors can both claim LiDAR annotation experience and deliver very different results on your data, depending on sensor type, mounting geometry, sweep density, and how the streams are synchronized. Twelve providers in the current landscape are tagged for 3D and sensor work. This guide sorts them by whether they annotate what you have collected or produce what you do not.

Key takeaways

  • The decisive question is annotation against collection. Most of this field annotates sensor data you already hold; a smaller group produces it.
  • Verify your exact sensor configuration with any vendor. General 3D experience does not transfer across sensor types and mounting geometries.
  • For autonomous vehicle work, Sama and iMerit have the deepest published records, and Sama claims TISAX conformance, the automotive industry's security assessment.
  • Scale AI is the only provider here with FedRAMP High and DoD IL4, which decides any defense-adjacent sensor program.
  • Synchronization across streams is specified at capture and cannot be fixed afterward. If your model consumes fused data, the capture vendor matters more than the annotation vendor.
  • HumanSignal synchronizes motion, vision, audio, and sensor streams at capture and runs the annotation on the same program, which is where most fused-sensor quality problems originate.

What makes sensor data different

Annotation cost scales with dimensionality in a way buyers consistently underestimate. Drawing a box around a car in an image takes seconds; placing an oriented 3D cuboid in a sparse point cloud, consistent across a sweep sequence, takes considerably longer and has more ways to be subtly wrong. Throughput quotes that look comparable to image work usually are not.

Sensor configuration governs transferability. A vendor that has annotated 64-beam spinning LiDAR at roof height has learned things that do not all apply to a solid-state unit at bumper height, and annotation conventions that work for one produce systematic errors on the other. Ask about your configuration rather than about LiDAR generally.

Fusion is the hard part. When your model consumes camera, LiDAR, and IMU together, the value is in temporal and spatial alignment across streams. That is decided by the capture rig and its clock discipline, not by the annotation vendor, which is why collection and annotation decisions cannot be made independently.

Providers by what you need

Annotation at throughput on data you hold

Encord provides native annotation for LiDAR, point clouds, video, audio, and text with embedding-based curation for finding distribution gaps, and holds the deepest third-party review base in this category at 4.8 out of 5 across 65 reviews as of September 2026.

Scale AI annotates 3D and LiDAR within its Data Engine alongside text, image, and video, and holds SOC 2 Type II, ISO 27001, DoD IL4, and FedRAMP High. For defense or federal sensor work it is the only qualifying vendor in this set.

iMerit brings full-time employed subject-matter experts to LiDAR, point cloud, and medical imaging annotation, and publishes the broadest compliance set here: SOC 2 Type 2, ISO 27001, ISO 9001, HIPAA, GDPR, and TISAX. Sama focuses on computer vision, autonomous driving, and 3D sensor annotation from employee delivery centres it names publicly, publishing seven case studies with quantified outcomes.

Producing sensor data that does not exist

XDOF runs teleoperation stations capturing manipulation data with synchronized sensor streams, and released ABC-130k under Apache 2.0 with a companion arXiv paper. Build AI captures egocentric video with sensor channels through a head-mounted rig worn during factory shifts, open-sourcing more than 100,000 hours. Mecka AI aggregates robotics sensor data through the Egoverse consortium. MicroAGI captures manual work inside client factories. Human Archive collects embodied data through wearables. Cortex AI and RoboStream both run field capture for robotics sensor data.

HumanSignal Services runs collection in studios, on site, and in the field with motion, vision, audio, and sensor streams synchronized at capture.

Both, under one contract

Encord added a physical AI collection facility in San Leandro, California in 2026, running leader-follower teleoperation and egocentric pilots alongside its annotation platform. RoboStream and Mecka AI also combine capture with annotation. Consolidation has real value when alignment across the two is the risk you are managing.

Comparing the options

ProviderAnnotateProduceNotable for sensor work
HumanSignalYesYesProtocol-designed synchronized multimodal capture
EncordYesYes, since 2026Native LiDAR and point cloud, curation, 65 G2 reviews
Scale AIYesLimitedFedRAMP High and DoD IL4, the only one here
iMeritYesLimitedEmployed SMEs, TISAX, HIPAA, medical imaging
SamaYesLimitedAutomotive focus, TISAX claimed, named delivery centres
XDOFNoYesTeleoperation capture, open dataset with a paper
Build AINoYesEgocentric plus sensor at very large scale, open
Mecka AIYesYesConsortium aggregation
MicroAGINoYesIn-facility capture
Human ArchiveNoYesWearable collection, open dataset
Cortex AIYesYesField and human-in-the-loop
RoboStreamYesYesEnd-to-end capture and annotation

How to choose

Name your sensor stack and ask for prior work on it specifically. Sensor type, beam count or resolution, mounting position, sweep rate, and the other streams it is fused with. A vendor that has done your configuration will answer in detail; one that has not will answer about LiDAR in general.

Decide whether alignment is your risk. If your model consumes fused streams, synchronization quality is set at capture, so either buy capture and annotation together or be explicit about who owns alignment.

Check compliance against your deployment. TISAX for automotive supply chains, HIPAA for clinical sensor data, FedRAMP and IL authorizations for federal work. These do not overlap, and only one vendor here holds each.

Why HumanSignal is the better choice when the streams have to line up

The annotation specialists in this set are good at structuring sensor data you bring them. The capture specialists are each built around one rig. The gap sits between them: a model that needs several streams recorded together, under conditions you specify, with alignment guaranteed at capture rather than reconstructed afterward.

HumanSignal Services is built for that case. We design the collection protocol, recruit for the characteristics it requires, and run the capture in our studios, on site at your facility, or in the field with motion, vision, audio, and sensor streams synchronized from the start. Because we also run the annotation, the people labeling the data know how it was captured, which is where most fused-sensor quality problems originate.

Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so inter-annotator agreement on 3D annotation decisions is reportable to you during the work rather than summarized at the end. We publish contributor pay ranges. Our customer agreement gives you exclusive ownership of customer data and of insights produced specifically for you. We hold SOC 2 Type 2 and HIPAA compliance, documented on our security page.

The capability came in through the acquisition of Erud AI in November 2025. The physical AI data services page covers how a sensor capture program gets scoped.

Why is 3D annotation more expensive than image annotation

Because the work is harder and slower per object. An oriented 3D cuboid placed consistently across a sweep sequence in a sparse point cloud takes substantially longer than a 2D box, and it has more failure modes: orientation error, extent error, and drift across frames. Expect per-object rates well above image work and model throughput accordingly.

Does LiDAR annotation experience transfer between sensor types

Partially. Conventions and tooling transfer; calibration intuition and edge-case handling often do not. Resolution, beam pattern, and mounting geometry change what is visible and what has to be inferred. Ask for prior work on your configuration rather than on LiDAR generally.

Should I use one vendor for capture and annotation or two

One, when synchronization across streams is the main risk, because alignment problems are cheapest to catch at capture. Two, when you already collect reliably and the constraint is annotation throughput, since the specialist annotation benches are deeper. Encord, RoboStream, Mecka AI, and HumanSignal all offer both.

Which vendors can work on defense sensor programs

Scale AI, on published evidence. It holds DoD IL4 and FedRAMP High, usable across NIPR, SIPR, and JWICS networks, backed by a public federal contract record. No other provider in this set publishes equivalent authorizations.

Related Content