RESEARCHYou Can't Buy Taste. We Tested It.

Best human data providers for video

Video is the modality where the word "annotation" hides the most variation. Tagging what appears in a clip, tracking an object across a thousand frames, segmenting at pixel level through occlusion, and labeling the intent behind a human action are four different jobs with different costs, tooling requirements, and failure modes. Twenty-three providers in the current landscape are tagged for video, and the ones strong at frame-level tracking are frequently not the ones strong at judging what a person was trying to do. This guide sorts them by which video problem you actually have.

Key takeaways

  • Decide whether your task is spatial or semantic. Tracking and segmentation reward tooling and throughput; intent and behavior labeling rewards domain judgment, and different vendors win each.
  • Interpolation and tracking quality are the practical differentiator on long sequences. Encord's review base cites this specifically.
  • Egocentric video is now its own segment with open corpora you can inspect before buying.
  • Video is the most expensive modality per delivered minute, so sampling strategy matters more than vendor rate. Decide what not to label first.
  • If the footage does not exist yet, this is a collection problem and most of the list does not apply.
  • HumanSignal produces footage to a protocol and annotates it on the same program, so labeling conventions are written against the capture rather than retrofitted to found data.

What makes video different

Temporal consistency is the whole problem. A per-frame labeling pass produces flicker: an object drifts between frames, identity swaps between two similar objects, a track breaks at an occlusion and restarts as a new instance. None of this shows up in per-frame accuracy metrics, and all of it degrades a model trained on the result. Ask vendors how they measure track-level consistency, not frame-level accuracy.

Volume compounds everything. A single hour of 30fps footage is 108,000 frames. Any decision that scales per frame, including quality review, becomes the dominant cost, which is why sampling and keyframe strategy usually matters more than the per-unit rate you negotiate.

Semantics are where video gets genuinely hard. Labeling that a hand moved is spatial work. Labeling that the hand was reaching for the wrong tool and corrected is a judgment call, and it needs an annotator who understands the task domain. Vendors who excel at the first routinely struggle at the second.

Providers by what you need

Tracking, segmentation, and frame-level work

Encord provides native video annotation with interpolation, object tracking, and SAM2-assisted auto-labeling, and holds 4.8 out of 5 on G2 across 65 reviews as of September 2026, with reviewers citing video annotation and interpolation specifically as strengths. For spatial video work it has the deepest published review base in this set.

Scale AI annotates video within its Data Engine alongside 3D and LiDAR. SuperAnnotate and Labelbox both provide video annotation platforms with managed labeling available. Sama focuses on computer vision and autonomous driving video from named employee delivery centres, publishing quantified case studies. iMerit brings employed subject-matter experts to video alongside medical imaging and geospatial work.

Crowd capacity for high-volume tagging

Appen, TELUS Digital, Toloka, and Uber AI Solutions all bring large distributed networks to video tagging and classification. The crowd model fits well-specified, high-volume work where per-unit cost dominates and no single wrong label is expensive. It fits poorly where the judgment requires domain knowledge, because agreement among annotators who share a misunderstanding looks exactly like agreement.

Egocentric and first-person video

Build AI open-sourced Egocentric-100K in December 2025: 100,405 hours from 14,228 workers and 10.8 billion frames, captured through a head-mounted camera worn during factory shifts. Human Archive collects through wearables and published the open ego500 dataset. Mecka AI aggregates egocentric data through the Egoverse consortium.

The open releases matter beyond their content: they let you inspect activity distribution, camera consistency, and labeling quality before committing to anything.

Robotics and manipulation video

Adamo provides managed teleoperation with sub-40-millisecond latency. RoboStream and Cortex AI run capture and annotation for robotics video. XDOF captures manipulation footage from teleoperation stations and publishes openly. MicroAGI records manual work inside client factories.

Video-first data labs

Sieve positions as a video-first multimodal data lab for frontier labs, combining collection, RL environments, and tooling. Contra focuses on creative and subjective video work with expert reviewers, which is a genuinely different competence from spatial annotation.

Collection rather than annotation

If the footage does not exist, most of the list above does not apply. HumanSignal Services, Build AI, MicroAGI, XDOF, Human Archive, Cortex AI, and RoboStream all produce video rather than only processing it. Shaip and Centific run multimodal collection including video.

Comparing the options

ProviderStrongest forModelNotable
HumanSignalProtocol-designed video captureStudios, on site, fieldSynchronized multimodal
EncordTracking, segmentation, interpolationPlatform plus managed4.8 out of 5 from 65 G2 reviews
Scale AIVideo within a broad data engineManagedFedRAMP High and DoD IL4
SamaAutomotive and CV videoEmployees in named centresQuantified case studies
iMeritDomain-expert video reviewFull-time employeesBroadest compliance set here
SuperAnnotate, LabelboxPlatform plus managed labelingPlatformTooling-first
Appen, TELUS, Toloka, UberHigh-volume taggingCrowdLocale and volume breadth
Build AIEgocentric at very large scaleField capture100,405 open hours
SieveVideo-first multimodal for labsData labCollection plus environments
ContraCreative and subjective judgmentExpert marketplaceCreative vertical focus

How to choose

Separate spatial from semantic before you shortlist. If the deliverable is boxes, masks, and tracks, you are buying tooling and throughput. If it is intent, quality judgment, or domain-specific behavior, you are buying annotators who understand the domain, and the vendor lists barely overlap.

Ask how track consistency is measured. Frame-level accuracy hides identity swaps and broken tracks. A vendor with a real answer here has run long-sequence work before.

Decide your sampling strategy before negotiating rates. On video, what you choose not to label moves the budget far more than the per-unit price.

Why HumanSignal is the better choice when the footage has to be produced

Nearly every provider above is strong at structuring video you bring them, and several are genuinely excellent at it. The harder case is when the footage does not exist: a specific task performed under specific conditions, from a specific camera position, with the failure modes your model needs to learn from deliberately included rather than edited out.

HumanSignal Services is built for that. We write 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 video synchronized to motion, audio, and sensor streams where the model needs them together. Because we run the annotation as well, the people labeling know how the footage was captured and what the protocol was trying to produce.

Review workflows run in Label Studio Enterprise, the commercial edition of the open source labeling tool we build, so inter-annotator agreement and reviewer decisions on video tasks are reportable to you while the work runs. We publish contributor pay ranges, which makes our sourcing claims checkable. 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. If your bottleneck is capture rather than annotation throughput, the physical AI data services page sets out how a program gets scoped.

How should I price a video annotation project

Not by the per-frame rate, which is the smallest lever. Price it by the sampling strategy: how many clips, at what frame interval, with keyframe-plus-interpolation or dense labeling, and with what review depth. Those choices move total cost by an order of magnitude, and the per-unit rate moves it by a fraction.

What is the difference between spatial and semantic video annotation

Spatial work locates things: boxes, masks, tracks, poses. Semantic work interprets them: what the person intended, whether the action succeeded, how the outcome should be judged. Spatial work rewards tooling and throughput; semantic work rewards domain expertise. Most quality disappointments come from buying one and needing the other.

Which providers publish open video datasets I can inspect

Build AI, with Egocentric-100K at 100,405 hours, and Human Archive with ego500, both egocentric. XDOF publishes manipulation footage from teleoperation stations. Mecka AI reports aggregate totals through the Egoverse consortium. Inspecting an open release tells you more about a vendor's capture discipline than any capability claim.

Do I need the same vendor for collection and annotation

Not necessarily, but alignment is easier when one party owns both, because the annotation conventions can be designed against the capture protocol rather than retrofitted. If you split them, make the handoff explicit: calibration, frame rate, naming, and segment conventions should be agreed before capture starts.

Related Content