Egocentric Video Data From Real Workers Doing Real Jobs

Egocentric data from real workers is first-person video recorded while a person does their actual job at a working site. The work is real, the setting is a place where the work normally happens, and the clips come from normal operations.

It is a sourcing choice, separate from a task or a camera. Any task page on this site can be collected this way. This page explains how real-worker capture is run and what the site states about it.

In staged capture, a person is asked to act out a task. In real-worker capture, a person is recorded doing the task they would be doing anyway. The order of steps, the pauses and the way tools are picked up come from the job, not from a script.

The two can be combined. A spec can ask for real work for most of a program and a scripted edge case where the real job rarely produces one. Either way the spec, not the contributor, sets what counts as an accepted clip.

A real worker is someone trained in the work being captured, such as a store associate, a technician, a member of a kitchen crew, a tradesperson or a warehouse operator. The role comes from the task in your spec.

The site does not state how contributors are recruited into a given program, how many there are, or in which countries, so those details are Not stated. They are confirmed per engagement.

The site task catalog groups commercial work by sector and environment, and its environments hub lists each one with its tasks. It includes settings such as cafes and bars and catering and event kitchens, alongside many other working environments.

This page does not claim a setting that the catalog does not list. Most catalog rows are Available on request, so reach for a given site is quoted against your spec.

The written spec names the task, the setting, the camera setup, the labels and the accept criteria. Clips are bounded by semantic task, so a clip covers one task from start to finish rather than an arbitrary stretch of a shift.

The spec also sets what is out of scope. A real worker will do unrelated things during a shift, and clips outside the spec are not accepted.

Sessions are scheduled around real shifts. A session is planned for the part of the shift where the target task happens, so capture adds as little disruption as possible to the working day.

Session length, the number of sessions per contributor and the shift patterns are Not stated. They are agreed in the spec for each program.

Before capture, site access and permission are secured and a written release is signed. Identity is verified and the rate is agreed. The contributor is briefed on the spec so they know which tasks and clip boundaries count.

The consent record is part of the delivery. Every episode links to a signed contributor release through a documented consent chain.

Each contributor is certified with a test clip before their data counts toward a program. The clip is checked against the spec, so a contributor who does not follow the clip boundaries or camera setup is caught early.

The pass criteria for the test clip are Not stated. They are part of the written spec.

Contributors are scored weekly on acceptance rate, which is the share of their clips that pass validation. A contributor whose rate falls can be re-briefed or re-certified before more data is collected from them.

Acceptance-rate targets and the actions taken at a given score are Not stated. The site states the weekly scoring and the validation step it rests on.

Every batch is validated against signed accept criteria before it ships, with a reject log. Rejected items are recaptured at our cost, and billing is per validated hour.

A human reviewer also confirms the privacy pass before anything ships. The checks and thresholds used inside validation are set in the spec.

Real work can carry things a script leaves out: the order a worker actually follows, interruptions, tool changes and the pace of a normal shift. Whether those matter depends on your policy, and the spec decides which of them to keep.

The site does not state a measured effect of real-worker data on model performance, so no such claim is made here. Test it in a Spec Pilot against your own evaluation.

Yes. The public Egocentric-10K dataset card describes itself as the first dataset collected exclusively in real factories, with each worker having a folder of clips and a calibrated camera intrinsics file. It is a published dataset, not a collection service, so the spec, tasks and labels are fixed by its authors.

A custom collection starts from your spec instead, so the tasks, settings and labels are the ones your policy needs.

The site does not state contributor counts, recruitment channels, the working relationship between contributors and the sites they record in, partner names, or the countries a program draws from. It also states no pay figure on this page.

Those details are confirmed per engagement. The table below lists each one and where it is settled.

Every episode links to a signed contributor release through a documented consent chain. Faces, plates and screens are detected and irreversibly blurred at ingest, and a reviewer confirms the result before anything ships. The anonymization record is delivered with the data.

The site states no further privacy rule specific to real-worker capture, so this page makes none.

Where coverage is already deep, the first validated batch lands in a median of 48 hours from signed spec. Thin environments typically take three to four weeks, because vetting contributors there is the bottleneck, and we quote that up front.

On-demand top-ups run against the same spec and accept criteria after each training run.

You describe the work, the setting and the clips your policy needs. We turn that into a written spec covering tasks, environments, capture tier, labels and accept criteria, then quote reach, timeline and price before capture begins.

A Spec Pilot, a small first batch, can prove the spec and the contributor checks before you scale. Every collection ships under a perpetual, buyer-owned commercial license.