Validated Egocentric Video Data

Validated egocentric video data is first-person footage — head- and wrist-mounted video, with its depth, IMU, and annotation streams — that has already passed a defined acceptance test before it counts toward a delivery. That test checks whether the footage is in sync, whether calibration held, whether it covers a condition the spec actually needs, and whether consent was documented for every participant in frame.

That is a different thing from simply having a lot of egocentric footage. A validated hour is 60 minutes of footage that has passed QA against your spec; a raw hour is unfiltered recording time before that check runs. Teams routinely discard most bulk egocentric footage as unfit once it is checked against a real spec, which is why budgeting and billing in validated hours keeps a coverage or volume number honest.

A raw hour says nothing about whether the footage is in sync, whether calibration held, whether it covers a condition your spec needs, or whether consent was documented — only that a camera was rolling. A validated hour has already cleared all of those checks: sync within the measured ceiling, calibration solved and within drift tolerance, coverage cell confirmed against the spec, and consent documented.

On this site, billing is per validated hour, not per raw hour. A take that fails QA is not billed; it is rejected and recaptured at Firsthand’s cost. That means a quoted coverage or hour number is already the number you can train on, rather than a gross figure you still have to filter down after delivery.

Four gates run before footage counts as a validated hour. First, task acceptance: the footage is checked against the acceptance test written into the skill spec, not a reviewer’s general impression. Second, stream sync: every stream — head video, wrist video, depth, and IMU — is hardware-triggered onto one PTP clock domain, and any episode breaching the measured sync-error ceiling is rejected. Third, calibration: every episode ships its intrinsics, SE3 extrinsics, and RMS reprojection error, held below 0.28 pixels, so geometric soundness is recorded per episode rather than asserted once for the whole dataset. Fourth, coverage and consent: the footage is checked against the spec’s condition matrix, and consent is validated by confirming each episode links to a signed contributor release.

All four gates are documented per episode, not summarized once across a whole batch, so a buyer can check any single episode against the same rule the reviewer applied.

Footage is rejected when it misses the acceptance test written into the skill spec, when a stream breaches the sync-error ceiling, when calibration does not hold within drift tolerance, when a coverage cell is not actually filled by what was captured, or when consent or de-identification cannot be confirmed. Rejected footage is logged with a reason and, where the spec calls for it, recaptured at Firsthand’s cost — it is not billed.

If the same failure keeps recurring, the more useful fix is often revisiting the spec rather than repeatedly recapturing — for example, if the geometry that must stay in frame turns out to be impossible to hold in a particular environment. That decision is made with you, and the data card records it.

A second reviewer covers a sampled or full pass depending on the annotation’s density and stakes. Action segments ship as {start_ns, end_ns, verb, noun} with nanosecond timestamps aligned to every other stream, so a reviewer can check that boundaries land on the actual contact moment rather than drifting into transit. Hand pose ships as handJoints[2][21][3] with a visibility flag per joint, which lets a reviewer confirm that occluded joints are marked rather than invented — one of the most common silent failures in pose labeling.

Stream timing is checked because every stream carries its own PTP-domain timestamp, so sync between head video, wrist video, depth, and IMU can be verified rather than assumed. Calibration is checked per episode against the intrinsics, SE3 extrinsics, and RMS reprojection error it ships with.

Every episode links to a signed contributor release, and that link is checked as part of validation — consent is confirmed, not asserted after the fact. Faces, license plates, and screens are detected, tracked, and irreversibly blurred at ingest, with a human reviewer confirming the result before anything ships.

The de-identification pass fails closed: if a reviewer is unsure a region is fully covered, the episode is held or discarded rather than shipped. That is part of what makes validated footage usable for a commercial model rather than footage with unresolved privacy exposure.

Every delivery carries a data card recording which coverage cells were captured and what was rejected, so coverage against the spec’s condition matrix is documented rather than summarized as a single hour count. Each episode’s meta.json also carries its own reject_flags, keeping the record at the episode level, not just the batch level.

Because the data card ties validation back to the spec’s condition matrix, a buyer can see exactly which conditions are covered, which were thin, and which cells were rejected and why — the same record Firsthand uses internally to decide whether to recapture or revisit the spec.

A Spec Pilot runs a small first batch against the agreed spec and is reviewed against every acceptance axis — task, sync, calibration, coverage, and consent — before volume scales, with the same reject log and validated-hour billing a full collection gets. That means a coverage gap or an acceptance test that does not hold up on real footage is caught in the pilot, not discovered after a much larger collection has already shipped.

Running the pilot against the exact acceptance rule the full order will be billed against is what makes the pilot a real test of the spec, rather than a smaller sample that could pass or fail on different terms than the main collection.

Start with a written skill spec describing the tasks, environments, objects, and streams your model needs, including the acceptance test a reviewer will apply. Most orders begin with a Spec Pilot — a small, fast first batch that proves the spec and acceptance criteria hold up on real footage before scaling to the full order.

Where coverage for the requested tasks and regions is already deep, the median time from signed spec to first validated batch is a median of 48 hours. Where it is thin, turnaround runs three to four weeks while recruiting expands in parallel across more regions. Either way, what you receive and what you are billed for is validated hours, with the reject log visible alongside it.