What Is Responsible Egocentric Data Collection?

Egocentric (first-person) video is collected by putting a camera on a real person doing a real task in a real environment, which means the footage inherently involves people: the contributor wearing the rig, and often bystanders who happen to be nearby — a family member in another room, a coworker who walks past, a stranger visible through a shop window. "Responsible" collection is the set of practices that make that footage defensible to use commercially rather than a liability waiting to surface after a model has shipped. In practice it breaks into four parts: documented consent from the people who agreed to appear, de-identification of the people who did not, a provenance record that lets a buyer verify both after delivery, and a licence that makes clear who owns the resulting dataset and what they can do with it.

None of these four stand in for each other. A written release does not substitute for blurring an un-consented bystander's face; a blurred bystander does not substitute for the contributor's own signed release; a clean consent chain does not substitute for a clear commercial licence; and a permissive licence does not retroactively fix an undocumented consent gap underneath it. A dataset is only responsibly collected if all four hold at once, for every delivered episode — not as a policy statement, but as artefacts a buyer can actually check.

This is general information about how responsible collection is typically structured for commercial egocentric data, not legal advice. Confirm the specific requirements for your jurisdiction and use case with counsel.

Consent for the person wearing the camera has to be obtained before capture, not inferred afterward from the fact that they showed up. On Firsthand, that means a written release is signed, the contributor's identity is verified, and the terms of use — including that the footage may be used to train commercial models — are agreed in advance and recorded in Schedule A alongside the agreed rate. See consent requirements for AI training data for a fuller breakdown of what a consent chain needs to cover and how it differs from a data licence.

Identity verification is part of the same step as the release, not a separate optional check: a signed release that cannot be tied to a verified person is not a usable artefact if a claim is ever raised, because there is no way to confirm the consent actually came from the person recorded. Contributors can also withdraw: on Firsthand, withdrawal requests are honoured for undelivered footage, but delivered footage cannot be recalled — and that limit is stated in the release itself, rather than discovered later.

A head-mounted camera captures whoever happens to be in the wearer's view, and most of those people never signed anything. Protecting them starts with site permission — bystander signage posted where a site is semi-public — and continues with an automated detect-track-blur pipeline applied at ingest: faces, licence plates and legible screens are detected per frame, tracked across frames so coverage survives turns and motion, and irreversibly blurred into the delivered pixels. Spoken names, addresses and numbers are scrubbed from the audio track the same way. See anonymizing first-person video for the full mechanics of why egocentric footage is harder to de-identify than fixed-camera video.

Automated blurring is checked, not trusted blind: a human reviewer confirms that no un-consented person is identifiable in any frame of a finished episode, and any clip where coverage cannot be confirmed is discarded rather than delivered with a residual risk. That review step also catches a case blurring alone cannot: a distinctive kitchen, a visible address on mail, or a company badge left in frame can be identifying even with every face blurred, so the reviewer looks at the scene as a whole, not only the faces in it.

Consent and de-identification are only useful to a buyer if they survive delivery as checkable records, not as a one-time assurance made during a sales call. Provenance is what makes that possible: a per-episode data card listing the consent artefact IDs and review record behind that specific footage, so a buyer can point to exactly which release and de-identification pass covers any given clip rather than relying on a provider's word after the fact. See provenance, consent and licensing for the underlying release language and the full de-identification pipeline, and dataset provenance explained for how these fields fit into the broader provenance record a buyer should expect.

Provenance matters more, not less, the further the data gets from the point of collection. A model team evaluating a dataset months after delivery, or a legal team responding to a claim years later, has no way to independently reconstruct whether a given clip was properly consented and de-identified — they can only check the record that was made at the time. A dataset without that record is not necessarily improperly collected, but it is unverifiable, which is functionally the same risk from a buyer's side.

Responsible collection extends to how the people generating the footage are treated, not only to the footage itself. On Firsthand, contributors are paid hourly for all time in session — including setup, calibration and breaks — rather than per accepted episode, because payment contingent on acceptance creates exactly the wrong incentive on a marginal take. Payment terms are net 7 with no clawback on rejected footage, equipment is supplied rather than bought by the contributor, and the acceptance bar (calibration discipline, spec literacy, consent handling) is published rather than applied opaquely. See /for-collectors for the full published standards and rates.

Responsible consent and de-identification still leave an open question for a commercial buyer: what can they actually do with the resulting dataset? On Firsthand, the commercial licence is buyer-owned, perpetual, irrevocable and worldwide by default — covering training, evaluating, fine-tuning and deploying models, including commercial models, with no royalty or revenue share on derived weights. The consent chain behind every delivered episode is warranted, and claims arising from it are indemnified, rather than left as a disclaimed assumption. See /provenance for the full licence terms and a comparison against several public egocentric dataset licences, several of which are non-commercial or research-only despite sometimes being assumed otherwise.

Consent and licence answer different questions and neither substitutes for the other: consent establishes that the people in the footage agreed to be recorded and to the stated use; a licence is the separate grant defining what a buyer can legally do with the resulting dataset. A dataset can carry a permissive commercial licence and still have an undocumented consent problem underneath it, which is exactly the gap a documented consent chain is built to close.

Because every piece of this is meant to be evidenced rather than asserted, a buyer evaluating a collection partner can ask for the artefacts directly instead of accepting a verbal assurance. See the AI training data rights checklist for a fuller due-diligence list covering consent, licence and provenance together before training on any dataset; the table below focuses specifically on what a good answer to a responsible-collection question looks like.

Responsible practices add steps — a release, site permission, automated de-identification, human review, a provenance record ��� but they are built into the collection pipeline rather than bolted on afterward, so they do not have to trade off against turnaround for a custom, on-demand collection. A buyer commissioning a new egocentric collection for a specific skill, environment or embodiment can specify the consent, de-identification and licence terms up front as part of the spec, the same way they specify streams, labels and formats, so responsible handling is delivered as part of the batch rather than negotiated after the fact. See custom egocentric video data collection for how a commissioned collection is specified end to end.