How to Buy Egocentric Video Data

Teams sourcing egocentric (first-person) video have three broad routes, and most programs end up using more than one. The first is a public research corpus such as Ego4D or EgoDex — free, large, and useful for early experimentation, but captured for research goals that may not match your task, usually without a commercial license. The second is licensed, off-the-shelf data from a vendor that already has a relevant library — faster to access than a fresh capture, but you are still bound to whatever conditions, environments and label set the vendor already shot. The third is a custom collection built against your own written spec — slower to start than grabbing a public file, but the only route that lets you control exactly which skills, scenes, objects, contributors and labels show up in the data.

These routes are not mutually exclusive. A common pattern is to prototype on a public corpus, then commission a custom collection once the policy is working and needs targeted coverage the public data does not have.

A public corpus is enough when your task closely matches what it already captures, when you do not need a commercial license, and when you are still in early experimentation rather than shipping a product. Ego4D and EgoDex are useful references for benchmarking and for early-stage model work, but they were collected for research purposes with their own license terms, consent scope, and condition coverage — not for a specific buyer’s commercial skill spec. If your task is a close match to what a public corpus already contains and your license needs are research-only, that is the fastest and lowest-cost path.

The limits show up quickly once a project moves toward a shipped product: public corpora rarely carry a usage grant for training a model you then sell, they do not let you add scenes or objects specific to your deployment, and their labels stop wherever the original research team stopped. At that point, a licensed or custom route becomes necessary.

Commission a custom collection when your task needs coverage a public or off-the-shelf set does not have, when you need a commercial license that lets you train and ship a model you own, when you need manipulation-grade labels such as 3D hand pose, metric depth and frame-accurate action segments rather than just RGB and narrations, or when you need documented per-participant consent you can produce on request. A custom collection is also the right route once you know the specific skills, objects, environments and contributor profile your policy needs and you want a provider to fill a coverage histogram against that spec rather than hand you whatever they already have on the shelf.

• Your task needs scenes, objects or contributor profiles a public corpus does not cover
• You need a commercial license that lets you train and ship a model you own, not just a research license
• You need manipulation-grade labels: 3D hand pose, metric depth, frame-accurate action segments
• You need documented, per-participant consent you can produce per batch, not a general policy statement
• You need the provider to fill specific gaps in a coverage histogram rather than reuse a generic library

A spec for an egocentric purchase should state the skills or tasks to capture, the scenes and environments they happen in, the objects involved, the contributor profile needed, and the diversity or coverage targets across those dimensions — the same way a skill spec drives a custom collection program. It should also state which streams you need (egocentric video, depth, IMU, audio where relevant) and which labels you need delivered with the footage (hand pose, action segments, object or scene tags). Writing this down before you talk to any provider is what lets you compare routes and vendors on the same terms rather than against a demo reel.

The spec is also what a provider should pilot against. Without a written spec, there is no way to judge whether a sample batch — public, licensed or custom — actually matches what your policy needs, and no way to hold a vendor accountable for the gap.

Run a small paid or free pilot batch against your written spec before committing to volume, whatever the sourcing route. For a custom collection, this mirrors the spec-to-pilot step of a managed collection program: a provider captures a small batch, you review it against your spec and acceptance criteria, and only once it passes do you scale to the full order. For a public or licensed dataset, the equivalent step is pulling a representative sample and checking it against the same spec — sync quality, label depth, condition coverage, and license terms — before you build a training pipeline around it.

A pilot is the point to catch problems cheaply. Discovering a license, sync, or coverage gap after a model has already been trained on the full dataset is far more expensive than catching it in a small sample.

Every contributor whose footage you buy should have a documented, written release granting the specific rights your project needs, and a provider should be able to produce that consent artefact per batch rather than pointing to a general policy statement. Footage should also go through anonymization for bystanders and sensitive elements that are not part of the task — faces, license plates and visible screens are the categories this kind of review typically covers — before it is delivered. Ask where the provenance trail lives: who the contributors were (in aggregate, not identifying detail), when and where the batch was captured, and how the consent-to-footage link is maintained.

Public corpora carry their own consent and provenance terms set by the original research team, which is one more reason to read a public dataset’s documentation closely rather than assume it covers your commercial use case.

Confirm, in writing, that the license covers training a commercial model and that you — not just the provider — own the delivered data and whatever model you derive from it. A non-commercial or ambiguous research license will silently block a shipped product even if the footage itself is exactly what you need. For a custom collection, the commercial terms should be clear from the spec stage onward, with a perpetual, buyer-owned license on the delivered batch rather than a revocable or time-limited grant.

Pricing itself is driven by the scope of the spec — environments, label depth, contributor diversity, and volume — rather than a flat catalog rate; see /custom-collection/pricing for the specific factors that move cost on a custom order.

Ask any vendor — public, licensed or custom — which of these formats they can deliver in, and whether streams (video, depth, IMU, labels) arrive pre-synced in that format or need to be aligned yourself after delivery.

• Write a spec: skills, scenes, objects, contributor profile, diversity targets, needed streams and labels
• Decide the sourcing route: public corpus, licensed off-the-shelf set, or custom collection — or a mix
• For public or licensed data, pull a representative sample and check it against the spec and its license terms
• For custom collection, run a small pilot batch against the spec before committing to volume
• Require measured sync error, label depth matching the spec, and a visible reject policy
• Confirm per-participant written consent is documented and producible per batch
• Confirm anonymization of bystanders, faces, plates and screens where not part of the task
• Get the commercial license and ownership terms in writing before scaling the order
• Confirm the delivery format (RLDS, LeRobot, WebDataset, HDF5/zarr, .rrd) fits your training pipeline
• Only then scale from pilot to the full volume your program needs