Egocentric Inventory Counting and Stock Check Video Data

It is first-person video of a person counting or checking stock, recorded from a head-mounted camera, with two wrist cameras for a closer view of the hands. The contributor is a paid, consented person following a written skill spec, not an actor and not a robot.

The point of the data is what the hands do while counting: reaching into a location, handling items, aiming a scanner and recording a total. It is labeled so each step can be sliced out. Where the site states nothing, a table says "Not stated" and the answer lives in the spec or the quote.

The task catalog names two rows that are explicitly about counting. "Count and reconcile shelf inventory using a barcode scanner and count sheets" is described as "Scan shelf locations, count items and record the totals on the count sheet." "Cycle Counting & Inventory" is described as "Count stock in assigned locations with a scanner and reconcile the numbers against the system." Both are listed as Easy, Ready to deploy.

Other rows sit next to counting but do not name it, so this page does not count them: order picking, receiving and unloading, backroom stock sorting and damaged-goods sorting. The second table lists them for context. Ask for a counting step inside any of them in the spec.

Both counting rows are listed under the Warehouse & Distribution Center environment, whose catalog examples are an e-commerce fulfillment hub, a distribution center, a cold storage facility and a freight warehouse. Both list a head-mounted phone or multiview rig.

The site lists no counting row in a store environment. Its store backroom rows, under Retail Store -- Back of House, cover sorting stock onto racks and bins, not counting it. If you need counting in a store, say so in the brief. The site does not otherwise state where capture happens or who hosts it, so the spec settles location and access.

An episode is one contributor performing one task from the skill spec, recorded with ten hardware-synchronized streams that line up within 2 ms. The site documents a canonical episode schema, and every delivery format is a view of it.

Hand pose is stored as 21 joints per hand in metres in the head frame, with a visibility flag for each joint. Contact events record when a hand meets an object. The site does not say whether a carton, a scanner or a count sheet counts as the object, so settle it in the spec.

Action segments carry a start, an end, a verb and a noun from a taxonomy the site states as 97 verbs and 512 nouns, human-reviewed. Every episode links to its contributor’s signed release by a consent artefact ID, and each batch ships with a data card.

The unit of labeling is the action segment. A spec that says "count the location" gives you one long segment. A spec that names scan, count, record and reconcile gives you a sequence you can slice by step.

Every episode is machine-checked, then reviewed by a human against its skill spec, and rejected episodes are recaptured at our cost. The site states no label for the number counted, the count accuracy or any variance against the system, and no force or tactile reading, so ask for any such label at the start.

Only what the site already states applies. Every contributor signs a written release, and each episode links to it. Faces and screens are detected and blurred irreversibly, a human reviewer confirms that no un-consented person is identifiable, and footage with bystanders who cannot be de-identified is discarded rather than shipped.

The site does not state how customers, other staff or screens showing stock data are handled beyond that general rule, so settle it in the spec before capture. This is not legal advice, and the executed agreement controls.

Data is delivered in RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, or as raw streams with metadata, with the episode schema as the canonical layout. QA runs once, against the canonical episode, before any format is cut. The formats table sets the site’s delivery next to each maintainer’s own documentation, linked under Sources.

The license is perpetual, irrevocable and worldwide, buyer-owned and non-exclusive by default, with exclusivity available as a separate term and the executed order form governing.

The site states one published timeline: from a signed spec to a first validated batch, a median of 48 hours where coverage is already deep, and typically three to four weeks where it is thin because operators have to be staffed. The site does not say which applies to the counting rows and promises nothing faster, so ask when you request the quote.

Collection is on demand, so a focused brief, such as one aisle of locations counted and reconciled against a list, is a normal starting point.

Start with the free sample pack to check the loader and label quality, then move to a Spec Pilot: a small, fast first batch against the written spec that proves the fill rate and reject rate on your conditions before you scale.

A good counting brief names the setting, the order in which locations are visited, how each location starts and ends, what counts as a reject and the privacy rule for the space. The site’s skill-spec guide walks through that structure. Then request a custom dataset.