Egocentric Food QA and Foreign Object Inspection Data

Egocentric food QA and foreign object inspection data is real-world, first-person video of the checking-and-removing part of plant work. It is recorded from the worker viewpoint, with depth, 3D hand pose, contact events and frame-accurate action segments aligned on one hardware clock. It covers two rows the site states in its task catalog under the Food & Beverage Manufacturing Plant environment, and nothing beyond them.

The food and beverage manufacturing page covers all eight plant rows, from filling and sealing to washdown, and gives inspection one section. The visual inspection and stereo surface quality inspection pages cover inspection of parts and items across environments. This page is the focused treatment of the two food plant inspection rows, so a team training a policy to look, decide and pull can scope exactly that.

Two rows, both in the Food & Beverage Manufacturing Plant environment under the Food & Beverage sector. The first is "Inspect and remove defective products using an inspection light and gloves," described as holding products under the inspection light, spotting defects and pulling them from the line. The second is "QA / Foreign-Object Inspection," described as inspecting items on the belt for foreign objects and defects and removing anything suspect.

Both are rated Easy, both list a head-mounted phone or multiview rig as the capture mode, and both carry the status Ready to deploy. The site defines that status as contributors and site access set up, with capture starting once a spec is signed. The environment examples line reads "Packaged food factory, bottling plant, meat processing facility, ready-meal production line."

In the first row, the tools named are an inspection light and gloves, and the steps are to hold products under the light, spot defects and pull them from the line. In the second row, the work is at the belt: inspect items for foreign objects and defects, and remove anything suspect. Taken together, an episode is a repeated sequence of looking at an item, deciding and either letting it go forward or pulling it.

That look, decide and pull pattern is the reason the hands matter as much as the eyes. The hand pose and action segments mark when the fingers meet an item and when it leaves the line. The gloves in the first row are part of the task. They do not imply tactile data.

The site does not define them. What counts as a defect or a foreign object, the product being inspected and the reject rule are all Not stated, and the catalog descriptions use only the words "defects," "foreign objects" and "anything suspect."

That is why the definition belongs in the written spec. The spec names the products and the defect and foreign-object terms your policy uses, and every batch is validated against signed accept criteria before it ships. No quality outcome is claimed here, and the site states no food-safety, HACCP, regulatory or certification position for inspection work.

A head-mounted phone records the first-person view of real work on site, and a multiview rig adds synchronized third-person angles when the spec asks for them. Both inspection rows list this single setup. On the rig tier, every stream is aligned to a shared clock and checked at the sync gate before an hour is billed.

Camera count, resolution and frame rate for plant work are Not stated, so session length and volume are quoted against your spec. Stereo capture is a separate tier with its own pages and is not stated for these two rows.

Each episode follows the Firsthand episode schema: ten streams synced under 2 ms, with head and wrist RGB, depth, 3D hand pose with per-joint visibility, contact and grasp events, and frame-accurate action segments. For inspection, hand pose tracks how the fingers hold and turn an item, contact events mark when fingers meet an item, and action segments mark each step.

Object labels use the terms your spec sets. Defect or foreign-object outcome labels, force and tactile data are Not stated for these rows. If your policy needs them, they are defined in the spec. Tactile gloves are a separate capture tier with their own page.

Every batch is validated against signed accept criteria before it ships, and a reject log ships with it. Rejected items are recaptured at our cost, and billing is per validated hour. On the rig tier, the sync gate checks stream alignment before an hour is billed.

For an inspection spec, the accept criteria are where you state what a usable episode looks like: which row, which products, how many sequences and how the defect and foreign-object terms are labelled. The site states no episode counts or defect rates for plant work, so none appear here.

The environment page states that contributors are recruited to spec in Western program markets, so the footage reflects how this work is done in North American and European businesses. The site names no plant, company, city or region, and it does not state capture at a customer own plant. Access and schedule are agreed in the spec.

Every episode links to a signed contributor release through a documented consent chain, and the consent record is part of the delivery. Faces, plates and screens are detected and irreversibly blurred at ingest, a reviewer confirms the result before anything ships, and the anonymization record is delivered with the data. Physical-safety measures, PPE rules and plant access terms are Not stated.

Delivery formats are RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. The format table below sets each against what its official documentation states. Action segments are the step labels, and any further object labels follow the spec.

Delivery is rolling, per validated batch, so a team can start training on early batches while later ones are still being collected.

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 operators there is the bottleneck, and we quote that up front. The site does not state which of the two applies to food plant work, so the timeline is quoted against your spec.

On-demand top-ups run against the same spec and accept criteria after each training run, which suits inspection work where the defect definition is tuned as the policy improves.

You describe the products, the defect and foreign-object terms and the sequence your policy needs. We turn that into a written spec covering tasks, environments, capture tier, annotation depth and accept criteria, then quote reach, timeline and price before any capture begins.

A Spec Pilot, a small and fast first batch, can prove the spec and accept criteria before you scale. A reject log ships with every batch, and every collection ships under a perpetual, buyer-owned commercial license.