Egocentric Hair Coloring Data

It is first-person video of a person doing hair coloring work in a salon, 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 across the sequence the catalog row names: mix colour, section the hair, apply with brush and foils, time and rinse. It is labeled so each step can be sliced out. Where the site states nothing, a table below says "Not stated".

One. The row is named "Hair coloring" and sits in the Hair Salon, Barbershop & Beauty Studio environment of the Beauty & Personal Care sector. Its description reads "Mix colour, section the hair, apply with brush and foils, time and rinse." It is rated Hard, its capture setup is "Phone (head-mounted) / multiview rig", and its status is Active collection.

The site’s meaning for Active collection is that recording in this environment is running now, so a spec can join an existing capture stream. It does not say how many hours exist or have been delivered for this row, so this page claims none. This page names only the tools and steps that appear in the row: colour, a brush, foils, sectioning, timing and rinsing. The first table below reproduces the row as the catalog states it.

The hair salon and beauty page covers the whole environment, 13 rows across hair, nails, skin and workspace sanitation, with hair coloring as one row in a table. This page centres that single row and gives it its own step map, failure modes and not-stated list.

The environment also lists separate hair rows: "Cut and style hair using scissors, clippers, and combs", "Hair Cutting & Styling", "Hair styling" and "Wash and blow-dry hair using shampoo stations and a blow dryer". They are separate listings and are not described here. The comparison table below sets them side by side.

An episode is one contributor performing one task from the skill spec, recorded with ten hardware-synchronized streams that line up within the 2 ms ceiling the site documents. The rig is a head-mounted global-shutter 4K/60 camera with a 120 degree lens, two wrist-mounted 1080p/60 cameras, an active-stereo depth sensor at 848×480/30 and 200 Hz 6-DoF IMUs.

Every episode carries int64 nanosecond timestamps per stream, camera intrinsics and extrinsics, hand joints for both hands at 21 joints each, 24 body joints, instance masks and action segments. The annotation record also carries contact events. The one language label the site states is a task instruction for each episode.

The site does not say what counts as the object in a contact event, or how long the row runs, so settle both in the spec and the quote.

The unit of labeling is the action segment, with a start, an end, a verb and a noun, annotated to the frame and human-reviewed against a taxonomy of 97 verbs and 512 nouns. A spec that says "colour the hair" gives you one long segment. A spec that names mix, section, apply, time and rinse gives you a sequence you can slice by step.

The step table below sets each step in the row against the stream that carries it. The site states no label for the outcome of a colouring, such as how even the application is, so any outcome label would be written into the spec.

The failure modes the site already describes apply here. A brush, a foil or a held section can hide the fingers, which is why hand pose carries a per-joint visibility flag and contact events mark when a hand meets an item. The site states that episodes are checked against reject criteria before anything ships.

Those criteria are a sync breach beyond the ceiling for more than three consecutive frames, calibration drift beyond tolerance, the manipulated object out of frame for more than 15 percent of the episode, more than 35 percent invalid depth pixels in the manipulation region, an undocumented bystander, or a mismatch against the named skill spec. A human reviewer then checks each episode against the task spec. Rejected episodes go to a reject log that ships with every batch and are recaptured at our cost.

Data is delivered in RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd. Converters ship as readable source, not a binary, so you can retarget the schema to your own loader.

You receive a perpetual, irrevocable, buyer-owned commercial license, the consent artefacts for the batch and a data card that lists jurisdictions of capture. Billing is per validated hour, meaning footage that passed sync and QA against your spec; rejected episodes and recaptures are never billed.

Only what the site already states applies. Every participant signs a written release granting commercial training rights before capture begins. Faces and licence plates are detected and irreversibly blurred, a human reviewer confirms that no un-consented person is identifiable in any frame, audio is scrubbed of identifying speech, and a clip where a bystander cannot be de-identified is discarded rather than shipped. Consented subjects stay in frame.

The site does not state how a salon client is handled beyond that general bystander rule, or whether a client is in frame for this row. Settle both in the spec. This is not legal advice, and the executed agreement controls.

The site states one published timeline: from a signed spec to a first validated batch, a median of 48 hours in environments where coverage is already deep, and typically three to four weeks where coverage is thin because operator vetting is the bottleneck.

The row is in Active collection, but the site does not say which timeline applies to it and promises nothing faster, so ask when you request the quote.

It supplies first-person human demonstrations of a fine, multi-step hand task, with the hands, the tool and the work in view. The site’s guides explain that human footage has no robot joint-angle log, so wrist pose supplies the coarse trajectory and hand pose the fine grasp shape.

Firsthand captures people, not robots. Using these episodes for a specific robot means your own retargeting step, and the site states no robot result, benchmark or teleoperation data for hair coloring. It states no force or tactile reading either; contact is described only by contact events.

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

A good brief names the setting, the order of steps, how each step starts and ends, and what counts as a reject. The site’s skill-spec guide walks through that structure. Then request a custom dataset.