About Fourthuman
We build the capture layer for physical AI.
Fourthuman turns egocentric video of real physical work into annotated training data for humanoid and VLA models. We capture rather than generate, and we build in India.
We compete on two things. The first is provenance you can verify: every clip we deliver is cut from a Recording that carries a consent receipt. The second is bespoke capture at an Indian cost structure: you define the brief, review a sample, then scale. Data quality is table stakes, so it is not on the list.
Standing
What we have proven
Every row below resolves to a committed artifact. The full register of published figures is on Numbers.
| What | Standing | Where it comes from |
|---|---|---|
| End-to-end annotation runs | 1 · Measured | One committed clip, annotated all the way through: 1,843 frames and 5 timeline tracks. It is on the Reference clip page, with the JSON. |
| Per-hand confidence on that run | 0.8712 mean · Derived | Arithmetic mean over 294 kinematic entries — one score per detected hand, not per joint — on 185 sampled frames. It aggregates the tracking model's own confidence output, which makes it derived, never measured. The lowest score is 0.508. |
| Delivery formats implemented | 5 · Measured | Detected in the services directory by the writer or library each format requires, not asserted. An export bench reads every one back before we quote it. |
Method
How a brief becomes a dataset
These stages take a Capture Brief to a delivered dataset, in order. Each one links to how it works.
Capture
A Capture Partner records against the brief.
Redact
Faces the detector finds are blurred before any human review.
Segment
The Recording is cut into usable Clips in one pass.
Annotate
A model estimates 3D hand pose and confidence, both derived.
Deliver
The annotated data ships in the container the brief names.
Five containers ship. MCAP and WebDataset do not. What arrives, and in what container.
Ethical sourcing
How Capture Partners are paid
Capture Partner pay: from ₹229/hr of accepted capture, paid directly by UPI. What a brief pays above that floor depends on the brief.
Clear pay terms are meant to keep partners. A partner who stays learns the brief, and that keeps capture consistent across a long collection run.
A floor printed on a public page is a floor a partner can hold us to.
At the floor: ₹229 / hr
₹34,350
To the Capture Partner, paid directly by UPI.
Method: hours × the ₹229/hr floor. The result is a minimum, never an estimate of what a brief pays.
Not a price. It is the least a Capture Partner receives for those hours. What a brief costs a Commissioning lab is quoted against that brief.
Not a payroll record. The floor is policy; the total is arithmetic.
The company
A small team in India, running the whole chain
Fourthuman is an India-based training-data company. The people who scope a brief are the people who run the capture, the redaction, the annotation and the delivery — there is no handover from a sales team to a data team, because there is no sales team. Your first email to labs@fourthuman.ai is answered by someone who works on the pipeline it describes.
What we own end to end
Everything between a Recording and a delivered dataset is built and run in-house: the ingestion boundary and its face-redaction stage, the 3D annotation pipeline, the FH-Ego v1 schema, the export bench that verifies a format before we quote it, and the hash-chained consent ledger. The reference run linked from this page came off that chain.
On the other side of it sits how a brief is scoped and priced, the catalogue, Commissioning-lab relations, and Capture Partner recruitment and pay. The content-integrity rules this site is built under came out of an internal audit, and every published number is generated from a committed artifact rather than typed.
Pipelines · Schema · Consent ledger · Briefs · Partner relations
Redaction before review
Face redaction that fails closed
Sourcing egocentric video inside homes, clinics and factories means recording people who never agreed to be in a training set. Under India's Digital Personal Data Protection (DPDP) Act 2023, data principals hold consent and deletion rights, and the Act sets penalties of up to ₹250 crore for a breach of the security-safeguards obligation.
That is the exposure we engineer against. Every clip passes automated face redaction at the ingestion boundary, before annotation, before any reviewer opens the file. YuNet, OpenCV's neural face detector, runs where the host provides it; an older OpenCV cascade runs where it does not. Consent is granted per Recording and written as a receipt carrying a SHA-256 audit hash, and a Clip without an active receipt is refused at delivery. A Capture Partner asks to withdraw consent by emailing partners@fourthuman.ai.
Covered: human faces only — by YuNet where the host provides it, an OpenCV cascade where it does not — server-side at ingestion, before any human review.
Fails closed: if the detector cannot load, the clip fails the stage rather than passing through. There is no unredacted path to delivery.
Not EgoBlur: this is our own stage, not Meta's EgoBlur, and we do not restate EgoBlur's accuracy as ours.
Next
Bring us a brief
The useful first conversation is a Capture Brief: the task, the environment, the volume and the format you train against.