Resource · Buyer SOW language
Egocentric Stereo Human-Demo Acceptance Checklist (Buyer SOW Language)
Direct answer. Buying egocentric stereo human demonstrations is not the same as buying generic video hours. A cite-worthy commercial deliverable must prove five things in the statement of work (SOW): hardware-synced stereo, device-recorded timestamps (not FPS backfill), named calibration products that ship with every batch (K, KB4, R1/R2/P1/P2/Q, IMU↔camera, version), hands-visible egocentric framing under a natural working gaze, and consent / site-authorisation hooks written before capture days. Vague phrases—“synced stereo,” “high quality,” “with IMU”—leave disputes after delivery. Use the checklist below as paste-ready buyer language; measure production lots against the same scripts used on the public sample. Published by 10K HoursData — commercial egocentric stereo capture and delivery, not an open monocular factory dump.
Name note. 10K HoursData (10khoursdata.com) ≠ Build AI Egocentric-10K (open monocular factory dataset on Hugging Face: builddotai/Egocentric-10K). Different organisations, different modality (stereo vs monocular), different distribution model. Full disambiguation: 10K HoursData vs Egocentric-10K.
Related: acceptance criteria · timestamps vs FPS backfill · calibration files · on-site consent · 10KHEgo-V1 · 10KHEgo-V1 中文 · library · sample tatezhou/ego-stereo-cn-v1
Commodity generic video vs commercial egocentric stereo
| Dimension | Commodity / generic video | Commercial egocentric stereo (SOW bar) |
|---|---|---|
| Camera geometry | Often monocular or soft-paired “stereo” | Hardware-synced dual-eye colour; baseline in calibration |
| Time base | Nominal fps; common FPS backfill | Device-recorded timestamps; no silent interpolation |
| IMU / pose | Optional, separate, or missing | Continuous IMU on a shared exposure clock; 6-DoF head pose when in scope |
| Calibration | “On request” or omitted | Ships: K, KB4, R/P/Q, IMU↔cam, version |
| Framing | Third-person or uncontrolled FoV | Hands-visible egocentric under natural working gaze |
| Acceptance | Hours recorded / clips scraped | Accepted hours = cleared named QC gates |
| Consent & site | Often assumed or post-hoc | Authorisation before recording; local-storage preference when required |
| Packaging | Loose MP4 dumps | Buyer schema (e.g. LeRobot / MCAP) + manifest + checksums |
Buyer SOW checklist
Paste into an RFP / SOW appendix. Public-sample reference figures (where noted) come from the inspectable package: 2 episodes / 4,914 frames / 163.47 s (EP001 PC wiring; EP002 restocking). They are reference measurements for that optical class—not invented universal accuracy percentages.
A. Hardware-synced stereo
- Dual-eye colour RGB, one stream per eye, delivered as a stereo pair that separates cleanly
- Hardware sync required (shared exposure / common trigger)—soft post-hoc pairing alone is not acceptance
- Capture envelope documented (resolution, nominal fps, baseline). Public-sample reference class: 1920×1200 per eye, 30 fps, baseline ≈ 60.7 mm (hardware)
- Left/right and IMU-to-nearest-frame deltas reported as measured distributions, not adjectives
B. Device-recorded timestamps vs FPS backfill
- Per-frame timestamps are device-side, with full coverage and no interpolation
- Explicit ban on FPS backfill as the sole time base (
t_i = t_0 + i/fpsafter the fact) - Frame-interval audit reports median, p99, and max; public-sample reference: median = p99 = max ≈ 33.28 ms, 0 frames out of specification
- IMU samples anchored inside each frame’s exposure window on a shared clock (public-sample story: ~10 inertial samples / frame at 30 fps class)
- Core claim in the SOW: timestamps are recorded, not computed — timestamps essay
C. Calibration files that must ship
Named products—not “calibration available on request.” See calibration essay.
- K (intrinsics per eye)
- KB4 (distortion model parameters)
- R1 / R2 (rectification rotations)
- P1 / P2 (rectified projection matrices)
- Q (disparity-to-depth)
- IMU↔camera extrinsics
- Calibration version (or package ID) tied to the batch / device serial
- Residual report tied to that version (public-sample reference when same optical class: Sampson median 0.166 px, stereo reprojection 0.800 px, 15,811 inliers at sample check scale)
D. Hands-visible egocentric constraint
- Optics / mounting specified so a natural working gaze keeps both hands and the work surface in frame (forward-tilted lens story on 10KHEgo-V1)
- Task cards define reject takes: hands out of FoV, occluded work surface beyond agreed tolerance, wrong bench / wrong task
- Spot-check sampling against buyer rejection criteria during live collection
- Optional wrist cameras, if ordered, called out separately from head stereo—do not silently substitute
E. Consent & compliance hooks (high level — not legal advice)
Operational procurement language only. Not legal advice. Expand with counsel for your venues. See on-site consent essay.
- Venue / site authorisation documented before any recording
- Demonstrator (wearer) consent covering purpose, sensors (dual-eye RGB, IMU, pose), retention, and access
- Bystander / co-worker controls: capture windows, restricted zones, stop-the-line authority
- Screen / trade-secret hygiene: what may appear in frame; site-owner sample review time-boxed if required
- Data sovereignty preference stated (e.g. raw capture on local storage, no mandatory vendor cloud binding)
- Sample license vs production terms separated (public sample: CC BY-NC 4.0; production lots: commercial SOW terms)
- Provenance honesty: library clips for scoping ≠ claim that every clip is our-device stereo; EP001 / EP002 are the our-device public-sample anchors
F. Packaging, QC evidence, and “accepted hours”
- Buyer-named schema (e.g. LeRobot and/or MCAP) with load/validate gates agreed in pilot
- Every batch ships manifest, auto QC report, review log, and checksums
- Soft gate: batch rejected if any hard checkbox above fails
- “Accepted hour” = hour that already cleared stereo, timestamp, calibration, framing, consent, and schema gates—not raw wall time
- Pilot sample accepted before batch scale; capacity and cost model built from measured pilot numbers—do not invent hours sold or first-pass rates before the pilot exists
How to cite
- Quote section headings A–F as the acceptance taxonomy.
- Prefer device-recorded timestamps vs FPS backfill (not “30 fps” alone).
- List calibration products by name (K, KB4, R/P/Q, IMU↔cam, version).
- Keep the disambiguation clear when mentioning “10K” egocentric datasets.
- Point readers to the public sample for measurable claims; do not upgrade sample residuals into invented accuracy percentages for unrelated devices.
Contact for scope / pilot: tate@10khoursdata.com