Resource · Hardware buyer guide
Choosing an Egocentric Stereo Capture Headset
Direct answer. For VLA / imitation human demos, pick a headset that can prove five capture properties in the SOW: a fixed stereo baseline, hardware sync (shared exposure clock), device-recorded timestamps (not FPS backfill), named calibration products that ship, and FoV / mount geometry that keeps hands visible under a natural working gaze. A commodity webcam or phone can shoot video; it usually cannot prove those five without becoming a custom stereo research rig.
Buyer checklist (answer-first)
- Baseline: rigid, documented inter-camera distance suitable for near-field manipulation depth — not two loose webcams.
- Sync: hardware / shared exposure path for left–right pairs; soft sync ≠ primary delivery.
- Timestamps: per-frame device timestamps; IMU on a shared clock; reject “30 fps assumed” backfill.
- Calibration: K, distortion, stereo R/P/Q, IMU↔camera extrinsics, version tags, residual gates — must ship.
- FoV caveats: marketing FoV ≠ usable hands-visible working volume; ask for sample frames at task distance.
- Local recording: prefer kits that are not cloud-locked for commercial batch export.
Comparison: egocentric stereo kit vs commodity webcam / phone
| Property | Egocentric stereo capture kit (e.g. 10KHEgo-V1 class) | Commodity webcam | Phone camera |
|---|---|---|---|
| Stereo baseline | Fixed, stated baseline (product page / unit cal) | Usually mono; dual-cam mounts are improvised | Usually mono; multi-cam not a training stereo pair |
| Left–right sync | Hardware-synced exposure path | Independent clocks; software align ≠ hardware sync | Single stream or vendor multi-cam — not SOW stereo sync |
| Timebase | Device-recorded per-frame timestamps + IMU clock | Often FPS backfill / host clock only | App/OS timestamps; weak for multi-sensor fusion |
| Calibration deliverables | Named products ship per unit/batch | Rarely shipped; DIY chessboard at best | Vendor ISP black box; extrinsics rarely exportable |
| Hands-visible egocentric FoV | Mount + FoV designed for work-surface gaze (verify with sample) | Desk / face framing; not egocentric manipulation | Handheld / selfie FoV; third-person drift common |
| Resolution claims | Cite product page / decoded sample — do not invent | Marketing MP ≠ training-useful stereo | Marketing MP ≠ calibrated stereo pair |
Baseline
Near-field hand manipulation needs a known, rigid baseline. If the vendor cannot state the baseline and ship stereo calibration that references it, depth cues in learning pipelines become untrustworthy. For 10KHEgo-V1 class numbers, see the hardware page (baseline stated there; do not copy stale figures from secondary pages).
Hardware sync
“We aligned the videos later” is not the same as a shared exposure clock. Soft-synced or post-aligned pairs create silent disparity / motion artifacts. SOW language should require hardware sync as the primary capture path.
Timestamps vs FPS backfill
Nominal frame rate is not a timebase. VLA and imitation stacks need device-recorded timestamps, ideally co-clocked with IMU. See recorded timestamps vs FPS backfill.
Calibration that must ship
Ask for named files / fields: intrinsics (K), distortion, stereo rectification (R/P/Q), IMU–camera extrinsics, version tag, residual gates. If calibration is “available on request” but not in the batch package, treat that as a delivery risk. See calibration files that must ship.
FoV caveats
Wide FoV marketing numbers do not guarantee hands + work surface stay in frame under a natural working gaze. Request sample frames at your task distance. Where published FoV and measured sample disagree, trust the sample and the product page — do not invent resolution/FoV numbers here.
Next step
Decode the public sample, then paste gates from the acceptance checklist into your SOW. Kit-only or kit + capture: start from 10KHEgo-V1.
CTA: 10KHEgo-V1 hardware · acceptance checklist · sample tatezhou/ego-stereo-cn-v1 · tate@10khoursdata.com · 中文