Technical Field Note #25 · November 2026

How much of 2.6 hours of egocentric work video is usable? A per-second QC index: 85% sharp, about half with hands at work

By Hoonhoe Kim, Co-Founder & CTO, Origin Data Lab

Field Note #25 · Usable-time QC for egocentric stereo+IMU capture (ODL-EGO S1): one row per second, raw footage untouched

Buyers pay for usable hours, not recorded hours, so we measure usable time second by second. For 2.57 hours of real produce-portioning work recorded on 2026-11-08 (four sessions, 9,246 seconds), we built a QC index with one row per second and four checks: exposure from the image, camera motion blur from the gyroscope and exposure time, view direction from the accelerometer, and hand presence from a hand detector. Exposure passed in all but one second. Camera blur stayed within 8 px (90th percentile in the second) for 85.4% of seconds; median blur was 1.6 px at a median head speed of 19.6°/s. The camera pointed down at the work in 99.8% of seconds. A hand was detected in 56.7%. Combined, 49.2% of seconds (1.26 h) passed every check, 39–59% per session. Hand detection is the limiting check and misses hands (Field Note #23): in 60 reviewed seconds, 55% actually had a visible hand with a usable image versus 37% by the detector, so the true usable share is nearer 55% (about 1.4 h). Usable stretches are short: half last 2 s or less, and stretches of 10 s or more carry 34–47% of usable time. The raw footage is never cut; the index is stored next to it and applied only at export.

Published November 2026⏱ 11 min readStatus Preliminary · automated analysisOpen data GitHub · Hugging Face · Kaggle
85.4%Seconds with sharp, well-exposed imagescamera blur ≤ 8 px, 2.19 of 2.57 hVerified (automated)
49.2% (1.26 h)Seconds usable for manipulation (all checks)39–59% per session, detector-basedVerified (automated)
≈ 55% (≈ 1.4 h)Same, corrected for detector recall33 of 60 reviewed seconds (95% CI 42–67%)AI visual review, not human-verified
34–47%Usable time in stretches of ≥ 10 sper session; median stretch 2 sVerified (automated)

Why this matters

Recorded hours are not usable hours

Every egocentric recording contains time a buyer cannot use: the camera swinging during a head turn, the wearer walking to fetch a box, hands out of view. A vendor that reports recorded hours leaves the buyer to discover the difference. We want to report usable hours, define them explicitly and let the buyer change the definition.

This note builds that measurement on one real working day and shows what the numbers look like.

Setup

One row per second, four independent signals

Recording: 2026-11-08, produce portioning for school meals, four sessions (24, 33, 53 and 44 min, 9,246 s in total), R8 stereo+IMU rig (ODL-EGO S1) on the head, 1600×1200 at 30 fps, 500 Hz IMU on the same clock.

For each second: image exposure from one left-eye keyframe; camera blur from the gyroscope over each frame's exposure window (rotation rate × exposure × 482 px focal length, as in Field Note #15); camera pitch from the averaged accelerometer (as in Field Note #17); hands from MediaPipe with the full-frame and lower-region passes of Field Note #23. Thresholds are listed below and are parameters, not fixed truths.

Table 1. One row per second, four independent signals
Check (per second)SourceRuleSeconds passing
Exposureleft image, 1 frame/smean 40–200, < 10% saturated, < 20% black100.0%
Camera motion blurgyroscope × exposure time × focal length90th percentile within the second ≤ 8 px85.4%
Looking at the workaccelerometer (gravity) → camera pitch≤ −30° (pointing down)99.8%
Hand presentMediaPipe, full frame + lower work region (Field Note #23)≥ 1 hand found56.7%
Usable for manipulationall of the above49.2% (1.26 h)

Finding 1

Image quality is rarely the problem; hands are

Exposure failed in one second of 9,246. The camera pointed down at the work bench (pitch ≤ −30°) in 99.8% of seconds, median −66°. Camera blur was the main image-quality loss: 14.6% of seconds had a 90th-percentile blur above 8 px, caused by fast head turns; median blur was 1.6 px.

Hand presence decided the rest. With all four checks, 49.2% of seconds (1.26 h) were usable for manipulation learning, from 39.5% in the last session (more walking and fetching) to 58.9% in the longest one.

Horizontal bars showing recorded hours dropping from 2.57 to 1.26 hours as each QC rule is added
Figure 1. Usable time as each rule is added, 1-second resolution.
Table 2. Image quality is rarely the problem; hands are
SessionMinutesBlur OKHand foundUsable for manipulationMedian head speedMedian camera blur
09:322483.5%55.2%47.5%19.5°/s1.6 px
10:003384.8%54.5%47.8%18.1°/s1.5 px
10:495386.0%67.2%58.9%17.7°/s1.5 px
11:484486.1%46.3%39.5%22.5°/s1.9 px
All15485.4%56.7%49.2%19.6°/s1.6 px

Finding 2

Correcting for the detector: about 55%, not 49%

Field Note #23 showed that even the improved detector finds only part of the visible hands. On the same 60 reviewed seconds, 33 (55%) had a visible hand, a usable image and the camera on the work; the detector-based rule passed 22 (37%). The detector-based figure is therefore a lower bound; the reviewed estimate is about 55% (95% CI 42–67%), roughly 1.4 of 2.57 hours.

Finding 3

Usable time comes in short pieces

Usable seconds alternate with brief head turns and reaches out of view. The median usable stretch was 2 seconds, and stretches of 10 s or longer carried only 34–47% of usable time per session. A buyer who needs continuous 10-second episodes gets less than half of the usable total; one who trains on short clips or single frames gets all of it.

Four horizontal timelines, one per session, coloured second by second as usable, image OK without a detected hand, or blurred
Figure 2. Every second of the four sessions. Blue: usable for manipulation; grey: image OK, no hand found; red: camera blur above 8 px.
Histogram of per-second camera blur with a dashed line at 8 pixels
Figure 3. Per-second camera blur (90th percentile within the second) from the gyroscope.

How this is built

Raw stays raw; the index travels with it

The index is a CSV with one row per second per session: time, exposure, brightness, blur percentiles, head speed, pitch and hand count, plus the pass/fail columns. It is computed on the server after upload and stored next to the untouched recording. Cutting, privacy blurring and episode selection happen only at export, with the buyer's thresholds.

  • Change a threshold, re-export: no re-recording, no re-processing of video.
  • Every delivery states recorded hours, usable hours and the rules that produced them.
  • Hand presence is reported with its reviewed recall (Field Note #23).

What this means

Usable time, measured per second and adjustable

Verified: on this working day, 85% of recorded seconds had sharp, well-exposed images and about half to 55% were usable for hand-manipulation learning under the stated rules.

For buyers: ask any vendor for usable hours and the rules behind them. Ours come as a per-second table you can re-threshold yourself.

SCOPE AND LIMITS

Limitations

What this note does not show, stated up front.

  • One wearer, one task, one rig, one day.
  • Image checks use one keyframe per second; blur comes from camera rotation only (not object motion or defocus).
  • Hand presence relies on a detector that misses hands; the corrected estimate rests on a 60-second AI visual review, not human annotation.
  • Thresholds (8 px, −30°, exposure bounds) are our working defaults, not a standard.
  • Preliminary analysis: AI-assisted automated analysis; no third-party review.

FAQ

Questions data buyers ask

Q01How much of egocentric work video is usable for robot learning?
On 2.57 h of real head-mounted produce-portioning work, 85% of seconds had sharp, well-exposed images and 49% passed every check including a detected hand; correcting for detector misses, about 55%.
Q02How do you measure motion blur without looking at every frame?
From the gyroscope: rotation rate during each frame's exposure × exposure time × focal length gives camera blur in pixels, frame by frame.
Q03What removes the most usable time?
Hands out of view or not detected, then fast head turns (14.6% of seconds blurred above 8 px). Exposure and view direction removed almost nothing.
Q04Is the raw video cut?
No. The QC index is stored beside the untouched recording and applied only when exporting, using the buyer's thresholds.
Q05How long are the usable stretches?
Short: median 2 s. Stretches of 10 s or more held 34–47% of usable time per session.

OPEN DATA AND CODE

Reproduce and inspect the evidence

Every number in this note can be checked: derived tables, analysis code, figures and a runnable notebook are public.

GitHubAnalysis code and reproduction stepsorigin-data-lab/per-second-usable-time-qc-egocentric-workView repository Hugging FaceDataset card, tables and figuresorigin-data-lab/per-second-usable-time-qc-egocentric-workOpen dataset KaggleData explorer and runnable notebookorigindatalab/odl-egocentric-usable-time-qcExplore on Kaggle
Run the notebook on Kaggle

Licence: CC BY 4.0 (data and figures) · MIT (code)

Cite this note

Origin Data Lab (2026). Technical Field Note #25: How much of 2.6 hours of egocentric work video is usable? https://origindatalab.io/technical-field-notes/per-second-usable-time-qc-egocentric-work.html

CONTINUE READING

Related field notes

FIELD NOTE #23Hands are in view 75% of the time; an off-the-shelf detector finds a quarter of themFIELD NOTE #19Four days of real work: does a head-mounted stereo+IMU rig behave the same every day?FIELD NOTE #17Where a head-mounted camera points during 2 h 25 min of real work

Want this measured on your data?

Origin Data Lab plans and runs egocentric, stereo+IMU and field video capture with documented QA, for physical AI and robotics teams.

Request sample dataContact usData services