Technical Field Note #19 · November 2026

Four days of real work, 8.9 hours, 957,336 stereo frames: does a head-mounted stereo+IMU rig behave the same every day?

Field Note #19 · Day-to-day consistency of a head-mounted stereo+IMU rig over four days of real work (ODL-EGO S1)

Field Note #16 showed that one morning of real work came through intact. A buyer of hundreds of hours needs more: the same behaviour every day, after the rig has been charged, carried, put on and taken off again. We recorded the same produce-portioning job on four consecutive mornings (2026-11-01 to 11-04) and ran the same checks on all of it: 17 work sessions, 8.9 hours, 957,336 frames per eye. Every day had zero dropped frames, a 15 µs median left/right exposure difference (worst case 61–70 µs per day), an IMU rate of 501.4–501.7 Hz with no sequence breaks and no gap above 2.2 ms, seamless 5-minute file boundaries, and a 10.04 ms median exposure. Head-motion statistics from the IMU were also nearly identical across days. The only irregularities were operational: three accidental recordings shorter than 2 s and one session whose device log lacks a stop entry although its data are complete.

Published November 2026⏱ 12 min readStatus Preliminary · automated analysisOpen data GitHub · Hugging Face · Kaggle
0 of 957,336Dropped framesper eye, 17 sessions on 4 daysVerified (automated)
15 µs median, every dayLeft/right exposure differencedaily maximum 61–70 µsVerified (automated)
501.4–501.7 Hz · 2.2 msIMU rate and largest gapno gaps over 5 ms, no sequence breaksVerified (automated)
27.5–29.7 °/sHead rotation, daily medianper-second mean gyro speed; p95 91–95 °/sVerified (automated)

Why this matters

One good morning is not a production rig

Most capture-rig validations stop at one recording. Production data are collected day after day: batteries are recharged, the rig is carried, worn, taken off and powered up again, and the storage card fills and empties. Each of these can change behaviour.

This note repeats the integrity checks of Field Note #16 on four consecutive days of the same real job and asks one question: is day four indistinguishable from day one?

Setup

Same job, same operator, four mornings

Produce portioning for school meals in a food supplier's workroom: weighing, trimming, sorting and bagging vegetables. The rig was worn on the head; recording was started and stopped by the operator. Site, people and labels in the images below are blurred or cropped.

Six head-mounted camera frames of hands weighing cabbage, sorting peppers, trimming green onions, portioning potatoes and packing carrots
Figure 1. Left-eye frames from each day (1600 × 1200, fisheye). Text on boxes and documents is blurred.
Three bar charts of recorded hours, frames per eye and sessions for each of the four days
Figure 2. Volume per day: 1.9–2.4 hours, 3–6 sessions, 209,455–261,390 frames per eye.
Table 1. Same job, same operator, four mornings
DaySessionsHoursFrames per eyeDroppedL/R median (µs)L/R max (µs)Gyro rate (Hz)IMU max gap (ms)Exposure median (ms)Head rotation median (°/s)p95 (°/s)
2026-11-0162.42261,39001570501.571–501.6442.1910.0427.595.1
2026-11-0241.94209,45501564501.582–501.6232.1110.0428.191.4
2026-11-0332.33250,93501565501.357–501.6372.210.0428.592.2
2026-11-0442.18235,51501561501.598–501.6882.1110.0429.794.3

Timing and IMU

Every day looked like the first day

Frames: no drops on any day by interval counting; for every session with a stop entry, the device log's frame counts matched the frames in the files (to within the one final frame) and reported zero strobe drops. All 5-minute file boundaries followed after exactly one frame period (33.3 ms).

Stereo sync (pairs by frame order, see Field Note #18): the exposure-start difference had a 15 µs median every day and a daily maximum between 61 and 70 µs. Exposure: 10.04 ms median every day, 95th percentile 10.54–10.56 ms.

IMU: 501.4–501.7 Hz in every session (the device's nominal rate is 500 Hz), largest inter-sample gap 2.2 ms, zero sequence breaks, and gyro and accelerometer sample counts identical in every file.

Three box plots by day for left-right sync, exposure time and head rotation, nearly identical across days
Figure 3. Distributions by day. Boxes span the 25th–75th percentile.
Scatter of IMU sample rate per session between 501.3 and 501.7 Hz
Figure 4. IMU sample rate per work session.
Table 2. Every day looked like the first day
DayStartMinutesFilesFrames per eyeDroppedL/R max (µs)Gyro (Hz)IMU max gap (ms)File boundary gap (ms)Device log stop
11-0109:4827.5649,473057501.6442.1133.3-33.3user_stop
11-0110:2017.9432,225052501.5822.1133.3-33.3user_stop
11-0110:4221.7538,998055501.5822.1133.3-33.3user_stop
11-0111:1114.1325,334070501.5792.1933.3-33.3user_stop
11-0111:2832.2758,015067501.5722.1133.3-33.3user_stop
11-0112:0531.9757,345056501.5712.1133.3-33.3missing
11-0209:0627.8650,105064501.6232.1133.3-33.3user_stop
11-0209:3622.2539,957051501.5842.1133.3-33.3user_stop
11-0210:0417.6431,605059501.6092.1133.3-33.3user_stop
11-0210:4248.81087,788057501.5822.1133.3-33.3user_stop
11-0310:3735.8864,380059501.6372.1133.3-33.3user_stop
11-0311:1630.1754,130059501.3572.233.3-33.3user_stop
11-0311:5273.615132,425065501.5722.1133.3-33.3user_stop
11-0409:3840.0972,024057501.6262.1133.3-33.3user_stop
11-0410:2223.0541,431057501.6882.1133.3-33.3user_stop
11-0410:4940.4972,802061501.5982.1133.3-33.3user_stop
11-0411:5527.4649,258056501.6182.1133.3-33.3user_stop

Work rhythm

The same job leaves the same motion signature

From the gyroscope we computed the mean head-rotation speed for every second of recording. The daily median was 27.5–29.7 °/s and the 95th percentile 91–95 °/s; 5.6–7.4% of seconds were nearly still (< 5 °/s). The rhythm of reaching, weighing and bagging is visible as a steady band, interrupted by breaks between sessions.

Note on method: these are per-second means, so they are higher than the per-frame median of 22 °/s reported in Field Note #15 for day 1, which weights still frames more heavily. The two notes agree when the same method is used.

Four stacked time series of head rotation speed by clock time for each day
Figure 5. Head-rotation speed (10 s smoothing) by clock time, one panel per day.

Irregularities

What did go wrong, and why it did not cost data

  • Three accidental recordings of 0.2–1.2 s (one on day 2, two on day 4), most likely from a double press of the record button. They contain 0–24 frames and are dropped automatically by a 60 s minimum-length rule.
  • One session (day 1, 12:05) has no stop entry in the device log, consistent with power-off before stop. Its 57,345 frames per eye and its IMU stream are complete; only the log line is missing.
  • In 8 of 118 files the left and right frame counts differ by one; all are the final file of a session (see Field Note #18).

Calibration was not re-checked in this note. The calibration file saved inside each recording folder belongs to a different resolution (Field Note #20); the correct-mode check over a full shift is reported in Field Note #16.

What we changed

Daily QA, not one-off validation

Verified: across four days of real work the rig's frame, sync, IMU and exposure behaviour did not change measurably.

Process: recordings are now uploaded automatically every night with SHA-256 verification, and the same analysis runs on a short-lived cloud server, so these numbers are produced for every capture day rather than once.

SCOPE AND LIMITS

Limitations

What this note does not show, stated up front.

  • One rig, one operator, one workroom, four mornings; not a population estimate.
  • Head-rotation statistics are per-second means of gyro speed and are not directly comparable to per-frame statistics.
  • Calibration and metric scale were not re-checked in this note.
  • All checks are automated: AI/algorithm-only preliminary analysis; no human re-verification.

FAQ

Questions data buyers ask

Q01Does the stereo+IMU rig behave the same from day to day?
In four consecutive days of real work, yes: 0 dropped frames, 15 µs median left/right exposure difference each day, IMU 501.4–501.7 Hz with no gap above 2.2 ms, and a 10.04 ms median exposure every day.
Q02How much data was recorded?
8.9 hours in 17 work sessions, 957,336 frames per eye at 1600 × 1200, 30 fps, with a 500 Hz IMU.
Q03Were there any problems?
Three accidental recordings under 2 s and one session without a device-log stop entry. None caused data loss.
Q04Does head motion differ between days?
Very little: the daily median head-rotation speed was 27.5–29.7 °/s and the 95th percentile 91–95 °/s.
Q05Was calibration checked over the four days?
Not in this note. Field Note #16 checks the correct-mode calibration over a full shift; Field Note #20 explains why the file saved in each recording folder must not be used.

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/four-day-real-work-stereo-imu-consistencyView repository Hugging FaceDataset card, tables and figuresorigin-data-lab/four-day-real-work-stereo-imu-consistencyOpen dataset KaggleData explorer and runnable notebookorigindatalab/odl-four-day-stereo-imu-consistencyExplore 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 #19: Four days of real work with a head-mounted stereo+IMU rig. https://origindatalab.io/technical-field-notes/four-day-real-work-stereo-imu-consistency.html

CONTINUE READING

Related field notes

FIELD NOTE #17Where a head-mounted camera points during real workFIELD NOTE #16Recording integrity and calibration over 2 h 25 min of head-mounted workFIELD NOTE #15Head motion and motion-blur budget in 2 h 25 min of real egocentric 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