Technical Field Note #17 · November 2026

Where does a head-mounted camera point during real work? View direction, hands and self-occlusion over 2 h 25 min

Field Note #17 · Camera view direction, hand position and self-occlusion, head-mounted stereo+IMU rig (ODL-EGO S1)

Head-mounted egocentric data is usually described by hours and resolution, not by where the camera was actually pointing. We measured it for every second of 2 h 25 min of real work, using the gravity direction from the rig's 500 Hz accelerometer. The optical axis was a median 56° below horizontal; 71% of the time it was 45–75° down and only 8% of the time within 30° of level. The IMU-predicted horizon lined up with vanishing points in the images, and the same desk posture at session starts and ends gave 43–48° across four sessions, so the rig did not slide on the head. In sampled frames the hands were visible 83 of 96 times, typically 180 px (21°) below the image centre, well inside the central field, while the wearer's own jacket covered a median 10% of the frame and more as the head tipped down. From the next delivery on, every delivery includes a per-frame view-direction file.

Published November 2026⏱ 14 min readStatus Preliminary · automated analysisOpen data GitHub · Hugging Face · Kaggle
56°Optical axis below horizontal, median8,738 s; 71% of time at 45–75°Verified (automated)
83 of 96Frames with a hand visible95% CI 78–92%; visual labelsPreliminary (AI-labelled)
65% from topHand height in image, median≈21° below the optical axisPreliminary (AI-labelled)
10%Frame covered by the wearer's own body, median256 sampled frames; 0% under 30° to 14% over 75°Verified (automated)

Why this matters

Hours of video say nothing about what is in view

For robot learning from egocentric video, what matters is whether the hands and the objects they touch are in view, how much of the frame is the wearer's own body, and whether the image is upright. These depend on where the camera points, which in turn depends on the task and on how the rig is mounted.

The rig already records the answer: its accelerometer measures gravity in the camera's frame. This note turns that into a per-second view direction for a full morning of real work, checks it against the images, and looks at what it means for hands and self-occlusion.

Setup

8,738 seconds of real work, 256 sampled frames

Recording: 2026-11-01, 09:48–12:37, the same six sessions as Field Notes #15 and #16: produce portioning for school meals in a food supplier's workroom, rig worn on the head, one worker. The worker mostly stood at a bench, with about half an hour of carrying and staging produce at the start.

View direction: gravity from the 500 Hz accelerometer averaged over 1 s windows and rotated into the left camera frame with the factory camera–IMU calibration (a rotation of 0.25°). Pitch is the angle of the optical axis below horizontal; roll is the rotation of the image about that axis.

Images: 256 left frames (8 per 5-minute file). The wearer's blue jacket was segmented by colour (connected regions touching the bottom edge) to measure self-occlusion. In 96 of them (16 per session, evenly spaced) the vertical position of the hands was labelled visually.

Check against the images

The IMU horizon matches the scene, and the rig did not slide

In frames that show a corridor or room with a clear vanishing point, the horizon predicted from the accelerometer alone falls on the vanishing point to within a few pixels, about 1° (three frames checked by eye).

Sessions started and ended at the same desk with the phone in hand (starting or stopping the recording). The 6 sampled frames from these moments, in 4 sessions between 10:42 and 12:37, gave 43.2–47.7° with no trend, so the rig sat at the same angle on the head throughout.

Three frames of corridors and a doorway with a red line marking the horizon predicted from the accelerometer, lying on the vanishing point of each scene
Figure 1. Red: horizon predicted from the accelerometer only. Green dashed: optical-centre row. No other people are visible in these frames.

Result

The camera looked 56° down for most of the morning

Across all 8,738 seconds the optical axis was a median 56° below horizontal (middle half 47–63°). It was 45–75° down 71% of the time and within 30° of level only 8% of the time.

The first session differs because it was mostly carrying and staging produce (median 37°, 38% of the time under 30°). From 10:20, at the bench, the median was 58°, rising slightly through the morning (about 3° per hour); fatigue and task mix cannot be separated here.

Head turning about the vertical had a median of 19°/s and a 90th percentile of 65°/s.

Histogram of camera pitch peaking at 55 to 60 degrees below horizontal, and a time series rising from about 20 degrees at 9:50 to about 60 degrees from 10:30 onward
Figure 2. Left: optical-axis angle below horizontal for every second. Right: median and middle half per 5 minutes over the morning.
Table 1. The camera looked 56° down for most of the morning
Session startMinutesMedian below horizontalMiddle 50%Time < 30°Time 45–75°Rows tilted > 10°
09:4827.637°24–47°38%28%40%
10:2017.954°47–61°1%76%38%
10:4221.757°51–64°1%76%47%
11:1114.156°49–62°1%77%57%
11:2832.360°54–66°0%85%65%
12:0531.959°53–67°0%87%59%
All145.656°47–63°8%71%52%

Hands and own body

Hands near the centre; 10% of the frame is the wearer

A hand was visible in 83 of 96 labelled frames (78–92%, 95% CI). The hands sat a median 65% of the way down the image (middle half 62–73%), about 180 px or 21° below the optical centre. That is inside the central field used for the calibration check in Field Note #16 (within 500 px of the centre), where the fisheye calibration is tightest.

Because the head tips down, the wearer's own jacket, torso and sleeves, is often in view: a median 10% of the frame overall, rising from 0% under 30° to 10% at 45–60°, 11% at 60–75° and 14% beyond 75°.

The image was rarely upright: with the camera less than 70° down, the rows were tilted by more than 10° in 52% of seconds and by more than 20° in 20% (median 10°, mostly to one side).

Three workroom frames at 16, 60 and 72 degrees below horizontal with the wearer's jacket tinted orange, covering more of the lower image as the camera tips down
Figure 3. Examples at three angles; the wearer's own jacket is tinted. Dashed line: image centre row.
Scatter of own-body share rising with camera pitch, and a histogram of hand height peaking at 60 to 65 percent of image height
Figure 4. Left: share of the frame covered by the wearer's jacket against camera pitch (256 frames). Right: vertical hand position in 83 frames with a visible hand.
Table 2. Hands near the centre; 10% of the frame is the wearer
Below horizontalShare of timeSampled framesOwn body in frame (median)Hand-labelled framesHand visible
under 30°8%160%367%
30-45°14%266%1486%
45-60°45%11410%4486%
60-75°27%7911%2886%
over 75°7%2114%7100%

What we changed

A view-direction file with every delivery

Verified: the accelerometer gives a usable per-frame view direction for this rig, checked against vanishing points and a repeated posture.

Process: the delivery QA tool now writes per_frame_view_<session>.csv with pitch, roll and head-turn rate for every frame, and adds the pitch distribution to the QA report (tested on this recording). From the next delivery on, every delivery includes it, so buyers can separate bench work from walking and looking around before they open a single video.

Mounting: for standing bench work we keep the current mount, which points the camera 10–15° below horizontal when the wearer looks ahead down a corridor. The wearer's own head tilt already brings the hands to about 21° below centre; in this recording, frames taken with the camera pointing further down show more of the wearer's jacket, so a downward-tilted mount would likely add self-occlusion for little gain. A seated or floor-level task may need a different answer.

  • Per-frame pitch, roll and head-turn rate in every delivery.
  • Expect about a tenth of each bench-work frame to be the wearer's own body; mask it if your model needs to.
  • Keep the current mount for standing bench work; re-check for other postures.

SCOPE AND LIMITS

Limitations

What this note does not show, stated up front.

  • One rig, one worker, one task and one morning (2026-11-01); posture depends on bench height, the worker's height and the task.
  • Pitch and roll come from 1 s accelerometer averages; during strong accelerations (carrying, fast turns) they are less accurate. Roll is not reported for pitch above 70°, where it is ill-defined.
  • Hand positions were labelled visually on 96 frames by an AI model and not re-checked by a person; hand height is vertical position only, converted to degrees with an equidistant fisheye approximation. AI-only preliminary analysis.
  • Self-occlusion was measured by the colour of this wearer's blue jacket; another person in the same jacket touching the bottom edge could be counted, and a different jacket would need a different mask.
  • Walking versus standing was not separated: the accelerometer step-band signal was not validated against the images, so no walking share is reported.
  • Example frames show the workroom with the site owner's consent; no faces or feet are visible in them. Raw video is not published.

FAQ

Questions data buyers ask

Q01Where does a head-mounted camera point during manual work?
In this recording of standing produce portioning, a median 56° below horizontal; 71% of the time between 45° and 75° down and only 8% within 30° of level.
Q02How often are the hands visible in head-mounted egocentric video?
In 83 of 96 evenly sampled frames (78–92%, 95% CI), typically about 21° below the image centre.
Q03How much of an egocentric frame is the wearer's own body?
A median 10% in this recording, from about 0% when looking ahead to 14% when looking steeply down.
Q04Should a head-mounted camera be tilted down for tabletop work?
Not for standing bench work in this test: with a mount that looks 10–15° down when the wearer faces ahead, the hands already sat near the centre because the wearer tilts the head, and steeper views showed more of the wearer's own body.
Q05How is camera pitch measured without a vision model?
From the gravity direction measured by the rig's accelerometer, averaged over 1 s and rotated into the camera frame with the factory camera–IMU calibration.

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/head-mounted-camera-view-direction-real-workView repository Hugging FaceDataset card, tables and figuresorigin-data-lab/head-mounted-camera-view-direction-real-workOpen dataset KaggleData explorer and runnable notebookorigindatalab/odl-head-mounted-camera-view-directionExplore 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 #17: Where a head-mounted camera points during 2 h 25 min of real work. https://origindatalab.io/technical-field-notes/head-mounted-camera-view-direction-real-work.html

CONTINUE READING

Related field notes

FIELD NOTE #16Recording integrity and calibration hold-up over 2 h 25 min of real workFIELD NOTE #15Head motion and motion-blur budget in 2 h 25 min of real egocentric workFIELD NOTE #14IMU noise and bias of a stereo+IMU egocentric rig, measured from static recordings

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