Internal · method record · Lear Body 2 pre-study · 26–27 Aug 2026

Body 2 Method Record

What was tried, what was measured, what was discarded and why — the audit trail behind the work-content findings.

Where it landed

yolo11n-pose on CPU · region = the hand-drawn box with its bottom extended to the frame edge · engaged = wrist depth below shoulder > 0.6 × torso · 15 s sampling for levels, 1 s for events. Engagement MAE 0.28, presence 0.39, within one operator on 97% of frames.

Decision log

chronological · every branch that changed the answer

Ran the supplied pipeline at its documented settings

SAM2 person count, --interval 5 --confidence 0.9, region boxes drawn by hand on 12 cameras.

✗ Frames scored 0 operators visibly contained 2–4. Confidence sweep on B2C12: mean 0.84 at conf 0.9, 1.49 at 0.7, 2.28 at 0.5 — a 2.7× swing on identical footage, and 43% of frames read zero at the shipped default.

Lowered confidence to 0.5

Recovered the missed operators — and started counting people across the aisle.

✗ Ground truth on 4 frames: at conf 0.5 a station with 3 real operators read 9. The two errors partially cancelled at conf 0.9, which is what made the original settings look plausible.

Trimmed 35% off the top of 6 aisle-contaminated cameras

Chosen by eye from the crops; validated on 2 cameras against manual counts.

✗ Applied to only half the cameras, which introduced a systematic bias: the six untrimmed read 3.13 mean against 1.91 for the six trimmed, on the same plan of 2.

Anchored every region to the yellow floor stripe

Auto-detected the stripe per camera, set the top edge 5% of frame height above it.

✗ 4–5× too tight on all twelve. Measured apparent person height per camera: 62–450 px, against a margin of only 24–97 px. B2C11 read 0.84 operators because its people were outside the box entirely.

Re-drew all twelve, then tuned the top edges by optimisation

Swept each top edge to maximise "capture" — the share of engaged detections falling inside the box — capped at ≤1 extra aisle person.

✗ Circular. Capture was scored against pose's own detections, which include people who are not that station's operators, so maximising it maximised admitting bodies. Six-way test: tuned tops scored worst of all variants (engaged MAE 0.61).

Reverted to the original hand-drawn boxes, bottom extended to the frame edge

Six-way comparison against 36 hand-scored frames settled it.

✓ Best of six: bodies 0.39, engaged 0.28, 97% within one. Extending down adds board and no people; raising the top adds people who should not count. An independent person-height calculation reproduced the original y1 to within 4 px on two cameras and 40 px on five.

Switched the model from SAM2 to pose

Triggered by the question "would hands be more robust than whole bodies?" — the answer turned out to be a model swap, not a rule change.

✓ A pose model must fit a whole skeleton, so it rejects the partial far-aisle bodies that inflated every detector. At B2C17: SAM2 11, yolo11n 7, pose 2, truth 2. This solved the region problem that steps 03–05 had been attacking from the wrong side.

Defined "working" as wrists below the board line

Board zone taken as everything below the yellow stripe.

✗ Returned 0 on frames with operators plainly working. Keypoints were found at 0.74–0.98 confidence but sat 56–187 px above the line. The stripe is painted on the floor; the board is an elevated bench, so hands appear above a floor marking in image space.

Redefined "working" as posture-relative reach

(wrist_y − shoulder_y) / torso_length > 0.6, computed per person.

✓ Self-normalising, so immune to camera height, resolution and the floor-versus-bench geometry that broke step 08. Needs no board polygon at all, which removes a calibration step. MAE 0.28 — better than any method's error on presence.

Model comparison

36 frames hand-scored across 9 stations · same regions, same frames

Model	Runs on	MAE	Bias	Exact	±1	Latency
SAM2 persons	remote GPU	2.03	+1.92	27%	52%	~12,400 ms	dropped
yolov8n persons	CPU	—	+0.20	—	—	68.6 ms	not pursued
yolo11n persons	CPU	1.28	+0.78	33%	75%	15.7 ms	dropped
yolo11n-pose persons	CPU	0.39	+0.11	69%	94%	53–246 ms	shipped
yolo11n-pose reach>0.6	CPU	0.28	+0.06	75%	97%	same call	shipped

The expensive model bought nothing. On an earlier 10-frame set SAM2 scored MAE 1.00 against yolo11n's 0.90 — statistically indistinguishable at ~800× the latency. SAM2's value in this study was as a zero-shot labelling tool, not a runtime. Pose costs 3–15× plain detection and is still comfortably CPU-only at 1 Hz.

Region variants

six definitions · same 36 frames · same model

Region definition	Bodies MAE	Bias	Engaged MAE	Bias	±1
Original hand-drawn, bottom extended	0.39	+0.11	0.28	+0.06	97%	shipped
Second redraw, as drawn	0.42	−0.14	0.42	−0.25	91%	clips heads
Original hand-drawn, as drawn	0.67	+0.28	0.44	+0.00	94%
Second redraw, bottom extended	0.64	−0.31	0.58	−0.25	88%
Tuned tops, bottom extended	0.61	+0.22	0.61	+0.17	86%
Tuned tops, as drawn	0.81	+0.47	0.61	+0.22	83%	worst

The two redraws sit either side of the original: the second redraw under-counts (bias −0.25, clipping heads to keep the aisle out), the tuned tops over-count (+0.22, admitting bench workers). Aisle discipline in the redraw was excellent on its own terms — 6 aisle admissions out of 230, 98% purity — but purity was bought with capture.

Sampling interval

not enforced by the service · chosen from the excursion-duration distribution

Excursions away from a station's own baseline, measured at 5 s and at 1 s. The distribution decides the interval, not the compute budget.

Excursion length	5 s sampling	1 s raw	1 s median-3	1 s median-5
1 sample	34%	4,368	933	553
2 samples	19%	2,134	796	408
20–39 s	—	314	436	434
40–79 s	—	65	163	217
80 s+	—	10	30	49

5 s is a dead zone. A third of excursions there last exactly one sample — indistinguishable from detector noise at MAE ~1.0 — while the median-of-5 gate needed to suppress that noise spans 25 s and erases 80% of all excursions. Raw is untrustworthy, smoothed is blind.

At 1 s both work, and the smoothing behaves diagnostically: the 1–2 s population collapses 6,502 → 961 while the ≥20 s population grows 389 → 700. That growth is real events being reassembled from fragments a noisy detector had split. Chosen: 15 s for levels — unbiased at any interval, precision scales with √n — and 1 s for event structure.

Analyses withdrawn

reached a conclusion, then failed a check

Claim	What killed it
Extra operators stand significantly upstream	The null pooled all normal-frame positions into one distribution. Stations have multiple standard work positions (B2C17 shows 3 modes, IQR 0.560) — a second person simply occupies the other one. Also an order-statistic artifact inflated the first estimate ~3×.
Entry edge identifies work-ahead behaviour	Attributes to the wrong station. A person entering station N from downstream most likely belongs to N+1. Replaced by exit direction for people who first appear mid-frame — those were demonstrably working there.
Staffing is phase-locked to the cycle (R² 45%)	Held at Station 1, does not generalise. Tested at each station's own period on 1 s data: median R² 4.2%, only 2 of 23 station-hours above 15%. The period generalises; the rhythm does not.
Chase rate as "% of boards chased"	Exceeded 100% everywhere. ByteTrack fragments an occluded person into several tracks, inflating absolute counts by an unknown factor. Replaced by share of the station's own appearances, where fragmentation largely cancels — which reorders the ranking (rank correlation 0.62).
B2C17/B2C13 idle time means operators waiting	They also have the highest chase rates. Much of that "idle" is likely operators walking downstream after boards, not standing. Unresolved — needs a walking-versus-standing classifier, and it inverts the recommended intervention.

What survived checking

each confirmed by an independent method

Finding	Value	Independent confirmation
Pitch	175 / 176 / 179 s	Board QR tracking · occupancy autocorrelation at 15 s · same at 1 s. Three methods within 4 s.
Cameras do not overlap	0 collisions	ORB matching (4–9 RANSAC inliers of 6,000 keypoints) · visual inspection · no board ID in two cameras at the same instant.
Exit direction is behaviour	r = −0.01	Against mean working position. Net drift agrees: all six chasing stations drift downstream, B2C12 drifts upstream.
No cascade between stations	+0.14 / +0.02	Lag 0 and lag +1 across 9 adjacent pairs. Spatial pattern is build-then-absorb, never accumulate.
Work content ≠ chase index	r = +0.30	Weak by design — they measure work done versus work that did not fit. Both are needed to classify a station.
Line runs faster than takt	179 s vs 211 s	Takt supplied by Lear; pitch measured. Yields ~82 min/shift of downtime cushion.

Compute

measured on an M-series laptop, 10 cores

The remote endpoint was the only real constraint. Saturated at ~1.2 requests/sec regardless of client parallelism — 4 jobs gave 1.15/s, 8 jobs 1.20/s. Later in the day its median latency was 6.1 s per call. Local CPU sat at 3% while eight workers waited on it.
Inference is not the bottleneck; decode is. Streams are HEVC 2592×1944 at ~12 fps, 3.35 Mbit/s. One stream decodes at ~80 fps on 1.5 cores; 22 live streams is ~5 CPU cores, or near-free on a GPU's decode engine. Aggregate network ~45 Mbit/s.
Live inference fits on CPU. 22 stations at 1 Hz with plain detection is 0.35 s of CPU per wall-second. Pose is 3–15× that and still fits. The GPU case rests on decode offload and offline fine-tuning, not on keeping the line running.
Raising image size bought nothing — MAE 0.90 at both 640 and 960 for yolo11n, at 2.3× the time. A simulated camera substream (640×480 full-frame downscale) retained 7 of 8 detections.

Datasets produced

all retained · ~/Downloads/Lear

Directory	Rows	Configuration
results_pose_v2	5,381	pose, present + engaged, 15 s, final regions	authoritative
results_tracks	14,023	ByteTrack over yolo11n at 12 fps, entry/exit edges	authoritative
results_local_1s_orig	80,497	yolo11n persons at 1 s — event structure, period detection	supporting
results	5,381	SAM2 at 15 s, final regions — method comparison	supporting
results_tracks_t	in progress	as above plus per-track timestamps, for the sub-block cascade test	running
results_stripe_anchored	16,101	SAM2 5 s, stripe regions — superseded, kept for audit	superseded
results_trimmed35	16,101	SAM2 5 s, 35% trim on 6 cameras — superseded	superseded
results_preview	1,362	SAM2 60 s scoping pass at conf 0.9	superseded

Eight region definitions are on disk as boxes_*.json, every one reproducible via draw_all_boxes.py --configs-only. All batch runners are resumable — row-level for counting, chunk-level for tracking — and resume was verified by SIGKILL rather than assumed.

Open

known limits of this record

Ground truth is one scorer. 36 frames, 9 stations, scored by eye. Enough to separate a 5× accuracy gap; not enough to certify an MAE. A second scorer with disagreements adjudicated would firm it.
Per-camera accuracy varies and a single figure hides it. Pose is perfect on B2C5, B2C7 and B2C11 (MAE 0.00) and worst on B2C17 (+1.25, still admitting bench workers).
Two cameras are framing-limited, not tuning-limited. B2C11 views a walkway; B2C12 is unconfirmed and carries the lowest measured work content on the line.
The cascade test cannot resolve a fast whiplash. 5-minute blocks against 120 s transit. The timestamped tracker re-run now in progress removes that limit.
Everything rests on two hours of one evening, one product mix, with the line running ~40% leaner in the second hour for reasons nobody has explained.