Files
aimsight-labeling/auto/analyses/MICHIGAN_2026-10-03T0000_2026-10-12T1600/analysis.md
T

81 lines
15 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Labels: MICHIGAN, 2026-10-03T00:00Z → 2026-10-12T16:00Z (UTC)
Request: verify the three existing labels of the range (`work/existing.jsonl`: MICHIGAN-1, MICHIGAN-2, MICHIGAN-4) and the spanning label MICHIGAN-3; keep, correct, remove, add. DETECTORS: no.
Station: indoors, in the same building as BREITHORN and NEBRASKA (`stations.md`); one measurement every 4 h (59 instants in the range, none missing, a pose at every instant); no image-quality metrics (export columns peak to snr empty); **no weather** (`data/MICHIGAN/weather/` does not exist: no MeteoSwiss data for this station). Images: one per measurement, uniformly dark indoor scene, so no moment of the day can be judged from them.
## Verdict per existing label
| Old id | Verdict | Evidence |
|---|---|---|
| MICHIGAN-1 | **kept** | `summary.py corr_x/corr_y/x/y/error_value`: COLH21' and REFH22' carry identical values at every instant until 05/10 12:00, separate at 16:00 (raw y of COLH21' up by several px, of REFH22' down by more), their corr_y jumps from far below their neighbours to their level (15-16 noises, `ranges.py`), their error_value drops. Thumbnails `--fixed` 05/10 12:00 and 16:00 (`pair.png`): at 12:00 both crosses sit between two spots, at 16:00 COLH21' is on the top spot and REFH22' on the bottom one. At 16:00 the raw positions of all sensors shift (COLH18' by several px, the others and the references by about a pixel or less) with no change of pose and no jump of the other corr: a reconfiguration. Bounds right (last wrong point 12:00, end at the first right one). Minor nuance: "change a little" understates the step of COLH18', as large as that of the pair; not worth losing the validation. |
| MICHIGAN-2 | **corrected** (end 08:00 → 20:00, description) | The 04:00 dip is confirmed (6 targets beyond 2.5 noises in corr_y, COLH21' and COLH20' up in corr_x; median of the targets' corr_y −5.5 noises; housing temperature at its minimum of the range; tilt +4.5 and scale −7 noises). But the edge checked sensor by sensor: COLH18S''' corr_y stays low at 08:00, 12:00 and 16:00 (−3.0 and −3.8 noises at 12:00 and 16:00, `ranges.py --k 3`), POUH18' and COLH18' a little lower at 12:00-16:00, all back at 20:00; the pose comes back gradually (tilt 0.0161 → 0.0147 → 0.0140 → 0.0132°, scale back at 16:00), the temperature back at 16:00, and the error_value of COLH18', POUH18', COLH19N' raised from 04:00 to 12:00. "At the end of the night" removed: the indoor images do not show the moment of the day. |
| MICHIGAN-4 | **corrected** (start 09/10 12:00 → 08/10 00:00, description) | The camera drift already passes into the curves with three references: from 08/10 04:00, a common step of corr_x (median over the targets +1.1 to +1.9 noises at every measurement until 09/10 08:00, against −0.9 to +0.6 before; 4 to 7 targets beyond +2 noises at each instant against 0 to 2 before; `scripts/common.py`), COLH18', POUH18', POUH19N', COLH21', COLH19N', COLH18S''' beyond +3 noises repeatedly, up to +5.5 (`ranges.py --k 3`), POUH19N' corr_y −3.3 to −4.3 noises; the corr_x of REF42 and REFG43N move apart (+0.025 / −0.036 at 08/10 04:00 against about ±0.01 before). It falls on the clearer step of the camera between 07/10 20:00 and 08/10 08:00 (tilt 0.0152 → 0.0175°, scale down), the step NEBRASKA-1 dates on 7-8 October. Rigid (`overview.py --rigid 2026-10-05T20:00Z`): ty of the targets rises steadily from 0.1 to 0.4 px between 07/10 and 12/10, scale and rotation of the targets down, median target residual 0.01-0.03 px: the targets move as a block, the camera moves. One event across the base change (`labeling-rules.md` §3: a change of speed does not cut it). "Its scale drops": true of the targets in the image and of the pose until 9 October; the pose computed on two references after 9 October shows the scale rising, which the description now says. The expert sentence is kept word for word. |
| MICHIGAN-3 (spanning) | holds within the range | REFPOU22 rejected at all 20 measurements from 09/10 12:00 (error_value up to 0.93); thumbnails `--fixed` 09/10 08:00, 12:00, 12/10 16:00: its spot sharp at 08:00, gone at 12:00 and on the 12th, REF42 and REFG43N intact; corr of REF42/REFG43N at 0 (to 1e-13: two references); jump of corr_x and corr_y of all targets at 12:00 with no step of the raw positions (rigid ty 0.174 → 0.175 px). Nothing to change. |
Added: one label, 04/10 20:00 → 05/10 12:00, Thermal displacement (below). Removed: none.
## 1. My labels
### 2026-10-03T00:00Z → 2026-10-05T16:00Z (MICHIGAN-1, kept)
- **Phenomenon of the catalogue**: The configuration of a sensor has changed
- **Precedents**: none needed: the label is reviewed and its facts check out
- **Family**: Target issue: badly configured pair, then reconfigured (boundary "Target issue or Displacement")
- **Measured**: COLH21' and REFH22' identical (x 1534.7, y 1159.1) until 05/10 12:00, then x 1535.9 / 1532.4, y 1152.3 / 1170.8; corr_y −0.14 → +0.02; error_value 0.09-0.11 → 0.04
- **Images**: thumbnails 05/10 12:00 and 16:00 (`pair.png`), as the label says
- **Confidence**: high for the facts; the cause stays deduced, as written
### 2026-10-04T20:00Z → 2026-10-05T12:00Z (added)
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle (sheltered)
- **Precedents**: MICHIGAN-2 (reviewed, same station, the next night: same signature; the sign of the corr_y is opposite because the pose moves the other way at that instant); MICHIGAN-9, MICHIGAN-10 (reviewed, same station later in October: brief housing drop, camera nods, corr of all targets move together, Thermal displacement); BREITHORN-29 (not reviewed, the same morning at the neighbouring camera, thermal). They gave the family and the way of writing it.
- **Family**: Thermal displacement: follows a drop of the housing temperature and comes back with it, nothing on the images, error_value at its level; not Displacement (it returns at equal temperature), not Visibility (no error rise, no nuisance)
- **Measured**: housing temperature 34.0 °C at 04/10 20:00 → 32.5 at 00:00, 32.5 at 04:00, 32.4 at 08:00 → 33.75 at 12:00 (−5.5 to −6 noises of the 4 h series). corr_y at 00:00: all 14 targets up by 1.8 to 3.7 noises (level 03/10 12:00 – 04/10 20:00); `ranges.py --k 3`: COLH19N' +5.9, POUH19N' +5.0, COLH21'' +4.4, COLH18S''' +4.1, COLH19N'' +3.4, POUH18' +3.4, COLH21 +3.3, POUH20' +3.2; median of the targets +5.7 noises of its series. 04:00: corr_y mostly back, corr_x up 2 to 2.6 noises on COLH20', COLH20'', COLH21'/REFH22', POUH21'', COLH19N'. 08:00: corr_y up again on COLH19N' (+4.8), POUH19N' (+4.1), COLH19N'', POUH20', COLH18', COLH18S''' (+2.2 to +2.7). 12:00: all within 1.6 noises. Pose: 00:00 rotation −0.0105 → −0.0088° (+3.9 noises), tilt 0.0128 → 0.0112°; 04:00 and 08:00 tilt 0.0135, 0.0139°, scale 0.999943, 0.999930 (−3.3, −4.3 noises); 12:00 scale 0.99997. Rigid from 04/10 20:00: targets ty +0.05, +0.10, +0.14 px, scale down, median residual 0.01-0.02 px; REFPOU22 dy −0.037 px against the targets at 00:00, −0.024 at 08:00. error_value of the targets flat.
- **Images**: whole images 05/10 00:00 and 08:00: dark indoor scene, all spots present, nothing in front
- **Confidence**: medium: a small deviation (a few noises) at a 4 h cadence, but common to all the targets, with the signature of the reviewed MICHIGAN-2 of the next night
### 2026-10-06T00:00Z → 2026-10-06T20:00Z (MICHIGAN-2, corrected)
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle
- **Precedents**: MICHIGAN-2 itself (reviewed); MICHIGAN-9, -10 for the shape (dip, slow return)
- **Family**: Thermal displacement (unchanged)
- **Measured**: housing temperature 33.75 (05/10 20:00) → 32.9 (00:00) → 31.75 (04:00, minimum of the range) → 32.5 → 33.0 → 34.0 at 16:00. At 04:00 corr_y −2.5 to −4.3 noises on COLH20', COLH21', REFH22', COLH21'', POUH21'', COLH18S'''; corr_x COLH21' +3.0, COLH20' +2.5. COLH18S''' corr_y 0.168 (05/10 20:00) → 0.106 (04:00), 0.127, 0.117, 0.106 (16:00), 0.167 (20:00); POUH18' 0.126 → 0.073, 0.127, 0.097, 0.098, 0.129. error_value COLH18' 0.048 → 0.060 (04:00) → 0.047 (16:00), POUH18' 0.057 → 0.070 → 0.057, COLH19N' 0.047 → 0.060 → 0.048. Pose tilt 0.0116 (00:00), 0.0161, 0.0147, 0.0140, 0.0132°; scale 0.99988 at 04:00, back at 16:00. Rigid: ty +0.116 px at 04:00, the references sliding a little more than the targets (REF42 dy +0.064, REFPOU22 +0.044, REFG43N +0.031 px).
- **Images**: whole image 06/10 04:00: dark, nothing in front of the spots
- **Confidence**: medium for the end at 20:00 (one target clearly, two weakly), high for the event
### 2026-10-08T00:00Z → 2026-10-12T16:00Z (MICHIGAN-4, corrected)
- **Phenomenon of the catalogue**: The camera tilts little by little; then A compensation on two references
- **Precedents**: NEBRASKA-1 and BREITHORN-2 (reviewed, the neighbouring cameras, Displacement, same drift; NEBRASKA-1 dates the clearer step of 7-8 October and says its three references absorbed it, which is not fully the case here); MICHIGAN-4 itself for the expert sentence
- **Family**: Displacement: a drift of several days that does not come back at equal temperature (expert sentence), the targets do not move relative to each other; not Thermal displacement
- **Measured**: medians of corr_x 05/10 16:00 – 07/10 16:00 → 08/10 04:00 – 09/10 08:00 (`scripts/steps.py`): COLH18' 0.101 → 0.137 (+3.1 noises), POUH19N' 0.101 → 0.141 (+2.9), COLH21' 0.044 → 0.074 (+2.6), POUH18' 0.172 → 0.219 (+2.5), COLH20' +2.0, COLH19N' +1.7, COLH18S''' +1.5; COLH20'', COLH21, COLH21'' unmoved. Housing temperature 32.9 → 32.1-32.4 °C over 08/10-09/10, flat. After 09/10 12:00: corr_x of all targets rises again by roughly 0.07-0.10 between 09/10 12:00 and 12/10 (COLH18' 0.14 → 0.24, COLH21 0.12 → 0.19), with alternating swings from one measurement to the next that follow the gap between REF42 and REFG43N (`scripts/refgap.py`: gap_x 321.47 → 321.52 px between 11/10 08:00 and 12:00, median corr_x of the targets 0.10 → 0.29). Pose on two references: pan −0.0105 → −0.0174°, tilt noisy upward, scale 1.00003 → 1.00009; rigid on the targets: scale 0.99995 → 0.99988, ty 0.17 → 0.41 px.
- **Images**: none useful (no visible cause; REFPOU22 thumbnails under MICHIGAN-3)
- **Confidence**: high for the drift after 9 October; medium for the start (a shift of a few noises, but common, lasting and aligned with the camera step)
## 2. What I left without a label
- **03/10 08:00**: corr_x of COLH18', POUH18', POUH20' about −2.5 noises (5 targets beyond −3 noises on the whole-batch noise of `ranges.py --k 3`, at most −2.7 on the window noise), back at 12:00; the corr_y of the three references alternate around this instant (±4 to 6 noises at 00:00, 04:00, 08:00: jitter of the three-reference fit, REF42 off the targets by a few hundredths of a pixel in the rigid fit); housing temperature dip under half a degree, pose within 3 noises. One measurement, below the deviations of 05/10 and 06/10 and below the 5-noise threshold the rules use for cold mornings: not labelled.
- **04/10 08:00**: corr of the references ±3.4 noises, targets within 1.7: nothing on the curves of the targets.
- **The agitation after 09/10 12:00** (swings of all the targets from one measurement to the next, largest on 11/10 08:00-12:00 in corr_x, 11/10 20:00 and 12/10 08:00 in corr_y): the two-reference fragility described by MICHIGAN-3 and MICHIGAN-4. The largest swing (11/10 08:00 → 12:00) is about 4 noises of the post-loss step-to-step noise in corr_x, about 2 from the local level; smaller in corr_y than MICHIGAN-8 (17-18 October, labelled apart for a corr_y plunge). Not labelled apart; a reviewer may decide otherwise for 11/10 04:00 → 16:00.
- **Cold measurements of 11/10 04:00 and 12/10 00:00-08:00** (housing 31.75-31.9 °C): on two references the pose and the corr are too agitated to isolate a thermal response (tilt and corr_y within 1.5 noises of the post-loss noise); BREITHORN-31 labels 12 October at the neighbouring camera. No label here.
## 3. The documentation
- `labeling-rules.md` §2 ("each edge checked sensor by sensor"; a camera cycle ends at the later return of the pose and of the curves) moved the end of MICHIGAN-2; §1 ("a signature found once is searched over the whole batch") found 05/10; §3 ("an event in one piece over several days makes one label; a change of speed does not cut it") kept MICHIGAN-4 as one label across the base change.
- `families.md`, boundary "Thermal displacement or Displacement", and the phenomenon "drifts when the housing temperature drops, without a daily cycle (sheltered)" settled the family of 05/10.
- `tools/GUIDE.md`, `overview.py --rigid`, proved the targets move as a block both on 05/10 and on 08/10-12/10.
- The precedent NEBRASKA-1 says the clearer camera step of 7-8 October was absorbed by three references at NEBRASKA; at MICHIGAN it was not fully absorbed (common corr_x step from 08/10 04:00). Not a contradiction of NEBRASKA-1, a difference between the two cameras.
- Missing: a threshold in noises for a lasting Displacement remainder (the 5-noise rule is written for thermal cycles only); at a 4 h cadence the decision rests on persistence over several measurements and on the number of targets.
- The rule "the moments of the day are judged on the images" cannot be applied on this indoor station: the images are always dark. MICHIGAN-2's "at the end of the night" was removed for that reason.
## Difficulties
- `tools/GUIDE.md` (Conventions) restricts Bash to the tools; the trade allows scripts in `scripts/`: I used `scripts/steps.py`, `common.py`, `pertarget.py`, `pose.py`, `refgap.py`, `write_labels.py` (from the summary tables and the raw exports, read-only).
## Sources read
- `agents/agents/analyst.md`, `instructions/analyst.md`, `work/prompt.txt`, `work/existing.jsonl`, `work/spanning.jsonl`, `request.json`
- `tools/GUIDE.md`, `tools/DATA.md`; `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md` (in full)
- `labeling/auto/MICHIGAN.jsonl` (labels of the range and edges); `labeling/auto/BREITHORN.jsonl`, `NEBRASKA.jsonl` (labels of 3-16 October)
- `precedents.py` ×2 (cold-night common corr shift; drift on three references)
- Tools: `inventory.py`, `overview.py` (range; `--rigid` 05/10 20:00 and 04/10 20:00), `summary.py` (corr_x, corr_y, x, y, error_value), `ranges.py --k 3 --instants 3`, `metrics.py`, `plot.py`, `thumbnails.py --fixed` (05/10 12:00-16:00 pair; REFPOU22, REF42, REFG43N 09/10 08:00, 12:00, 12/10 16:00); whole images 05/10 00:00, 05/10 08:00, 06/10 04:00
- Raw exports `data/MICHIGAN/displacement/MICHIGAN_advanced_202610*.csv` (by `scripts/refgap.py`)
- Weather: none (no `weather/` folder for MICHIGAN)