auto: the analysts verified the labels of 32 periods (11 stations, up to 14 days each): 710 labels -> 835, 587 unchanged; one analysis folder per period with its report (section Verification)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
1 parent
df979301b5
commit
2a5d87659b
279 files changed
+10758
-710
No files matched your search
@@ -0,0 +1,100 @@
|
||||
# Labels: MICHIGAN, 2026-10-23T20:00Z → 2026-11-01T00:00Z (UTC)
|
||||
|
||||
Request: verify the 6 existing labels of the range (`work/existing.jsonl`, all validated by the user on 2026-11-17), keep / correct / remove / add; the 2 spanning labels (`work/spanning.jsonl`) are context only. DETECTORS: no.
|
||||
|
||||
Station context: indoors, in the same building as BREITHORN and NEBRASKA (according to the expert); one measurement every 4 h (50 instants in the range, none missing, every instant has a camera pose); REFPOU22 rejected at all 50 measurements, so the compensation runs on REF42 and REFG43N only (reprojection error at zero all along); no image-quality metrics (export columns peak to snr empty); `sigma0` empty. **Weather**: no `weather/` folder for MICHIGAN, so no MeteoSwiss data for this range (no MeteoSwiss station name or distance to give); the station is indoors anyway. Images: one per measurement, all dark (indoor scene lit by the camera), spots sharp.
|
||||
|
||||
## Verdicts on the existing labels
|
||||
|
||||
| Old id | Verdict | Evidence |
|
||||
|---|---|---|
|
||||
| MICHIGAN-10 | **corrected** (end 24T20 → 25T00, description) | the corr_x of all targets is lowest at 24T20 (COLH21 0.138, COLH18' 0.266, median of the targets −0.04 against +0.12 at 25T00); the camera scale is still low at 24T20 (0.99987, as all of 24 Oct) and comes back at 25T00 (1.00003) with the corr_x; the hourly housing temperature is still 29.0–29.25 °C at 21–23 h and back to 29.5–29.75 at 25T00–03. 24T20 is a touched measurement, so the end is 25T00. corr_y is back at 24T20, as the label said. |
|
||||
| MICHIGAN-11 | kept | corr_y drops from 25T20 (median 0.06 → 0.01), lowest 26T04 (−0.15); corr_x drops 26T04–26T08 (COLH21 0.32 → 0.07), 26T12 halfway back, 26T16 back on both; tilt 0.0240 → 0.0378 → 0.0254, scale down to 0.99980; housing 29.5 → 28.5 °C then 30 °C at 26T14. Image 26T04 clean. |
|
||||
| MICHIGAN-12 | kept | corr_y of all targets up from 27T00 (median +0.10, +0.20 at 27T04, +0.13 at 27T16), back at 27T20/28T00; corr_x flat; tilt drops to 0.0180 at 27T04 while the housing temperature stays at 30–30.5 °C as on 26T16–20 (tilt 0.0254–0.0264 there): not explained by the temperature at equal value. Vertical translation of the targets (rigid `ty`) unchanged (0.04–0.05 px as on 26T16–28T00), so the pose change comes from the two references. NEBRASKA tilt flat that day (0.038–0.039), BREITHORN flat (0.0285–0.0288). Image 27T04 clean. Displacement, cause not established, holds. |
|
||||
| MICHIGAN-13 | kept | corr_x 28T04 a little down (COLH21 0.31 → 0.22), 28T08 plunge (−0.04), 28T12 back; corr_y 28T08 −0.11, 28T12 still −0.08, 28T16 back; housing 29.25 → 27.5 °C (06–08 h) → 30 °C at 12 h; tilt 0.0264 → 0.0388, scale 0.99971. Image 28T08 clean. |
|
||||
| MICHIGAN-14 | kept | corr_x of all targets 28T20 → 29T08 down (median 0.00 → −0.39), corr_y down to −0.23; 29T12 corr_x up to a lower plateau (−0.08, kept until 01T00), corr_y back at 29T16; housing 29.5 → 25.75 °C (strongest drop, 29T05–07) → 28.5–29 °C at 29T13–14; tilt up to 0.0507, scale 0.99945; rigid: REFG43N dx off the targets (+0.10 to +0.155 px); error_value of COLH21'', COLH21', REFH22', COLH20' up 10–15 noises, far from the 0.7 threshold. Image 29T08 clean. |
|
||||
| MICHIGAN-15 | kept | corr_y 29T20 a little down (−0.07); 30T00–30T08 corr_x down to −0.38 (30T04), corr_y lowest −0.23 at 30T08; 30T12 back (corr_x −0.06 = the plateau level of 30T16–01T00; corr_y −0.06 vs −0.07/−0.03 after); housing 28.25 → 25.75 °C (30T06) → 29.5 °C (30T13); tilt 0.0505, scale 0.99945; REFG43N dx +0.13 px. Image 30T04 clean. |
|
||||
|
||||
Added: **2026-10-31T00:00Z → 2026-10-31T16:00Z**, Thermal displacement (section 1).
|
||||
|
||||
## 1. My labels
|
||||
|
||||
### 2026-10-23T20:00Z → 2026-10-25T00:00Z (MICHIGAN-10, corrected)
|
||||
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle; A compensation on two references.
|
||||
- **Precedents**: MICHIGAN-2, MICHIGAN-9 (validated): same station, same signature (corr_y of all targets down for one or two measurements with a housing temperature drop, camera nods forward), labelled Thermal displacement, bounded from the measurement before to the one after the curves return.
|
||||
- **Family**: Thermal displacement: the pose (tilt up, scale down) and the curves follow the housing temperature and come back with it; not Displacement (it comes back at equal temperature), not Visibility (error_value flat, image clean).
|
||||
- **Measured**: corr_y median of the targets (metrics.py): 23T20 +0.03, 24T00 −0.09, 24T04 −0.06, 24T08 0.00, 24T12 +0.02, 24T16 −0.03, 24T20 +0.09. corr_x (summary.py, COLH21): 23T12 0.30, 23T16 0.23, 23T20 0.17, 24T00–16 0.14–0.18, **24T20 0.138**, 25T00 0.314; all 14 targets lowest or near-lowest at 24T20 in corr_x and back at 25T00. Pose: tilt 0.0278 → 0.0348 (24T00) → 0.0292 → 0.0320 (24T16) → 0.0270 (24T20) → 0.0228 (25T00); scale 0.99991 → 0.99986–0.99989 → 0.99987 (24T20) → 1.00003 (25T00). Housing temperature hourly: 29.5–31 at 23T18–21, 28.5–28.75 at 24T01–07, 29–30.25 at 24T08–12, 28.75–29.25 at 24T13–23, 29.5–29.75 at 25T00–04. Rigid (base 23T20): target residual median ≈0.01 px, references within their usual residuals. The corr_x decline starts at 23T16, before the range: there is no label of the previous batch to extend (the labels before stop at 19T12), so the label starts at the range edge (labeling-rules 2, last rule). error_value flat.
|
||||
- **Images**: 24T00 and 24T20 (zone of all the spots): identical to the other days, nothing in front.
|
||||
- **Confidence**: medium: small amplitude (about two noises on the median of the targets), but the same signature as the larger cold events of 28–30 Oct, pose and scale following the temperature.
|
||||
|
||||
### 2026-10-25T16:00Z → 2026-10-26T16:00Z (MICHIGAN-11, kept)
|
||||
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle.
|
||||
- **Precedents**: MICHIGAN-9.
|
||||
- **Family**: Thermal displacement (as above).
|
||||
- **Measured**: see the verdict table; COLH21 corr_x 26T00 0.32 (a little high), 26T04 0.12, 26T08 0.07, 26T12 0.13, 26T16 0.26; corr_y COLH21 25T16 +0.045, 25T20 −0.045, 26T00 −0.117, lowest 26T04/08, back 26T16.
|
||||
- **Images**: 26T04: nothing in front.
|
||||
- **Confidence**: high.
|
||||
|
||||
### 2026-10-26T20:00Z → 2026-10-27T20:00Z (MICHIGAN-12, kept)
|
||||
- **Phenomenon of the catalogue**: A compensation on two references; The base changes without a reference lost (closest).
|
||||
- **Precedents**: MICHIGAN-8 (validated, same station: agitation of the pose computed on two references without raw movement of the targets, Displacement, cause not established).
|
||||
- **Family**: Displacement: the tilt changes at equal housing temperature (30–30.5 °C from 26T16 to 27T20, hourly), so the thermal boundary ("comes back at equal temperature") is not met.
|
||||
- **Measured**: see the verdict table; stations.py: NEBRASKA tilt 0.0383–0.0394, BREITHORN 0.0283–0.0288 on 26T16–28T00.
|
||||
- **Images**: 27T04: nothing in front.
|
||||
- **Confidence**: medium (cause not established, as the label says).
|
||||
|
||||
### 2026-10-28T00:00Z → 2026-10-28T16:00Z (MICHIGAN-13, kept)
|
||||
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle.
|
||||
- **Precedents**: MICHIGAN-9.
|
||||
- **Family**: Thermal displacement.
|
||||
- **Measured**: see the verdict table.
|
||||
- **Images**: 28T08: nothing in front.
|
||||
- **Confidence**: high.
|
||||
|
||||
### 2026-10-28T16:00Z → 2026-10-29T16:00Z (MICHIGAN-14, kept)
|
||||
- **Phenomenon of the catalogue**: idem; A compensation on two references (REFG43N off the targets at the cold peak).
|
||||
- **Precedents**: MICHIGAN-9.
|
||||
- **Family**: Thermal displacement; the lower corr_x plateau after 29T12 belongs to the spanning MICHIGAN-7 (said in both labels).
|
||||
- **Measured**: see the verdict table.
|
||||
- **Images**: 29T08: nothing in front.
|
||||
- **Confidence**: high.
|
||||
|
||||
### 2026-10-29T16:00Z → 2026-10-30T12:00Z (MICHIGAN-15, kept)
|
||||
- **Phenomenon of the catalogue**: idem.
|
||||
- **Precedents**: MICHIGAN-14 (the day before).
|
||||
- **Family**: Thermal displacement.
|
||||
- **Measured**: see the verdict table.
|
||||
- **Images**: 30T04: nothing in front.
|
||||
- **Confidence**: high.
|
||||
|
||||
### 2026-10-31T00:00Z → 2026-10-31T16:00Z (added)
|
||||
- **Phenomenon of the catalogue**: The camera drifts away when the temperature of its housing drops, without a daily cycle.
|
||||
- **Precedents**: MICHIGAN-2 (validated: a single measurement, "a little", labelled Thermal displacement), MICHIGAN-9 and MICHIGAN-10: same signature, weaker here; they show that small drops are labelled on this station.
|
||||
- **Family**: Thermal displacement: the pose and the curves follow a small drop of the housing temperature and come back with it.
|
||||
- **Measured**: corr_y of all 14 targets lower together at 31T04, lowest at 31T08, partly back at 31T12, back at 31T16 (at 00/04/08/12/16/20: COLH21 −0.044 / −0.087 / −0.160 / −0.088 / −0.070 / −0.045; POUH18' 0.168 / 0.076 / 0.047 / 0.095 / 0.143 / 0.167; COLH20' −0.055 / −0.110 / −0.139 / −0.084 / −0.061 / −0.049); median of the targets −0.02 / −0.06 / −0.11 / −0.06 / −0.04 / −0.01 (metrics.py). corr_x median −0.05 / −0.09 / −0.04 / −0.09 / −0.03 / −0.03. Pose: tilt 0.0335 / 0.0375 / 0.0390 / 0.0368 / 0.0341 / 0.0330; rotation −0.0255 / −0.0280 / −0.0310 / −0.0276 / −0.0268 / −0.0251; scale 0.99983 / 0.99977 / 0.99982 / 0.99980 / 0.99986. Rigid `ty` of the targets 0.100 → 0.147 / 0.148 / 0.123 → 0.103 px: the targets move together in the image (the camera nods), target residual median 0.014–0.016 px as usual; REF42 dy −0.073 at 31T08 (usual −0.04 to −0.06). Housing temperature hourly: 28.75 at 31T00, 28–28.5 at 01–06, 27.75 at 07, 28 at 08, 29 at 09, 28.5–29.5 at 10–12. error_value flat, no rejection except REFPOU22. Amplitude about 1.5 noises on the median of the targets: weaker than 24 Oct (about 2), but coherent on all 14 targets, with the pose and the temperature, over three measurements. Bounds: 31T00 at level (corr_y median −0.02, tilt as 30T20), 31T16 back (tilt 0.0341, corr_x −0.03, corr_y −0.04 within noise of the −0.01/−0.02 level).
|
||||
- **Images**: 31T00 and 31T08 (zone of all the spots, and the whole image of 31T08): identical, nothing in front.
|
||||
- **Confidence**: medium: small, but the same signature as 24 Oct, treated the same way (labeling-rules 1, "a signature found once is searched over the whole batch").
|
||||
|
||||
## 2. What I left without a label
|
||||
- **25T12**: one measurement, the corr_y of all targets a little lower (median −0.04 vs +0.04/+0.06 around) with the tilt up a little (0.0295 vs 0.0264/0.0240); no drop of the housing temperature (29.25–29.5 °C). The size of the measurement-to-measurement wobble of the pose on this two-reference compensation (as on 20T04, 21T08, 23T08, unlabelled before the range). Not labelled.
|
||||
- **01T00**: the housing temperature starts going down (28.25 °C) and the tilt rises a little (0.0351 vs 0.0330), corr_y a little down (−0.04): the start of the cold event of 1 November (01T04: corr_x −0.13, corr_y −0.11, tilt 0.0436), outside the range; its label (start at 31T20 or 01T00) belongs to the next range.
|
||||
- The slow trends (tilt higher, corr_x lower plateau after 29T12, corr_y slowly lower) are carried by the spanning MICHIGAN-7; REFPOU22 by MICHIGAN-3.
|
||||
- The error_value rises of 29–30 Oct are inside MICHIGAN-14/15 and far from the threshold.
|
||||
|
||||
### Spanning labels, within the range
|
||||
- **MICHIGAN-3** (REFPOU22 lost): holds: REFPOU22 rejected at all 50 measurements of the range, raw positions wandering (up to +55 px); compensation on two references, reprojection error at zero.
|
||||
- **MICHIGAN-7** (camera tilt, Displacement): holds within the range: tilt from ≈0.025 (24–28 Oct) to ≈0.034 (30–31 Oct), corr_x high until 28 Oct then on a lower plateau from 29T12 to the end; corr_y lower at the end (≈ −0.02 vs ≈ +0.05). Nothing to change.
|
||||
|
||||
## 3. The documentation
|
||||
- `labeling-rules.md` section 2 ("a measurement that has come back halfway is still touched", "the later of the two returns") decided the MICHIGAN-10 correction: the corr_x return, not the corr_y one, sets the end.
|
||||
- The rule "no thermal label when neither the pose nor the corr deviate by more than 5 noises from the night level" does not fit this station (4 h cadence, indoors, no day/night): `days.py --bounds` puts the 20–3 h night levels across the events themselves, and none of the validated labels 10, 11, 12 exceeds 5 noises there. I followed the validated precedents of the station (MICHIGAN-2, -9) instead. A sentence for indoor stations at 4 h cadence would help.
|
||||
- `reading-rules.md` 5 (neighbouring cameras read together) and `stations.py` confirmed MICHIGAN-12 (NEBRASKA, BREITHORN flat that day).
|
||||
- `tools/DATA.md`, the weather: MICHIGAN has no weather folder.
|
||||
- The hourly housing temperature shows one-hour spikes at the measurement hours (00, 04, 08…, +1 to +1.5 °C, the computing heating the board): read the hours around a measurement, not the hour of the measurement.
|
||||
- No precedent contradicted by the data.
|
||||
|
||||
## Difficulties
|
||||
none
|
||||
|
||||
## Sources read
|
||||
`agents/agents/analyst.md`, `instructions/analyst.md`, `work/prompt.txt`, `work/existing.jsonl`, `work/spanning.jsonl`, `tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`, `labeling/auto/MICHIGAN.jsonl` (ids, bounds, descriptions), precedents (`precedents.py`: MICHIGAN-2, -8, -9, JUNGFRAU-112, NEBRASKA-1, WYOMING-4). Tools on `data/MICHIGAN`: `inventory.py`, `overview.py --rigid 2026-10-23T20:00Z` (common, rigid, residuals PNGs), `metrics.py` (4 h and 1 h, 19 Oct–2 Nov), `summary.py` (corr_x, corr_y, error_value at 4 h), `days.py --bounds`, `stations.py` (MICHIGAN, BREITHORN, NEBRASKA), `thumbnails.py --zone` (24T00, 24T20, 26T04, 27T04, 28T08, 29T08, 30T04, 31T00, 31T08), whole image 31T08, `plot.py`. Script: `scripts/write_labels.py`.
|
||||
Reference in new issue
Block a user