119 lines
15 KiB
Markdown
119 lines
15 KiB
Markdown
# Labels: NORDEND, 2026-08-10T01:15Z → 2026-08-23T05:15Z (UTC)
|
||
|
||
This is a verification of the 49 existing labels (`work/existing.jsonl`, all reviewed by the user on 2026-11-17), plus the spanning NORDEND-24. Verdicts: **47 kept, 2 corrected (NORDEND-40, NORDEND-49), 0 removed, 0 added.** `labels.jsonl` holds the 49 labels: the kept ones exactly as they were (review field included, id removed), the two corrected ones without their review field.
|
||
|
||
Weather: `stations.txt` gives NORDEND → "Ouffet". `meteoswiss.py` resolved this to L'Ouffette (VS), station MAR Les Marécottes, 4.9 km. NORDEND is a quarry, and the place name may well be Ouffet in Belgium, so this weather is probably not the station's. It was not used as evidence; the images were.
|
||
|
||
## 0. Verdict per old id
|
||
|
||
| id | verdict | evidence |
|
||
|---|---|---|
|
||
| NORDEND-1 | kept | PR12 corr_x at 15 min: level about 0.06–0.08 until 01:15, dips from 01:30 (0.031) to 02:15 (0.017; lowest 0.005 at 02:00), back at 02:30 (0.051). The fast-deviation scan flags only PR12 x at 02:00 and 02:15. The bounds hold. |
|
||
| NORDEND-2 | kept | CTR01 rejected at 10/08 05:45 only (inventory). Image 05:45: sun in the field, scene dark. |
|
||
| NORDEND-3 | kept | Pose turning point 05:30 (tilt −117 → −115 → −105 m°, pan +9 → +6 → +4). Sun facing the camera on the 05:45 image. The chain to NORDEND-5 is at the next turning point. |
|
||
| NORDEND-4 | kept | CTR01 rejected 13:45 → 16:45 (13 measurements); label runs 13:30 → 17:00. |
|
||
| NORDEND-5 | kept | Turning point 05:45 (tilt −120 → −116, pan +13 → +12). Sun on the 06:30 image. |
|
||
| NORDEND-6 | kept | CTR01 rejected 11/08 06:30 only. Sun on the 06:30 image. |
|
||
| NORDEND-7 | kept | CTR01 rejected 13:00, then 13:45 → 16:45; label runs 12:45 → 17:00. |
|
||
| NORDEND-8 | kept | CTR01 rejected 12/08 06:00. Sun on the 06:00 image. |
|
||
| NORDEND-9 | kept | Turning point 05:45 (tilt −120 → −117 → −71). |
|
||
| NORDEND-10 | kept | CTR01 rejected 13:45 → 16:45. 14:30 image: sunlit rock face. |
|
||
| NORDEND-11 | kept | CTR01 rejected 13/08 06:00. Sun on the 06:00 image. |
|
||
| NORDEND-12 | kept | Turning point 05:45. Strongest pose day: pan lowest −0.050 on 13/08 15:00; housing temperature 58 °C, the highest of the batch. |
|
||
| NORDEND-13 | kept | CTR01 rejected 13:45 → 16:45. |
|
||
| NORDEND-14 | kept | CTR01 rejected 06:00 → 06:15; label runs 05:45 → 06:30. Sun on the 06:00 image. |
|
||
| NORDEND-15 | kept | Turning point 05:45. Ends at the next turning point, 15/08 05:15. |
|
||
| NORDEND-16 | kept | CTR01 rejected 13:45 → 16:45. |
|
||
| NORDEND-17 | kept | One step missing at 15/08 01:15. Images every 15 min over 13 → 19/08, so the image of that instant exists. |
|
||
| NORDEND-18 | kept | Turning point 15/08 05:15 (tilt −116 → −113 → −101). Cloudy jolts: the scan flags PR14, PR01, PR10 at 6–7 noises at 09:30, which the description covers. |
|
||
| NORDEND-19 | kept | CTR01 rejected 13:15, 13:45–14:00, 14:30–16:00, 16:45; label runs 13:00 → 17:00. 16/08 14:00 image: sunlit rock. |
|
||
| NORDEND-20 | kept | Overcast morning; slow turning point about 06:30 (tilt −111/−113/−110/−108). This is within a step of the uncertainty. |
|
||
| NORDEND-21 | kept | CTR01 rejected 16/08 09:30 only. |
|
||
| NORDEND-22 | kept | CTR01 rejected 13:15, 14:00–14:15, 14:45, 16:30–16:45; label runs 13:00 → 17:00. 14:00 image: sunlit rock. |
|
||
| NORDEND-23 | kept | PR07 rejected from 02:45, CTR01 05:00 → 10:15, PR09 05:00 → 08:45; label runs 02:30 → 10:30. The fast-deviation scan flags PR05, PR06, PR08, PR14, PR15, PR16 between 05:45 and 06:45, all inside the label. |
|
||
| NORDEND-25 | kept | Turning point 17/08 07:30 (tilt −119 → −107 at 07:45, pan +9 → +6). |
|
||
| NORDEND-26 | kept | CTR01 rejected 13:45, 14:15, 14:45 → 15:15; label runs 13:30 → 15:30. |
|
||
| NORDEND-27 | kept | Fast deviation on PR16 x at 06:30, so the 06:15 start holds. Rejections run to 12:00 (CTR01); end 12:15. 07:45 image: uneven light, no obstacle. |
|
||
| NORDEND-28 | kept | Turning point 08:15 (tilt −119 → −115 at 08:30). The night with CTR03 lost is correctly referred to its own label. |
|
||
| NORDEND-29 | kept | CTR01 rejected 15:30; PR07, PR09, PR11 15:45 → 16:15; fast deviation PR16 x at 16:30. Bounds 15:15 → 16:45 hold. |
|
||
| NORDEND-30 | kept | CTR01 rejections 17:45 → 03:45; label runs 17:30 → 04:00. 18/08 23:00 image: lit web. |
|
||
| NORDEND-31 | kept | CTR02 rejected 18:00 only, with 0 references retained at that instant. 15 sensors deviate together at 18:00 (scan). |
|
||
| NORDEND-32 | kept | CTR03 rejected 22:30 → 01:00 (11 measurements). Plateau of CTR02 y, CTR04 x, PR17, PR18 from 22:30 to 01:00, rebound at 01:15; 01:30 missing, so the end at 01:45 holds. |
|
||
| NORDEND-33 | kept | 01:30 missing; image gap 01:15 → 01:45 (inventory). |
|
||
| NORDEND-34 | kept | CTR01 rejected 07:00–07:15, 09:00, 10:30, 12:00; label runs 06:45 → 12:15. |
|
||
| NORDEND-35 | kept | Turning point 07:15 (tilt −115 → −109 at 07:30). Ends before the gap at 20/08 03:45, as the gap rule says. |
|
||
| NORDEND-36 | kept | 23 steps missing 14:30 → 20:00. Images missing over 14:15 → 20:15. No metric 15:00 → 19:00, voltage normal before and after. The claim about FLORIDA, MISSISSIPPI and VERMONT cannot be checked: no other station has August data. |
|
||
| NORDEND-37 | kept | CTR01 rejected 20:15–20:30 and 21:00 → 03:45; PR07 rejected 23:00 → 23:30. PR08 corr_x is low (0.004 to 0.04 against about 0.06 the day before) all night, most in the second half. "Departs in the second half" is a little narrow, but not wrong. |
|
||
| NORDEND-38 | kept | 20:45 missing; image gap 20:30 → 21:00. |
|
||
| NORDEND-39 | kept | 00:45 missing; image gap 00:30 → 01:00. |
|
||
| NORDEND-40 | **corrected** (description) | 9 steps missing, 04:00 → 06:00. But `20260820T0600Z_9852824.jpg` exists, and last_sync has a sync at 06:00:17, while the first measurement is at 06:15. The old text said "image and synchronisation come back with the first measurement": that is false by one step. The bounds stay. |
|
||
| NORDEND-41 | kept | Measurements resume at 06:15 with the pose already moved (tilt −91 m° against −120 the night before). PR18 corr_x rises all night (0.23 → 0.27). |
|
||
| NORDEND-42 | kept | 08:15, 08:30, 08:45 missing; image gap 08:00 → 09:00; only one sync in the 08:00 hour; voltage normal. |
|
||
| NORDEND-43 | kept | CTR01 rejected 20/08 14:30 only. |
|
||
| NORDEND-44 | kept | PR07 rejected 16:30. PR14, PR15, PR16 corr_x at 0.21 against 0.12 at 16:30, still high at 16:45 and 17:00 (0.14–0.16), back at 17:15. PR13 x at 16:00 is −4 to −6 noises on a single point; not enough to move the start. |
|
||
| NORDEND-45 | kept | 00:45 and 01:00 missing; image gap 00:30 → 01:15. |
|
||
| NORDEND-46 | kept | Turning point 05:15: pan leaves at 05:15, tilt at 05:00, the later of the two kept. |
|
||
| NORDEND-47 | kept | CTR01 rejected 12:30 and 13:15. |
|
||
| NORDEND-48 | kept | PR07 rejected from 16:00, CTR01 17:45 → 07:15, most targets 18:00–19:45. PR08 corr_x is at −0.07 at 20:00, about 0 overnight, 0.039 at 07:00 and back at 07:30 (0.044 against a day level of 0.045). |
|
||
| NORDEND-49 | **corrected** (description) | Bounds and family hold: turning point 06:00, then the next label at 23/08 05:15. But the cold-night part named PR10 alone. Most targets' corr_y rise from about 22:30, with a step at 00:00, and stay up until 05:30 (see §1). |
|
||
| NORDEND-50 | kept | 16:45 missing; image 16:45 exists. No metric 16:00 → 20:00, then DB_not_uploaded climbs from 86 to 229 until 23/08 07:00. |
|
||
| NORDEND-24 (spanning) | holds within the range | PR04: no rejection before 17/08 05:30 in the range; rejections on 17, 18, 20, 21 and 22/08, many on 21–22/08. corr_y is noisy and changes level from 17/08. Its start at 05:15 holds. |
|
||
|
||
## 1. My labels
|
||
|
||
### 2026-08-20T03:45Z → 2026-08-20T06:15Z (NORDEND-40, corrected)
|
||
- **Phenomenon of the catalogue**: The station runs, the measurement is missing (at its last step). It sits at the end of a longer silence whose cause is not established.
|
||
- **Precedents**: the other No data labels of the batch (NORDEND-17, 33, 42), written in the same style.
|
||
- **Family**: No data. All curves are missing; there are no rejections.
|
||
- **Measured**: inventory: 9 instants missing between 03:45 and 06:15. Images: one gap of 135 min, 03:45 → 06:00, then an image every 15 min from 06:00. last_sync: syncs every 15 min until 03:45:16, then one at 06:00:17. No metric in the 04:00 and 05:00 hours. BusVoltage is 14.06 V at 03:00 and 14.30 V at 06:00.
|
||
- **Images**: file list checked; `20260820T0600Z_9852824.jpg` exists.
|
||
- **Confidence**: high.
|
||
|
||
### 2026-08-22T06:00Z → 2026-08-23T05:15Z (NORDEND-49, corrected)
|
||
- **Phenomenon of the catalogue**: The camera moves through a thermal effect; for the night, the cold-night variant, as in its own next label NORDEND-51.
|
||
- **Precedents**: NEBRASKA-10 and NEBRASKA-11 (reviewed): the coldest night, the camera turns, the corr_y of most targets jump, and the references move slightly relative to each other; labelled Thermal displacement. MICHIGAN-15 (reviewed): the same, with one reference moving slightly away from the overall movement. These supported keeping the family. NORDEND-51 (outside the range, reviewed) says the next night is "like the previous one".
|
||
- **Family**: Thermal displacement. Why not Visibility: the error_value of the targets stays at or below its evening level (2 h medians); the spots of CTR02, CTR03 and CTR04 are sharp on the fixed thumbnails at 22/08 20:00, 23/08 00:00 and 03:00; the shift is smooth and follows the temperature, without sharp edges; and the curves come back with the morning warming. Why not Displacement: it comes back with the temperature.
|
||
- **Measured**:
|
||
- Housing temperature falls from 31.3 °C (21:00) to 26.1 °C (05:00), the lowest of the batch.
|
||
- Pose: scale 0.99903 at 04:00 (lowest of the batch), tilt −0.132 (lowest), pan +0.022 (highest); cam_sigma0 up to 0.00207 at 03:00 (highest of the batch).
|
||
- `overview.py --rigid 2026-08-22T21:00Z`: target residual median about 0.01 px, unchanged, so the targets do not move relative to each other. CTR04 dy slides to −0.04/−0.05 px from 00:00 (−0.0503 at 01:45). CTR02 dy rises to +0.02 (+0.0248 at 05:00). reproj_err of CTR02 and CTR03 rises from 0.09 to 0.12.
|
||
- corr_y, 30 min medians, 21:00 → 02:00: PR15 0.017 → 0.051; PR14 0.019 → 0.055; PR07 0.091 → 0.123; PR10 0.124 → 0.147; PR06 0.082 → 0.104; CTR02 0.092 → 0.117. These are about 4 to 6.6 noises.
|
||
- Back at 06:30–07:30 on 23/08 as the housing warms up (PR15 0.0007, PR14 0.0076 at 07:30).
|
||
- No fast deviation in the scan (rolling ±4, 6 noises) during the night.
|
||
- **Images**: 22/08 20:00 and 23/08 03:00 whole images: lit cobweb threads cross the image and the area of the targets. Reference thumbnails as above.
|
||
- **Confidence**: medium. The family and the bounds are sure. That the web plays no part in CTR04's slide rests on its sharp spot and its steady error_value.
|
||
|
||
## 2. What I left without a label
|
||
- **Night 20→21/08 (under NORDEND-41)**: lit web threads cross the targets (23:00 image), but no target except PR04 is rejected, and PR04 is under NORDEND-24. The corr stay within noise, except PR09 corr_y at +5.3 noises for 04:30–04:45 against its level outside the night. That is two points at the threshold; PR09 is noisy (noise 0.0088) and there is no fast deviation at 6 noises. Not labelled. NORDEND-41 does not mention the web, which is correct, since the web touches nothing.
|
||
- **Single-sensor fast deviations of 6 to 8 noises outside non-thermal labels**:
|
||
- 10/08 06:00 (PR16–18): the sunrise jump, covered by NORDEND-3.
|
||
- 11/08 11:15 PR09, 12/08 08:15 PR17, 14/08 08:30 PR10, 15/08 06:15 PR18, 15/08 09:30–09:45 (cloudy jolts, NORDEND-18), 19/08 05:15 PR12, 21/08 09:15–09:45 PR15/PR18, 21/08 14:30 CTR02, 22/08 16:15 CTR02: one point, one sensor each, inside the daily thermal agitation.
|
||
- PR04 on 18/08 14:30 and 20/08 11:00 and 13:00: covered by NORDEND-24.
|
||
- **PR13 corr_x at 20/08 16:00**: one point just before NORDEND-44, at −4 to −6 noises; it does not move the start.
|
||
- **12/08 23:00**: a bright blob at the top of the image, probably an insect; no effect on corr.
|
||
|
||
## 3. The documentation
|
||
- **Helpful**:
|
||
- `labeling-rules.md` §2 for the turning-point starts, the bounds at gap edges and "a measurement come back halfway is still touched".
|
||
- `families.md` boundary "Thermal displacement or Visibility". It cut both ways on 22→23/08, because references do move apart from each other there. The reviewed NEBRASKA and MICHIGAN precedents decided it, together with the steady error_value and the sharp references.
|
||
- `reading-rules.md` §5 ("a very stable error_value says the targets stay well visible").
|
||
- **Missing**:
|
||
- The catalogue's "A spider web in front of the lens" says "no signal at all is possible". It does not say what to do when the web is seen but the curves are untouched. The rule "only what touches corr" answers it: no label. I applied that to 20→21 and 22→23/08.
|
||
- `stations.txt` "Ouffet" geocodes to the Valais. The weather of NORDEND is probably not available this way; this should be checked by whoever owns `stations.txt`.
|
||
- **Not verifiable**: NORDEND-36 and NORDEND-50 cite "the data documentation" for gaps at FLORIDA, MISSISSIPPI and VERMONT. Nothing in my sources says so, and no other station has August data. I kept them (reviewed); this phrase is the only unverified part.
|
||
- No precedent was contradicted by the data.
|
||
|
||
## Difficulties
|
||
- `meteoswiss.py` refuses times with a trailing `Z`; it was run with `2026-08-10T00:00`.
|
||
- `metrics.py` needs `.venv/bin/python` (matplotlib), not `python3`.
|
||
|
||
## Sources read
|
||
- `instructions/analyst.md`, `tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`.
|
||
- `work/existing.jsonl`, `work/spanning.jsonl`, `labeling/auto/NORDEND.jsonl` (labels at the edges: NORDEND-51 to 53).
|
||
- Precedents: `precedents.py` on the cold-night signature (MICHIGAN-2, MICHIGAN-15, NEBRASKA-10, NEBRASKA-11, WYOMING-7, NORDEND-173).
|
||
- Data, through `tools/scripts/data.py` and the tools:
|
||
- `inventory.py`; `overview.py` (whole range, then `--rigid` on 22/08 12:00 → 23/08 05:15); `ranges.py` (`--rolling 4 --instants 3`; `--outside` on night 20→21); `summary.py` (corr_x, corr_y, error_value on the nights and windows cited); `metrics.py` (10 → 23/08, then around the gaps); `thumbnails.py` (references and PR15, `--fixed`).
|
||
- Whole images: 22/08 20:00, 23/08 03:00; sheets of the 23:00 images of every night and of sunrise and afternoon images (`work/nights23.png`, `work/sunrise.png`).
|
||
- Weather: `work/weather.csv` (MAR Les Marécottes, 4.9 km, see above).
|
||
- My scripts: `scripts/fast.py` (fast deviations against coverage by the labels), `scripts/morning.py` (morning pose table), `scripts/sheet.py` (contact sheets), `scripts/write_labels.py`.
|