Files
aimsight-labeling/auto/analyses/NORDEND_2026-10-24T1630_2026-11-07T1400/analysis.md
T

156 lines
18 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: NORDEND, 2026-10-24T16:30Z → 2026-11-07T14:00Z (UTC)
Request: verify the 29 existing labels of the range (`work/existing.jsonl`): keep, correct, remove, add. DETECTORS: no. Weather: **no `data/NORDEND/weather/` folder** (no MeteoSwiss data for this station); the weather is read on the images only.
Main finding: the existing labels came from two batches, one ending at 2026-11-01T00:00Z and one starting at 2026-11-04T21:15Z. **The 3.6 days in between (01/11 00:00 → 04/11 21:15) had no label at all**, although the measurements are complete there (inventory: no missing step in the range besides the three common gaps of 26, 27 and 29/10). Four labels were cut artificially at these batch edges ("until the end of the batch", "from the start of the batch"); they are corrected, and the hole is labelled.
## Verdict per old id
| Old id | Verdict | Evidence |
|---|---|---|
| NORDEND-215 | kept | Rejections of almost all targets and of CTR02/03/04 from 24/10 16:45 to 25/10 06:00, PR08 07:00–07:45, no pose 22:15–05:30 (inventory); targets back by 08:00–08:30, the remaining offsets of PR09/10/11 are NORDEND-216. Bounds and text hold. |
| NORDEND-216 | **corrected** (end 01/11 00:00 → 07/11 14:00, text) | Night levels (`days.py`, 0–4 h): PR10 corr_y about 0.11–0.13 before the rain, 0.19 after, still 0.16–0.19 on 02–06/11; PR09 corr_x about −0.005 before, −0.055 after, still −0.053 to −0.059 in November; PR11 corr_x 0.06 → 0.09, still 0.09. Drop marks still visible on the 02/11 22:00 image. The offset lasts to the end of the range, so the label too (labeling-rules §2: without a return, to the end of the batch). |
| NORDEND-217 | kept | Weak pose cycle 25/10, PR18 corr_x dip, consistent with the overview. |
| NORDEND-218 | kept | Pan not back on the morning of 27/10 (common PNG), corr flat. |
| NORDEND-219 | kept | Two steps missing 26/10 14:30–14:45 for all sensors (inventory). |
| NORDEND-220 | kept | Cycle 27/10 until the night of degrading visibility. |
| NORDEND-221 | kept | One step missing 27/10 09:00 (inventory). |
| NORDEND-222 | kept | PR07 rejected from 16:45 (27/10), PR08/PR09/PR15/PR18 at night, CTR03 06:45–07:15, CTR02 07:15, no pose at 07:15, last rejections at 09:00 → end 09:15. |
| NORDEND-223 | kept | Pose cycle 28/10, PR18 corr_x dip; the single PR04 rejection at 11:45 inside it was unlabelled → added apart. |
| NORDEND-224 | kept | Cycle 29/10. |
| NORDEND-225 | kept | One step missing 29/10 15:00 (inventory). |
| NORDEND-226 | kept | Cycle 30/10 (pan min −53 noises at 13:45, scale +50 noises; `days.py --bounds`). |
| NORDEND-227 | kept | Rejections 30/10 09:00 (PR18) → 10:30 (PR09), CTR02/CTR03 09:30–10:15, no pose 09:30–09:45 → 08:45–10:45 holds. |
| NORDEND-228 | **corrected** (end 01/11 00:00 → 01/11 06:00, text) | The pan keeps rising and the tilt dropping all night until 01/11 06:00 (pan max 0.0385 deg, tilt min −0.1519 at 06:00), then the next cycle starts; 00:00 was the batch edge, not a turning point. |
| NORDEND-229 | **corrected** (start 04/11 21:15 → 03/11 23:00, text) | 21:15 was the batch edge. The housing temperature rises again from 03/11 22:00 (22.4 → 24.4 °C at 06:00 → 28 at 14:00), the pan peaks at 03/11 23:00 (0.0311) and drops all night and morning to 0.0143 at 04/11 14:00, then comes back until 05/11 06:30 (NORDEND-234 starts there): one cycle, labelled from its start. |
| NORDEND-230 | **corrected** (start 04/11 21:15 → 01/11 00:00, text) | CTR01 is rejected through the hole: 01/11 06:00 → 02/11 16:00 (137 steps), then most of 02–04/11 (inventory). Its regime since 16/10 (NORDEND-195, spanning, ends 01/11 00:00) continues; one label per target and per batch, joined to 195. |
| NORDEND-231 | kept | CTR03 rejected 05/11 01:00–01:15, comb in corr_y. |
| NORDEND-232 | kept | PR09 rejected 05/11 03:15, 05:45, 06:15, 07:15, 07:45, 08:15–08:45, 09:30. |
| NORDEND-233 | kept | CTR03 rejected 05/11 03:30. |
| NORDEND-234 | kept | Cycle 05/11 (pan −44 noises at 14:15, tilt +32). |
| NORDEND-235 | kept | PR09 rejected 05/11 15:00. |
| NORDEND-236 | kept | CTR03 rejected 23:15 → 06/11 09:45 (14 measurements). |
| NORDEND-237 | kept | Cycle 06/11, pan not back next morning (−10.7 noises next night). |
| NORDEND-238 | kept | PR10 rejected 06/11 09:30. |
| NORDEND-239 | kept | CTR03 rejected 06/11 21:30 → 07/11 10:45. |
| NORDEND-240 | kept | Cycle 07/11 to the end of the range. |
| NORDEND-241 | kept | PR04 rejected 07/11 11:30. |
| NORDEND-242 | kept | PR07 rejected 07/11 14:00 (last measurement). |
| NORDEND-243 | kept | One-measurement pose jolt 05/11 12:00, corr_y of PR01 0.162 → 0.123 → 0.161. |
Spanning labels (not changed, not in labels.jsonl): NORDEND-152 (PR04) holds within the range (rejections stop 26/10 11:30; label to 20:45). NORDEND-169 (PR17) and NORDEND-195 (CTR01) end at 2026-11-01T00:00Z, a batch edge, while their condition continues: PR17's level stays the same (night corr_x 0.113–0.125, corr_y −0.004 to −0.016 on 01–04/11, as on 26–31/10) and CTR01 stays rejected. Their continuation is labelled here (new PR17 label from 01/11 00:00; NORDEND-230 moved to 01/11 00:00).
## 1. My labels
Only the corrected and added labels are detailed; the kept ones are justified in the table above.
### 2026-10-25T05:45Z → 2026-11-07T14:00Z (NORDEND-216, corrected)
- **Phenomenon of the catalogue**: The soiled glass
- **Precedents**: the label itself (reviewed); NORDEND-169 for "the curves stay shifted after the error_value comes back"
- **Family**: Visibility, unchanged (marks over a part of the scene, several targets and the references)
- **Measured**: night levels above (`work/days_nights.txt`); the offsets of PR09, PR10, PR11 do not come back on any night of 01–06/11 (07/11 is skewed by the CTR03 spikes)
- **Images**: 02/11 22:00 whole image: the blurred drop discs are still there, next to new cobweb threads
- **Confidence**: high
### 2026-10-28T11:30Z → 2026-10-28T12:00Z (added)
- **Phenomenon of the catalogue**: A faint spot lost when the light is missing
- **Precedents**: NORDEND-104, NORDEND-117 (one rejection of a weak target in daylight, spot barely visible, Visibility, reviewed); NORDEND-24 (PR04 a weak target)
- **Family**: Visibility, as the precedents; a single measurement with nothing in front, not a Target issue (no regime of rejections of PR04 in the range)
- **Measured**: PR04 error_value 0.046 (11:30) → 0.490 rejected (11:45) → 0.143 (12:00) → 0.054; neighbours PR03, PR05, PR12 unchanged; corr of PR04 at 12:00 already at the level of 12:15
- **Images**: thumbnails 28/10 11:00 and 12:00 (`--contrast`): PR04 spot very faint at 12:00, PR05, PR16, PR17 sharp
- **Confidence**: medium (cause not established)
### 2026-10-31T07:45Z → 2026-11-01T06:00Z (NORDEND-228, corrected)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NORDEND-226, NORDEND-224 (same days, reviewed)
- **Family**: Thermal displacement
- **Measured**: 15-min table of the overview: pan 0.0345 (04:00) → 0.0385 (06:00) → 0.0368 (06:30); tilt −0.149 → −0.1519 → −0.1499; housing temperature 21.6 °C at 00:00 → 19.75 at 06:00, still falling
- **Images**: none new
- **Confidence**: high
### 2026-11-01T00:00Z → 2026-11-07T14:00Z (PR17, added)
- **Phenomenon of the catalogue**: An obstacle covers part of the spot (mark on the glass in front of one spot)
- **Precedents**: NORDEND-169 (spanning, reviewed): the same target, the same mark, until the batch edge
- **Family**: Obstruction, the family of 169 (labeling-rules §4: a same target keeps its family)
- **Measured**: night levels of PR17 corr_x: 0.08–0.09 before 08/10, 0.112–0.125 from 26/10 to 04/11, 0.125–0.135 on 05–07/11; corr_y: −0.05 before 05/09, +0.09 to +0.12 on 06/09–08/10, −0.004 to −0.016 on 26/10–04/11; error_value 0.056–0.063 since 26/10, its level before 08/10 (0.06–0.07); neighbour PR16 stable
- **Images**: PR17 240 px thumbnails, `--contrast 40,120 --fixed`: 07/10 02:00 (clean), 26/10 03:00 (pale disc under the spot), 02/11 03:00 (a cobweb band over it), 04/11 03:00 (pale area under and left of the spot), 07/11 03:00 (darker exposure, a thread on the left)
- **Confidence**: medium: the offset is clear and continuous with 169; that the remaining level is the mark and not a return is not established, as 169 already says
### 2026-11-01T06:00Z → 2026-11-02T06:30Z (added)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NORDEND-218 (camera not back overnight, reviewed), NORDEND-237
- **Family**: Thermal displacement: the pose follows the housing temperature and comes back at equal temperature (03/11 22:00, 22.4 °C, pan 0.030, tilt −0.142, as on 31/10 22:00 at 22.7 °C, pan 0.032, tilt −0.147)
- **Measured**: housing 19.75 °C (06:00) → 28 (14:00), then 27 ± 0.5 all night (unusual: the nights before fell to about 21); pan 0.0385 → 0.0145 (12:30, −35 noises), back only to 0.019 at 20:00 and flat until 02/11 04:30; tilt −0.152 → −0.121 (11:15, +34 noises); rotation −0.0015 → −0.0094; scale 0.99878 → 0.99908. Corr (1 h medians, `summary.py --instants`): PR18 corr_x 0.303 → 0.259 → 0.290; PR13 corr_x 0.090 → 0.063 → 0.072; PR17 0.124 → 0.104 → 0.117; PR04 corr_x 0.025 → 0.069, corr_y 0.100 → 0.145; PR13 corr_y 0.093 → 0.118; PR15 corr_y 0.018 → 0.038
- **Images**: 01/11 22:00 whole image: cobweb threads lit across the scene and the group, spots sharp, no fog
- **Confidence**: high on the family; the cause of the warm night is not known (no weather data)
### 2026-11-02T06:30Z → 2026-11-03T07:00Z (added)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NORDEND-217 (weak cycle)
- **Family**: Thermal displacement (pose beyond 5 noises: pan −9.3, tilt +8.8 noises; labeling-rules §1)
- **Measured**: housing 26–28 °C; pan 0.0198 (06:30) → 0.0147 (14:30) → 0.0196 (21:00); start at the pan's departure 06:30–06:45; end at the next morning turning point 03/11 07:00 (pan max 0.0228, tilt min −0.1329 at 06:30). Corr: PR18 corr_x 0.290 → 0.261; PR04 corr_x 0.044 → 0.070, corr_y 0.111 → 0.144; PR09 corr_y 0.141 → 0.164; PR17 corr_y −0.011 → 0.015
- **Images**: 02/11 22:00 whole image
- **Confidence**: high
### 2026-11-02T19:15Z → 2026-11-03T06:00Z (added)
- **Phenomenon of the catalogue**: A spider web in front of the lens (a thread hides a few spots: Obstruction)
- **Precedents**: NORDEND-232, NORDEND-236, NORDEND-242 (cobweb threads lit by the infrared over a target, Obstruction, reviewed)
- **Family**: Obstruction: two targets touched, the rest intact
- **Measured**: PR07 error_value 0.15 → 0.94 rejected at 19:30, rejected 19:30–22:30, 23:45–00:15, 01:30, 02:00, 02:30, 03:45–04:30; kept points: corr_x 0.124 at 23:30 (level 0.075), 0.061 → 0.014 from 02:15 to 05:15, 0.034 at 05:30, 0.065 at 05:45, 0.071 at 06:00; corr_y 0.10–0.11 → 0.077–0.09. PR08 corr_x 0.054 (20:00) → 0.044 (20:15) → 0.034 (21:15) → 0.023–0.045 through the night, 0.033 at 05:45, 0.054 at 06:00; corr_y −0.03/−0.04 → −0.05/−0.06; error_value 0.05 → up to 0.13. PR06, PR13, PR14, PR03 flat
- **Images**: thumbnails PR07, PR08, PR06, PR13 at 02/11 18:00, 20:00, 22:00, 03/11 01:00, 04:00, 06:00 and 01/11 22:00: PR07 very faint at 20:00 and invisible at 04:00, a large bright blob under it at 22:00; whole image 02/11 22:00: a knot of lit threads between PR07, PR06 and PR08
- **Confidence**: high
### 2026-11-03T07:00Z → 2026-11-03T23:00Z (added)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: NORDEND-234 (pose fluctuating in mid-cycle), NORDEND-237
- **Family**: Thermal displacement; the corr_y dip of the afternoon follows the return of the pose on the same dates (labeling-rules §3, last item), so it is the same event
- **Measured**: housing 25.4 °C (07:00) → 27.1 (12–13) → 24.8 (15:00) → 22.4 (22:00); pan 0.0228 (07:00) → 0.0151 (11:00) → 0.0311 (23:00); tilt −0.131 → −0.123 → −0.143; scale 0.99896 → 0.99906 → 0.99887; jolts 12:30 (rotation −0.0099 → −0.00005 → −0.0088, tilt −0.124 → −0.133 → −0.126) and 13:30. corr_y 11:00 → 18:00: PR12 0.117 → 0.085, PR05 0.126 → 0.093, PR10 0.183 → 0.161, PR06 0.085 → 0.060, PR01 0.168 → 0.138, PR03 0.163 → 0.135, PR11 0.140 → 0.127, still low at 04/11 06:00, back at 04/11 14:00; PR09 corr_x −0.067 (04:00) → −0.042 (11:00), raised until 04/11 23:00 (ranges `--k 3`, up to +5 noises); PR18 corr_x 0.284 → 0.266 → 0.297
- **Images**: zone thumbnails 03/11 04:00 and 20:00: threads cross the group, one with a bright halo over PR18 at 20:00
- **Confidence**: high
### 2026-11-03T13:15Z → 2026-11-03T13:45Z (added)
- **Phenomenon of the catalogue**: A step of the camera that the compensation does not erase (here a jolt and back)
- **Precedents**: NORDEND-243 (same signature 05/11 12:00, Displacement, reviewed)
- **Family**: Displacement, as 243: a jolt of the pose without an established cause, leaving a remainder in the corr
- **Measured**: tilt −0.128 (13:15) → −0.137 (13:30) → −0.131 (13:45), rotation −0.0091 → −0.0036 → −0.0068; median raw y of all sensors and of the targets −0.21 px at 13:30 only; corr_y 13:15 → 13:30 → 13:45: PR01 0.171 → 0.142 → 0.161, PR12 0.114 → 0.095 → 0.105, PR16 0.005 → −0.011 → −0.009, PR15 0.037 → 0.022 → 0.027, PR02 0.179 → 0.165 → 0.167, PR05 0.116 → 0.104 → 0.117; all three references kept, CTR01 rejected as usual (12:15–15:45). At 12:30 the larger pose jolt leaves at most 2–3 noises (PR17 corr_y −0.013, PR13 −0.009): not labelled apart, written in the cycle label
- **Images**: hourly images only; 13:00 and 14:00 nothing in front
- **Confidence**: medium (amplitude smaller than 243)
### 2026-11-03T23:00Z → 2026-11-05T06:30Z (NORDEND-229, corrected)
- **Phenomenon of the catalogue**: The camera moves through a thermal effect
- **Precedents**: the label itself, NORDEND-234
- **Family**: Thermal displacement
- **Measured**: see the table above; corr_y of PR12 0.086 (03/11 23:00) → 0.113 (04/11 14:00); corr_x at 14:00: PR16 0.163 → 0.150, PR18 0.291 → 0.266, PR04 0.051 → 0.075; 05/11 05:00: PR18 0.313, PR16 0.180, PR15 0.165; corr_y PR17 −0.002 → −0.017, PR16 −0.018 → −0.028. CTR03 error_value 0.33 at 04/11 00:15, under its threshold, corr changes of 1–2 noises
- **Images**: zone thumbnail 04/11 02:00
- **Confidence**: high
### 2026-11-01T00:00Z → 2026-11-07T14:00Z (NORDEND-230, corrected)
- **Phenomenon of the catalogue**: A target right at the threshold, the whole batch
- **Precedents**: NORDEND-195 (spanning, same target)
- **Family**: Target issue, unchanged
- **Measured**: CTR01 rejections from the inventory (821 in the range); error_value 0.3–0.95 on 01–02/11 (multiples PNG)
- **Images**: as in the original label
- **Confidence**: high
## 2. What I left without a label
- 01/11–04/11: many error_value fluctuations of single sensors (`ranges.py --rolling 8 --instants 3`, `work/ranges_gap.txt`) without rejection and with corr changes of a few noises at most (PR08 corr_x/corr_y quantized steps of ±5–6 noises recurring every day; CTR02 error_value spikes in the evenings of 03 and 04/11, under its threshold).
- Cobweb threads cross the group every night from at least 30/10 22:00 to the end of the range (zone thumbnails); outside the PR07/PR08 night and the labelled CTR03/PR09/PR10/PR04/PR07 rejections of 05–07/11, they reject nothing and no corr step of their own was found. Written in the thermal labels where seen.
- 03/11 12:30 pose jolt: 2–3 noises on PR17, PR13 corr_y (in the 03/11 cycle label).
- 04/11 00:15: CTR03 error_value 0.33 (threshold 0.4) for one measurement, corr changes ≤ 2 noises.
- PR09 corr_x raised by up to 5 noises from 03/11 morning to 04/11 evening, with the thermal days (in the 03/11 cycle label, not a separate event).
## 3. The documentation
- Helped: `labeling-rules.md` §2 (thermal bounds on turning points, "without rest the cycle ends at the next start", "an event touching the batch edge extends the neighbouring label") and §3 (one Target issue per target and batch); `families.md` (Thermal or Displacement: comes back at equal temperature, which decided the warm plateau of 01–03/11). Precedents NORDEND-243 (jolt), NORDEND-232/236/242 (cobweb), NORDEND-104/117 (weak spot in daylight) applied directly.
- Missing: nothing says what to do with long labels cut at a batch edge when the next batch does not start right after (here a 3.6-day hole between two batches): the batch-edge rule assumes adjacent batches.
- The nights of 01–03/11 when the housing stays warm (about 27 °C instead of about 21) are not in the catalogue: "the camera does not come back at night because the housing does not cool" is a variant of "The camera moves through a thermal effect"; without weather data, its cause cannot be told.
- Spanning labels NORDEND-169 and NORDEND-195 end at the batch edge 01/11 00:00 while their condition continues; I could not change them, so the continuation is labelled separately (PR17) or by moving NORDEND-230.
## Difficulties
- No weather folder for NORDEND (`data/NORDEND/weather/` absent).
- The interval 01/11 00:00 → 04/11 21:15 had no label and no analysis to start from: it was labelled from scratch.
## Sources read
`tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`, `instructions/analyst.md`, `labeling/auto/NORDEND.jsonl` (labels of and around the range), `work/existing.jsonl`, `work/spanning.jsonl`, precedents (`precedents.py`, PR04 daylight rejection). Tools: `inventory.py` (window), `overview.py` (range with `--table no`; 31/10 06:00 → 05/11 08:00 with the table), `days.py --bounds` (30/10 → 07/11), `days.py` night levels (18/10 → 07/11; 20/08 → 07/11 for PR17, PR16, CTR01), `metrics.py` (30/10 → 05/11 12:00), `ranges.py` (31/10 18:00 → 05/11 `--rolling 8 --instants 3`; 25/10 05:00–12:00; PR09), `summary.py` (15 min and 1 h tables), `thumbnails.py` (PR07/PR08/PR06/PR13; PR04/PR05/PR17/PR16; PR17 fixed; zone 1600,700,1000,1000 at 8 times), whole images 01/11 22:00 and 02/11 22:00. Script: `scripts/build_labels.py`.