Files
aimsight-labeling/auto/analyses/ALASKA_2026-11-04T2358_2026-11-16T1217/analysis.md
T
2026-11-18 10:51:31 +01:00

19 KiB
Raw Blame History

Labels: ALASKA, 2026-11-04T23:58Z → 2026-11-16T12:17Z (UTC)

Station context (not in stations.md, read from the data and the images): a camera under a bridge over a river (Geneva; MeteoSwiss Genève / Cointrin, 4.9 km, 411 m), looking along a row of piers carrying 12 targets (01 at 20 m to 11 at 231 m, 02 and 02_g on the same pier at 41 m), the far ones crowded at the end of the row (08–11 about 15 px apart). Two references only, ref_av_01 and ref_av_02, on the far bank, to the right of the row and about 350 px apart: their corr is 0 by construction, reproj_err 0 from 2026-11-05 ~08:00 (tiny non-zero values before, no effect). Old camera: no image-quality columns (quality.py not applicable). Cadence 15 min, one IRON image per hour. Hints: none. No existing label in labeling/auto/ALASKA.jsonl (the file does not exist). The neighbouring camera WYOMING looks at the same piers (its reviewed labels coincide at several instants, below).

Method of the threshold. The own noise of every target (median |step|) is dominated by a common jitter: at each measurement all targets move together by a few noises (near targets more px, far targets more mm), because the two-reference compensation carries any small shift between the two references into all targets (overview.py --rigid: the references move against the rigid fit of the targets, the targets' residuals stay small). Literally, at least one target exceeds 5 noises (rolling ±8 measurements) at 60 instants of corr_y and 38 of corr_x, mostly one or two near targets in the tail of this jitter. To keep labels significant and the treatment the same everywhere, I labelled (scripts scripts/common_dist.py, scripts/touched.py):

  • every moment where the median of the targets exceeds 5 noises (19 instants corr_y, 6 corr_x; the 99th percentile of the median is 5.7 noises), grouped into agitation episodes (same night, same mechanism; bounds on any target beyond 5 noises next to them);
  • a target or group that clearly leaves the others (10 on 05/11, 04 on the cold spells), slow excursions beyond 5 noises lasting hours (13–15/11), the rain episode, and the No data. What stays under (median below 5, a few targets at 5–7 noises) is listed in section 2.

1. My labels

2026-11-05T02:45Z → 2026-11-05T05:00Z

  • Phenomenon of the catalogue: "A compensation on two references" (family of the cause; cause not established → Displacement, "The base changes without a reference lost").
  • Precedents: MICHIGAN-8 (reviewed: agitation of all targets from the two-reference compensation, Displacement); FLORIDA-72 (reviewed: one-measurement jump, reference shifted alone, Displacement); WYOMING-9/-13/-17 (reviewed, neighbouring camera: far references jump alone, Displacement). Found by keyword search (see Difficulties).
  • Family: Displacement: no cause seen (night image clean, reference error_value far from threshold, no rejection); not Visibility (no veil, error_value not up on most sensors).
  • Measured: 03:00 corr_y median of targets +6.1 noises (9 targets 5–6.4 noises: 01–08, 02_g), 04:45 +5.0 (02, 03, 04, 05, 08, 09); back at the next measurement. Rigid residual of ref_av_02 dy +0.14–0.15 px at 03:00–03:15, targets' median residual 0.02–0.03 px. ref_av_01 error_value 0.048 at 03:00 (level 0.024). Pose without step. In mm: a few tenths of a mm on near targets to about 2 mm on far ones.
  • Images: zone 1700,950,850,400 at 03:00: night, all spots sharp, nothing in front.
  • Confidence: medium: the mechanism is clear, the threshold barely exceeded.

2026-11-05T08:45Z → 2026-11-05T09:15Z

  • Phenomenon of the catalogue: "A pass-by hides the targets for the length of one measurement" (signature) vs "jump then return without cause"; kept Displacement.
  • Precedents: WYOMING-10, WYOMING-18 (reviewed: one-measurement disturbance of two neighbouring targets, Obstruction deduced, no image at that instant). Here an image exists at the same minute and shows nothing, so the passage is not deduced.
  • Family: Displacement (jump then return, cause not established); Obstruction would need the object seen or a deduction an existing image contradicts.
  • Measured: corr_y of 10 +27 noises (+0.68 px, about 14 mm), 11 +16, 09 +12; corr_x of the far targets +3–4 noises; error_value of 10 0.075 (level 0.025), 08, 09, 11 up too; references at their level. Rigid residual dy of 10 +0.47 px, 11 +0.18, 09 +0.09: these targets move alone in the image. Back at 09:15.
  • Images: thumbnails --fixed of 08–11, 06, the references at 08:00, 09:00, 10:00: spots intact, nothing in front at 09:00 (dawn lighting changing).
  • Confidence: high on the event and bounds; cause unknown.

2026-11-05T23:45Z → 2026-11-06T07:30Z

  • Phenomenon of the catalogue: "A compensation on two references".
  • Precedents: MICHIGAN-8 (agitation over several measurements as one label), FLORIDA-72.
  • Family: Displacement, as above.
  • Measured: corr_x median +6.9 noises at 00:00 (06, 07, 08, 11 at +8), +5.4 at 00:45; corr_y +4.4 at 02:00 (4 near targets +5), −5.3 at 03:45 (01–05 at −6), −5.2 at 07:15 (05–11, 11 at −6.5). Rigid: ref_av_01 dx −0.15 px at 00:00, ref_av_02 dx +0.12 at 00:45, ref_av_01 dy +0.11 at 07:15; references' error_value at level (0.021–0.030). Humid calm night (93–94 %).
  • Images: zone at 00:00: clean night.
  • Confidence: medium (agitation in 4 separate jumps; I grouped it as one night episode).

2026-11-05T23:45Z → 2026-11-08T10:00Z

  • Phenomenon of the catalogue: "The structure expands and contracts with the temperature" (each target its amplitude, references stable), here a single target.
  • Precedents: none useful for a single target following the temperature over days (WYOMING-1…8 are thermal labels of the camera and of a reference, another mechanism).
  • Family: Thermal displacement: 04 follows the housing temperature as a trend and comes back at equal temperature (cold: 05–08/11 and 13–15/11 high; mild 08–12/11 low). Not Displacement "targets move slowly alone then come back" because the temperature explains it in both directions on both cold spells; medium confidence for that reason.
  • Measured: corr_x of 04 vs its window level: ranges ≥ 5 noises on 06/11 00:00–00:45 (common), 02:30–03:00, 05:45–07:00, 07/11 03:30–06:15 (+8.0), 07:30–07:45, 20:45 → 08/11 03:30 (+11.3 noises at 23:45, +0.44 px ≈ 3.5 mm), 04:45–09:45 (+9.1). Night levels (days.py): 07/11 +0.195 px against −0.06…−0.12 on 08–12/11. Rigid residual dx of 04 (base 05/11 00:00): ≈0 on cold nights, −0.12…−0.19 by day, stuck at −0.2 px from 08/11 12:00 to 12/11; 03, 05 within ±0.05. Housing temperature 20.7–25 °C on the cold nights, 25–27 °C on 09–12/11. error_value of 04 at level (~0.11). The 01 night level also rises on 07/11 (+4 noises), not beyond.
  • Images: thumbnails --fixed of 04 and 05 at 07/11 23:00, 08/11 05:00, 10/11 23:00: spots sharp, nothing in front, no shape change.
  • Confidence: medium: clear in the data, the thermal link is a trend on two cold spells (r not computed per target), cause (support of the prism, pier) not seen.

2026-11-07T04:45Z → 2026-11-07T10:45Z

  • Phenomenon of the catalogue: "A compensation on two references".
  • Precedents: WYOMING-9 (reviewed, 07/11 04:45→05:15: far references jump alone, Displacement) at the same instant; MICHIGAN-8.
  • Family: Displacement.
  • Measured: corr_y median +7.2 noises at 05:00 (all 12 targets, 07 and 10 +9), +5.7 at 06:30 (11 targets), −4.4 at 09:30 (04–07), corr_x +5.8 and corr_y +6.8 at 10:00 (22 sensor-columns beyond 5), 10 +6 at 10:30. Rigid: ref_av_01 dy −0.11 at 05:00, ref_av_02 dy +0.14 and ref_av_01 dx −0.08 at 10:00. References' error_value 0.024–0.030 (level).
  • Images: zone at 05:00 (clean night) and 10:00 (sunny morning, spots sharp, no sun on the spots).
  • Confidence: medium.

2026-11-07T22:45Z → 2026-11-08T09:00Z

  • Phenomenon of the catalogue: "A compensation on two references".
  • Precedents: WYOMING-11 (reviewed, 08/11 06:45→08:30, dawn after the coldest clear night, far prisms shift together, Displacement); MICHIGAN-8.
  • Family: Displacement; considered Visibility (ref_av_01 error_value up all night ×2–3, near targets' error_value rising at dawn) but the boundary asks error_value rising on most sensors or a veil seen: neither (5–6 of 14 sensors, far from thresholds, images clean).
  • Measured: 23:00 11 corr_x −5; 23:15 corr_y median −6.5 (10 −10.7, 11 −10); 23:45 corr_x +4.7 (5 targets +6); 03:15 corr_y −5.2 (07 −7.9); 04:30 corr_x −4.9; 05:30 corr_x +5.2; 06:45 mixed; 08:30 corr_y median +8.8 (06 +9.7, 02 +9.5) with corr_x −6.5; 08:45 02, 03, 04, 02_g +5–6; 09:00 back. Rigid: ref_av_01 dy +0.21 at 23:15, ref_av_02 dx/dy +0.16/+0.27 at 08:30 (pan at its maximum of the range at 08:30: the pose follows the references). Air 0–3 °C, humidity 95 % (Genève-Cointrin), housing 20.7 °C minimum at ~06:00.
  • Images: zone at 23:00, 03:00, 07:00, 08:00: no veil, no obstacle, spots sharp; whole scene at dawn normal.
  • Confidence: medium.

2026-11-10T08:15Z → 2026-11-10T09:00Z

  • Phenomenon of the catalogue: "A gap common to several stations".
  • Precedents: WYOMING-12 (reviewed, same bounds, same wording). It says "a synchronization falls in the gap" for WYOMING; for ALASKA last_sync shows syncs at 08:00, 08:15 then 09:00, none inside: I wrote "synchronizations resume at the next measurement".
  • Family: No data (no measurement of any sensor).
  • Measured: 08:30 and 08:45 absent for all 14 sensors; WYOMING the same two; CLARIDEN misses 08:30, LENZSPITZE 08:25–08:35, NADELHORN 08:25–08:30 and 08:40–08:45; FLORIDA, JUNGFRAU, NORDEND complete. BusVoltage 14.09 V (08:00), 14.01 V (09:00); link up.
  • Images: 08:00 and 09:00 exist; none expected at 08:30 (hourly).
  • Confidence: high.

2026-11-12T12:00Z → 2026-11-13T00:00Z

  • Phenomenon of the catalogue: "Rain" plus "A compensation on two references" (the references poorly seen distort the base).
  • Precedents: FLORIDA-65 (reviewed: night of rain, two references, one label for the episode); JUNGFRAU-116 (reviewed: references poorly seen in rain, targets slip, Visibility).
  • Family: Visibility: veil seen on the image, error_value up on references and most targets.
  • Measured: error_value of ref_av_01 up to 0.59, ref_av_02 0.44 (levels 0.024, 0.023) from 11:15 to ~19:30 and again 21:30–23:45, 10–13 sensors above their usual band; 02 rejected at 12:00. corr_y: 12:15 a few targets ±5–6; 12:30 07–11 +5…+8; 12:45 median −8.0; 13:00 median −15.7 (02 −19.9 noises ≈ −5.6 mm, 04 −20.2, 11 −14.3 ≈ −7.7 mm); 13:15 back. 15:45–17:00 04 corr_y −11.6 (rigid dy −0.38 px: its own spot), 16:30–17:00 01, 02, 02_g, 03, 06 −6…−8; 19:45 09 −5.6; 23:15 02, 04 −5.7; 23:45 04 corr_x −5.1. Rigid at 13:00: ref_av_02 dy −0.44 px. Rain 0.5–1.5 mm/h all day at Genève-Cointrin.
  • Images: zone at 13:00: veiled grainy background, haloed spots (against 11/11 03:00 and 09/11 08:00 clean).
  • Confidence: high on cause and core; the label spans the whole rainy afternoon and evening (one continuous nuisance, three bursts of exceedances).

2026-11-13T19:15Z → 2026-11-15T10:00Z

  • Phenomenon of the catalogue: "The base changes without a reference lost" (slow part) and "A compensation on two references" (jumps).
  • Precedents: WYOMING-17 (reviewed, 14/11 20:45→21:15, the same jump at 21:00); NEBRASKA-6 (reviewed, long common drift from the base, Displacement); MICHIGAN-8.
  • Family: Displacement: targets do not move in the image; the base (ref_av_02 against the targets) changes; no established cause. Not Thermal: on the cold morning of 08/11 the common corr_y was low, not high, so it does not come back at equal temperature. Not Visibility: references' error_value at level or barely up, images without veil on the spots.
  • Measured: slow common corr_y rise (hourly medians +1…+2 mm on 03–06, 04 most), ranges ≥ 5 noises: 04 13/11 19:30–19:45 and 23:15 → 14/11 09:30 (25 measurements, +8.1), 02/05/06 14/11 02:00–05:30 (+6.5…+7.6), single instants of 6–9 targets at 02:00, 03:15, 07:00, 09:30, 10:45; 04 19:15–22:15 (+11.5). Jumps: 17:30 corr_x median +6.0 (9 targets); 21:00 11 corr_x −14 noises (≈ −4 mm), 08–11 −6…−8, corr_y far targets −8…−10, both references' rigid residuals 0.24–0.33 px; 22:15 corr_y median +9.8 (02 +12); 23:00, 00:30, 01:15 (+6.7), 02:45 (−5.6), 04:00 (+7.9), 04:15, 06:00–06:30 (+5…+6); 15/11 06:45 02 +6, 09:30–09:45 04 corr_y +5.5; afterwards nothing beyond. Rigid: ref_av_02 dy positive (+0.1…+0.38) through 14–15/11. Housing 20.2–22 °C; rain from 22:00 on 14/11 to 10:00 on 15/11 at Genève-Cointrin.
  • Images: zone at 14/11 17:00, 21:00, 22:00, 15/11 04:00 (spots sharp, ref_av_02 a little smaller at 04:00); near-lens zone 0,0,1600,1100 with contrast at 21:00, 23:00, 01:00–07:00: background brighter from 23:00 (rain backscatter likely), lit objects close to the lens at 03:00 (blob) and 04:00 (streaks), far from the references.
  • Confidence: medium: one label for a two-day episode mixing a slow base shift and many jumps; split by a reviewer is possible (WYOMING labelled 21:00 alone).

2026-11-14T03:00Z → 2026-11-15T09:30Z

  • Phenomenon of the catalogue: as the 05–08/11 label of 04.
  • Precedents: none useful.
  • Family: Thermal displacement, same reasoning, same target, same family (labeling-rules 4).
  • Measured: 04 corr_x ranges ≥ 5 noises: 14/11 03:15, 05:15, 09:45–10:30, 17:30, 21:15–21:45, 23:00–23:45, 15/11 01:00–02:00 (+9.2), 03:30, 04:45–09:15 (+8.9 at 08:00). Rigid dx of 04 ≈ 0…+0.06 on 13/11 06:00 → 15/11 06:00, then −0.14 at 12:00, −0.2 at 18:00 (warm position again); housing 20.2–22 °C then 29 °C on 15/11 ~13:00.
  • Images: thumbnails --fixed 04/05 at 14/11 05:00, 15/11 08:00, 15/11 14:00: spot sharp.
  • Confidence: medium.

2. What I left without a label

  • Common jitter under the median threshold (all targets together, median of targets below 5 noises, a few targets 5–7): 09/11 06:45 (05, 09, 11 +5…+6); 10/11 04:30 (rain peak, 02 and 03 rejected, 05, 06, 08, 09 corr_x +5…+7, median +4.9: borderline); 10/11 06:30 and 13:00 (two targets −5…−7 corr_x); 11/11 14:30–17:30 (near targets 01, 02, 02_g, 03 ±5…+7 corr_y repeatedly, median up to −4.8 at 16:00: borderline, same mechanism); 14/11 12:00, 12:45 (2–3 targets +5…+6 corr_x); 15/11 21:00 (01, 02, 03 −6 in the evening rain and wind); 16/11 08:30 (05, 07, 09, 10 +5…+6). About 35 other instants where one target, mostly 01 or 02_g (the noisiest), touches 5–6 noises alone.
  • Rain night of 10/11 01:00–05:00: error_value up on 10–14 sensors, 02 and 03 rejected at 04:30 (missing points, not labelled), corr at most the borderline 04:30 above.
  • Rain night of 15–16/11 (19:00 → 02:00, strong wind): error_value up on 9–11 sensors around 20:15; corr only 15/11 21:00 above.
  • Camera thermal cycle: pan, scale follow the housing temperature (r −0.67 and +0.82 over the range), x raw of all sensors swings by up to 1.4 px on warm afternoons (05, 07, 08, 10, 11, 14, 15/11); compensated (corr flat at those hours), not labelled.
  • Slow tilt rise 0.03 → 0.055° over the range: compensated.
  • Near targets' error_value ramps on cold nights (01, 02, 03 on 07–08/11 and 13–14/11; 02 rejected 08/11 10:15, 14/11 21:15; 03 rejected 13/11 15:15): without effect beyond the noise apart from what is labelled.
  • Pose missing 16/11 00:30 → 12:15 (48 instants without registration in camera/) while the corr are present and calm: probably the camera file of 16/11 cut by the pumping; no effect on the curves.
  • reproj_err tiny non-zero until 05/11 ~08:00 then exactly 0: a change of computation, no effect.

3. The documentation

  • Helpful: families.md "A compensation on two references" and "The base changes without a reference lost"; reading-rules.md 5 (rigid reading first, neighbouring cameras read together); overview.py --rigid was decisive (references moving against the targets, target 04 moving alone, 10 moving alone). labeling-rules.md 4 (same target, same family across batches) for the two 04 labels. The weather (Genève-Cointrin) separated the rain days.
  • Precedents: WYOMING's reviewed labels (neighbouring camera, same piers) fall at the same instants as three of my moments (05:00 on 07/11, dawn of 08/11, 21:00 on 14/11) and support Displacement for reference-driven jumps; WYOMING-12 for the common gap (I could not repeat its "synchronization in the gap" for ALASKA). FLORIDA-65 and JUNGFRAU-116 for rain on references.
  • Missing / unclear: the 5-noise rule with the station's own noise does not fit a two-reference station whose noise is a common jitter with heavy tails: literally, 60 instants (corr_y) and 38 (corr_x) would be labelled, most of them one near target in the tail. I applied the threshold to the median of the targets (stated above); a rule on how to treat a common-mode jitter would help. The rules also do not say whether a night of intermittent jumps of the same mechanism is one agitation or one label per jump (FLORIDA labels each jump, MICHIGAN-8 groups three): I grouped per night.
  • stations.md has no row for ALASKA (bridge over a river, two references on the far bank, neighbour WYOMING); worth adding.
  • The housing temperature is temperature_deg_c only (all sensor_*_deg_c empty), which metrics.py uses as temperature.

Difficulties

  • precedents.py could not run: the Ollama container ollama-bge-m3 is up but has no published port nor IP (docker inspect: Ports {}, no address), so http://10.0.2.140:11435 refuses the connection. I did not restart it (shared service). Fallback: keyword search (grep) on labeling/auto/*.jsonl ("two references", WYOMING labels, the 10/11 gap). The precedents cited come from that search, not from the RAG.

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 (in full).
  • labeling/auto/ALASKA.jsonl: absent. Keyword search in labeling/auto/*.jsonl (labels of MICHIGAN, NEBRASKA, BREITHORN, FLORIDA, JUNGFRAU, NORDEND, LENZSPITZE, WYOMING).
  • Data: data/ALASKA/ (displacement, camera, metrics, weather, targets.csv, images), and the displacement exports of 10/11 of the other stations (minutes of the gap).
  • Tools: inventory.py, overview.py (default, --rigid, --follow), ranges.py (k 5, --rolling 8, --instants 3, --min-measurements), summary.py, plot.py, days.py, metrics.py, stations.py (ALASKA, WYOMING), thumbnails.py (sensors, --fixed, zones, --contrast), whole images of 05/11 09:00 and 15/11 04:00.
  • Own scripts: scripts/weather.py, scripts/devtable.py, scripts/common_dist.py, scripts/touched.py, scripts/slowfig.py; outputs in work/.