Files
aimsight-labeling/auto/analyses/NEBRASKA_2026-10-13T0000_2026-10-22T0000/analysis.md
T

9.8 KiB
Raw Blame History

Labels: NEBRASKA, 2026-10-13T00:00Z → 2026-10-22T00:00Z (UTC)

Verification of the 4 existing labels of the range (work/existing.jsonl), all reviewed by the user on 2026-11-17. Cadence: one measurement every 4 h (52 instants in the range, 3 missing). REFCOLH22 rejected at every measurement of the range: the compensation runs on REFH22 and REFP''21 only (their reprojection error stays at zero), as NEBRASKA-2 says. No image quality metrics on this camera (the export columns from peak to snr are empty). Weather: NEBRASKA has no weather/ folder, so there is no MeteoSwiss data. The station is indoors anyway (stations.md, and the images show a dark indoor scene).

Verdict per old id

  • NEBRASKA-3 (13-10 04:00 → 17-10 04:00, Displacement): kept. The corr_y of the top targets steps up between 04:00 and 08:00 on 13 Oct. COLJ20 goes from −0.051 to −0.002 to +0.032, POUJ20 from 0.047 to 0.084, POUJ19 from −0.007 to 0.032, POUH19 from 0.123 to 0.143. It stays high through 17-10 00:00 (COLJ20 0.041, POUJ20 0.082, COLH19 0.108, POUH19 0.145). At 17-10 04:00 every one of these targets is back at its 12 Oct level (COLJ20 −0.010, POUJ19 −0.025, POUJ20 0.053, COLJ19 0.034, COLH19 0.064, POUH19 0.120). So both bounds sit on the measurement before the first touched point and the measurement after the last. Pose: the tilt drops from its 12 Oct peak (29.66e-3 deg at 12-10 04:00) to 23.9e-3 at 15-10 20:00. The pan reaches its October minimum (−3.25e-3 deg at 15-10 16:00), then both go up again from 17 Oct. Housing temperature: 33.5 to 34 °C, stable. BREITHORN and MICHIGAN show the same pause in tilt (stations.py).
  • NEBRASKA-4 (15-10 20:00 → 16-10 12:00, No data): kept. Three measurements are missing (16-10 00:00, 04:00, 08:00), with no image between 20:00 and 12:00. NEBRASKA's metrics stop after 22:00 and come back at 11:00. The voltage is 14.14 V before the gap and 14.31 V after. stations.py over all the stations: from 23:00 there is no measurement on BREITHORN, CLARIDEN, FLORIDA, JUNGFRAU, MICHIGAN, NEBRASKA, NIEDERHORN, NORDEND and SNCF01 (nine stations). LENZSPITZE starts at 02:10, WILDSTRUBEL has no data, and the stations ALASKA, NADELHORN, SUSTENHORN and WYOMING have no data in this window. On JUNGFRAU, DB_not_uploaded climbs from 17 to 124 between 00:00 and 09:00, then falls to 0 at 11:00. Every claim in the description holds.
  • NEBRASKA-5 (16-10 20:00 → 17-10 20:00, Displacement): kept. COLJ19 corr_x goes 0.199 → 0.273 at 17-10 00:00 (largest jump of the range, 7.2 noises, ranges.py) and stays at 0.257–0.272 until 16:00. The raw x of COLJ19 moves +0.06 px while its neighbours move 0.00 to 0.01 px. Its rigid residual dx (overview.py --rigid, base 13-10 00:00) goes from 0.0015 to 0.070 and stays at 0.061–0.076 until 16:00. At 17-10 20:00 it is down to 0.033 in the residual and to 0.247 in corr_x: about half way back, so 20:00 is a defensible end. At 18-00 it is 0.020, then the long drift of NEBRASKA-6 takes over. The y of COLJ19 also falls on 17 Oct, but POUJ19, COLJ20 and POUJ20 fall with it (this is the end of NEBRASKA-3), so that part is not specific to COLJ19 and the description is right to mention only corr_x. COLJ19's error_value stays flat (0.071–0.074). The pose shows no step at 17-10 00:00 that the other targets do not absorb. Fixed thumbnails of COLJ19, with POUJ19 and COLJ20 as witnesses, at 16-10 16:00 and 20:00, 17-10 00:00, 12:00 and 20:00, and 18-10 00:00: the spot keeps the same shape and nothing is in front of it. The shift is a few hundredths of a pixel, too small to see.
  • NEBRASKA-7 (18-10 16:00 → 19-10 16:00, Thermal displacement): kept. The housing temperature has its October minimum so far on 19 Oct, 29.75 °C at 07:00 (the minimum before 19 Oct was 31.0 °C at 18-10 22:00). From 18-10 20:00, the corr_y rises: POUJ19 0.045 → 0.073 → 0.110 → 0.128 → 0.151 (08:00) → 0.130 → 0.077 (16:00), COLJ20 0.021 → 0.102 → 0.017, POUH20 to 0.069 at 08:00. At 19-10 08:00 the corr_x has its peak: COLH19 0.182, COLH20 0.142, COLH21 0.070, POUH19 0.266, COLJ20 0.573, POUH21 0.084. At the same moment the tilt peaks at 37.14e-3, the pan jumps to 2.08e-3 and the scale drops to its lowest (−42.1e-5). In days.py --bounds, 18 and 19 Oct are the only days of the range where the pose goes past 5 noises (tilt, rotation, scale). The references stay well seen: REFH22 error_value 0.155–0.168, REFP''21 0.028–0.040. The 08:00 peak returns at 12:00 to 16:00 with the temperature (32.5 °C at 12:00). The tilt does not come all the way back, because the long drift of NEBRASKA-6 keeps going underneath. The whole image at 19-10 08:00 shows the scene intact, with no obstacle.

1. My labels

The four labels are kept unchanged (labels.jsonl is identical to existing.jsonl, without the id field). The evidence for each is in the verdicts above. I add no label.

2026-10-13T04:00Z → 2026-10-17T04:00Z (NEBRASKA-3, kept)

  • Phenomenon of the catalogue: "The camera tilts little by little" (here, a pause and slight backward movement), with "A compensation on two references".
  • Precedents: NEBRASKA-1 and NEBRASKA-6 (the same camera before and after).
  • Family: Displacement. It does not come back at equal temperature and its cause is not established, so it is not Thermal.
  • Measured: see the verdict.
  • Images: 13-10 04:00 and 08:00. Same scene, nothing in front.
  • Confidence: high on the bounds; the cause remains not established.

2026-10-15T20:00Z → 2026-10-16T12:00Z (NEBRASKA-4, kept)

  • Phenomenon of the catalogue: "A gap common to several stations".
  • Precedents: none needed.
  • Family: No data.
  • Measured: see the verdict.
  • Images: none between 20:00 and 12:00 (the image gap is 960 min).
  • Confidence: high.

2026-10-16T20:00Z → 2026-10-17T20:00Z (NEBRASKA-5, kept)

  • Phenomenon of the catalogue: "A target or a small group jumps and stays shifted" (here a partial return).
  • Precedents: none useful (precedents.py returns FLORIDA and WYOMING reference jumps, which are not this case).
  • Family: Displacement, because no cause is seen.
  • Measured: see the verdict.
  • Images: fixed thumbnails, see the verdict.
  • Confidence: high on the jump. The end at 20:00 is within one measurement of the alternative (18-10 00:00).

2026-10-18T16:00Z → 2026-10-19T16:00Z (NEBRASKA-7, kept)

  • Phenomenon of the catalogue: "The camera drifts away when the temperature of its housing drops" (sheltered, in a building).
  • Precedents: NEBRASKA-8 to NEBRASKA-11 (the cold mornings at the end of the month).
  • Family: Thermal displacement. It follows the temperature trough and comes back with it.
  • Measured: see the verdict.
  • Images: 19-10 00:00 and 08:00, indoor scene intact.
  • Confidence: medium to high. The peak sits on top of the long drift.

2. What I left without a label

  • 20-10 04:00 and 08:00: a small dip of the top targets' corr_y (COLJ20 0.065 → 0.010 → 0.000 → 0.031) and a slight scale move (−39.7e-5). The temperature has a small trough here (31.0 °C at 06:00). days.py --bounds puts it under 5 noises on both the pose and the corr medians (scale −2.4, corr_y −2.9), so by the rule there is no thermal cycle to label. It sits under NEBRASKA-6. Note that the readings at 04:00 and 08:00 are identical on most sensors, and so is the pose.
  • 17-10 08:00: a small corr_x peak (COLJ20 0.466, POUH19 0.217) with a mild temperature dip (31.75 °C at 07:00). It is under NEBRASKA-5 and NEBRASKA-6, and below 5 noises.
  • 17-10 00:00: a one-measurement corr_y rise on COLH21, POUH21 and COLH20 (about 4 to 5 noises on COLH21 and COLH20). It falls inside NEBRASKA-3 and NEBRASKA-6, at the moment the camera turns from straightening to tilting again. POUH20 then stays up, which is the deformation NEBRASKA-6 describes. ranges.py --rolling 3 --instants 3 finds no instant in the range where three sensors deviate together.
  • 18-10: POUH20 corr_x steps up by a little (0.073 → 0.094) and stays. NEBRASKA-6 describes this explicitly (POUH20 rising fast with the tilt from 17 to 19 Oct).
  • error_value: small slow drifts (COLH19 0.097 → 0.126) that do not affect the corr.

3. The documentation

  • families.md and labeling-rules.md were enough. Two rules settled the "nothing to add" calls: no thermal label below 5 noises, and a bump that is only the trace of a slow camera movement follows the cutting of that movement.
  • reading-rules.md section 5 ("drift specific to a few targets that follows the camera… [NEBRASKA-6, NEBRASKA-5]") matches the data.
  • stations.md: the NEBRASKA row is accurate.
  • The rigid residual alone made COLJ19's y look specific on 17 Oct. Reading the raw y target by target showed the whole top group moving. On a scene that deforms, the similarity fitted on the median of the targets marks a group as a residual.
  • The spanning labels: NEBRASKA-2 holds within the range. NEBRASKA-6 holds within the range (the drift from 17 Oct; corr_y down on COLH19, POUH19, POUJ20, and a little on COLJ19 and COLH20; up on POUH20 and COLH21; tilt rising fast until 19 Oct). Its description cites "ensemble.py --rigide", which is not the name of the tool (it is overview.py --rigid). This is a wording slip, not an error on the data.

Difficulties

  • none

Sources read

tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md, labeling/auto/NEBRASKA.jsonl, work/existing.jsonl, work/spanning.jsonl, instructions/analyst.md. Tools: inventory.py, overview.py (--rigid 2026-10-13T00:00Z), days.py --bounds, ranges.py --rolling, metrics.py (NEBRASKA, JUNGFRAU), stations.py (the building, then all stations on the gap), thumbnails.py --fixed, plot.py, precedents.py. Whole images 2026-10-13T08:00Z and 2026-10-19T08:00Z. My script scripts/table.py (per-instant tables of corr, x, y, error_value with the pose).