Files
aimsight-labeling/auto/analyses/BREITHORN_2026-10-13T0000_2026-10-22T1600/analysis.md
T

13 KiB
Raw Blame History

Labels: BREITHORN, 2026-10-13T00:00Z → 2026-10-22T16:00Z (UTC)

Task: check the 5 existing labels of the range (work/existing.jsonl) and look for what they miss. 4 h cadence, 56 instants present, 3 missing (16/10 00:00, 04:00, 08:00). Recent-camera image-quality columns are empty: no quality.py. Weather: BREITHORN has no place in ~/aimsight-data/extraction/stations.txt (its line is bare), so no weather was fetched. The camera is indoors (factory) anyway.

Verdict per old id

  • BREITHORN-10: kept. The passage at 13/10 04:00 is confirmed: COLH20''' and COLI24 are rejected, POUI22 jumps in corr_x (+33 noises) and corr_y (−25), POUI24 jumps in corr_x (+28) and COLI22 in corr_y (−19), and everything is back at 08:00. On the 04:00 image a car with a light on sits at the bottom between the pillars, and the bottom spots are gone. The references hold (inlier 4/4). Bounds 00:00 → 08:00 are right.
  • BREITHORN-11: kept. Tilt goes from its high at 13/10 04:00–08:00 (0.0197) down to a minimum at 16/10 16:00 (0.01717). Rotation rises from −0.0253 (13/10 08:00) to −0.02197 (16/10 16:00). POUI19 corr_x goes from 0.276 down to 0.238 (15/10 12:00), COLH20 and POUI20 also come back. Resumption starts at 16/10 20:00 (tilt 0.0176, then 0.0188 at 17/10 00:00). Starting at 04:00 rather than 08:00 is within one cadence step and matches the end of the spanning BREITHORN-2.
  • BREITHORN-12: kept. No measurement at 16/10 00:00, 04:00 or 08:00, and no image (the image gap is 960 min, from 15/10 20:00 to 16/10 12:00). BREITHORN metrics, hourly: voltage normal (13.5–14.4 V) and last_sync present until 05:00, empty at 06:00, metrics missing 07:00–10:00, back at 11:00. stations.py on all stations: from 15/10 23:00, measurements are missing on every station measuring then, and they resume at different times. Everything the description says is borne out.
  • BREITHORN-14: kept. Housing temperature has its range minimum on 19/10 (25.75–26.75 between 05:00 and 08:00, versus 28–29 the day before) and is back to 28.5 at 12:00. Scale is at its minimum of 0.999752 at 08:00 and back to 0.999767 at 12:00, and rotation steps down. In corr_y, COLH20 goes 0.099 → 0.125 → 0.163 → 0.102 and REFPOU22 goes 0.027 → 0.052 → 0.086 → 0.025. COLI21 corr_x is 0.026 at 08:00 against 0.089/0.107 around it, and its corr_y is −0.017 against 0.032/0.057. The 08:00 image is clean. A small rise at 18/10 20:00 (COLH20 +4 noises, REFPOU22 +3) is too weak to move the start (see section 2).
  • BREITHORN-15: kept. COLI24 is rejected at 21/10 20:00, 22/10 00:00 and 04:00. COLI22 corr_x is around −0.10 (−11 to −15 noises) and POUI22 corr_y around +0.15 (+12). Both edges are clean: at 21/10 16:00 and 22/10 08:00 every sensor is at its level. On the 20:00 image a vehicle cab is at the bottom left between the pillars, and the 08:00 image is clear. Wording note: "deviate a little" understates the deviations of COLI22 and POUI22, which are clear. I left it, since the wording is qualitative and the label was reviewed.

1. My labels

2026-10-13T00:00Z → 2026-10-13T08:00Z (BREITHORN-10, kept)

  • Phenomenon of the catalogue: a pass-by hides the targets for the length of one measurement
  • Precedents: BREITHORN-9 (same signature, vehicle at the bottom of the scene)
  • Family: Obstruction, because the object is seen
  • Measured: as in the verdict above
  • Images: 13/10 00:00 (clean), 04:00 (car with a light on at the bottom)
  • Confidence: high

2026-10-13T04:00Z → 2026-10-16T20:00Z (BREITHORN-11, kept)

  • Phenomenon of the catalogue: the camera tilts little by little (the return phase)
  • Precedents: BREITHORN-2 and BREITHORN-13 (spanning)
  • Family: Displacement. The drift does not come back with temperature over the month (expert, BREITHORN-13).
  • Measured: pose and corr_x as in the verdict
  • Images: zone sheets from 13/10 to 16/10, nothing in front
  • Confidence: high

2026-10-13T20:00Z → 2026-10-14T12:00Z (added)

  • Phenomenon of the catalogue: none exactly. Closest is "The base changes without a reference lost" or a short go-and-return of the camera.
  • Precedents: NORDEND-112 (reviewed: a small joint jump with immediate return and nothing on the images, Displacement, cause not established). It guided the family and the conclusion.
  • Family: Displacement. No cause is seen, the housing temperature is flat, and the error_value does not rise.
  • Measured: corr_y of COLI20 goes −0.054 (13/10 20:00) → +0.0025 (00:00) → −0.001 (04:00) → −0.021 (08:00) → −0.037 (12:00); ranges.py gives the largest jump as +7.7 noises. COLJ19 goes −0.050 → −0.005 → +0.010 → −0.007 → −0.033 (+5.8 noises). POUI20 goes −0.045 → −0.019 → −0.015 → −0.033 → −0.058. POUI19 goes −0.125 → −0.081 → −0.072 → −0.096 → −0.088, so it comes back only in part. COLI20 corr_x is +0.018 at 00:00. With overview.py --rigid 2026-10-13T16:00Z --follow …, the residuals of these four targets stay within about 0.04 px, so they don't move relative to each other. The rigid run based on 12/10 20:00 shows the references moving against the targets at 00:00–04:00 (REFG44N dx +0.043/+0.030, REFI'21 dx −0.041/−0.030). Tilt is 0.01881 → 0.01969 → 0.01916, sigma0 peaks at 0.00099 at 00:00, and housing temperature stays 29.5–30 all night.
  • Neighbours (stations.py BREITHORN, MICHIGAN, NEBRASKA at 4 h): at 14/10 00:00 all three cameras take a small forward tilt step at the same measurement. BREITHORN goes 0.01881 → 0.01969, MICHIGAN 0.0100 → 0.0117, NEBRASKA 0.0263 → 0.0273, and all three come back partly at 04:00. MICHIGAN's targets also move then, in corr_x at 00:00 and in corr_y at 04:00, but that station runs on two references and swings that much on other nights too. No neighbour label covers this instant specifically (NEBRASKA-3 and MICHIGAN-16 are multi-day). This supports a cause common to the building, so the description says so.
  • Images: zone 13/10 20:00 and 14/10 00:00, the same scene with nothing in front
  • Confidence: medium. Clear on the curves, and the common tilt step makes a building-wide cause likely, but it is not established.

2026-10-14T04:00Z → 2026-10-14T12:00Z (added)

  • Phenomenon of the catalogue: a pass-by, with the reference base disturbed ("A passage in front of a reference that makes the base jump is Obstruction when the object is seen")
  • Precedents: JUNGFRAU-50 and JUNGFRAU-89 (reviewed: a lit background behind a reference gives small one-measurement jumps of several targets). They are Visibility there because the light was the sun. Here the light comes from objects, so it goes by the boundary in families.md ("even if the object is lit… Obstruction").
  • Family: Obstruction. A local object is seen on the image, and it is not a nuisance spread over the whole scene.
  • Measured: at 14/10 08:00 only. REFPOU22 corr_y is +0.014 against −0.005/−0.020 around it, REFG44N corr_y 0.0067 against 0.021, and REFI'21 corr_y 0.0089 against 0.021 (rolling ±3: −5 noises). COLH20 corr_x is 0.093 against 0.074/0.073, POUI20 corr_x 0.085 against 0.068/0.076, and COLI21 corr_x 0.052 against 0.083/0.075. error_value of REFG44N is 0.136 against 0.124/0.125 (+5 noises), and COLH20 is up as well. Nothing is rejected, and inlier is 4/4. Scale is 0.999824 against 0.999845/0.99986, and the housing temperature is flat (29–30).
  • Images: zone and sensor thumbnails at 04:00, 08:00 and 12:00 (fig_14oct_lit_objects.png). At 08:00 alone, lit structures appear behind REFI'21, REFH22, REFPOU22, POUI24, COLI22 and below COLI21, and the background behind COLH20 changes. All spots stay visible.
  • Confidence: medium. The deviations are small (3–5 noises), but they come together, on the references, with a pose dip and the object seen at that same measurement only.

2026-10-15T20:00Z → 2026-10-16T12:00Z (BREITHORN-12, kept)

  • Phenomenon of the catalogue: a gap common to several stations
  • Precedents: none needed. The label is reviewed and borne out.
  • Family: No data
  • Measured: see the verdict
  • Images: none during the gap; the files were searched
  • Confidence: high

2026-10-19T00:00Z → 2026-10-19T12:00Z (BREITHORN-14, kept)

  • Phenomenon of the catalogue: the camera moves through a thermal effect (cold trough of the housing)
  • Precedents: BREITHORN-17, 21, 24, 25, 27 (same signature)
  • Family: Thermal displacement (according to the expert)
  • Measured / Images: see the verdict; the 08:00 image is clean
  • Confidence: high

2026-10-21T16:00Z → 2026-10-22T08:00Z (BREITHORN-15, kept)

  • Phenomenon of the catalogue: a vehicle or machine that stays in front of the targets
  • Precedents: BREITHORN-16 (the next night)
  • Family: Obstruction
  • Measured / Images: see the verdict
  • Confidence: high

Spanning labels, inside the range

  • BREITHORN-2 (ends 13/10 04:00) and BREITHORN-13 (from 16/10 20:00): consistent within the range. Tilt rises from 16/10 20:00 to 0.0257 on 22/10. COLH20, COLI20, COLI21, COLJ19, POUI19 and POUI20 go up in corr_x, POUI19 the most (0.24 → 0.39), COLI24 goes down in corr_y, and POUH21 stays flat. Nothing to correct.

2. What I left without a label

  • 15/10 12:00 → 20:00, the warmest hours of the housing (31–32.25). Scale is at its range maximum, the rotation steps up, and COLH20 corr_x drops (a drift target, kept by BREITHORN-11). REFPOU22, REFG44N, POUH21 and POUI24 move in corr_x at 12:00 by only 3–4 noises (ranges --outside). The corr_y of REFPOU22 (down) and of COLI21, REFG44N, REFI'21 (up) mirror the cold signature but stay under about 4 noises, and they are back after the gap. That is under the 5-noise floor in labeling-rules.md for a thermal label. It is the closest candidate to a missed label.
  • 18/10 20:00: COLH20 corr_y +0.028 and REFPOU22 +0.024 (3–4 noises), as the housing starts to cool. It is the edge of the 19/10 episode, too weak to move a reviewed bound.
  • 19/10 20:00: COLI20, COLJ19, POUI20 and POUI19 dip together in corr_y for one measurement (−4 to −5 noises, rolling). The image is clean and the pose is quiet. It is the same mode of the upper group as on 14/10, but weaker and only one measurement.
  • 20/10 12:00: the same group goes up by 2–3 noises, and the error_value of COLI20, POUI20 and POUH21 rises a little. Too weak.
  • 22/10 04:00: pan and scale move a little at the housing trough (27.25), and POUI24 corr_x is +0.025 (~3 noises). This falls under BREITHORN-15, and nothing clear shows in the corr of the other targets.
  • 13/10 08:00 and 17/10 08:00: lit objects at the bottom of the scene (morning activity), with no deviation on the curves.

3. The documentation

  • What helped: families.md, on the Visibility/Obstruction boundary for a lit object and "A passage in front of a reference… is Obstruction when the object is seen". labeling-rules.md, on "what falls under a long label is labelled apart", on bounds at the gap edge, and on the 5-noise floor. reading-rules.md §5, on reading raw positions with --rigid. stations.md, for the building and the neighbours.
  • ranges.py --rolling 3 missed the two-measurement bump of 14/10 00:00–04:00, because the window contains it. It was found with --outside, by reading the targets' corr_y tables, and in the overview residuals. On a 4 h cadence, --rolling 1 would suit better.
  • The precedents: JUNGFRAU-50 and -89 are close in signature but differ in family. The difference is the light source (sun versus object), which families.md settles.
  • No contradiction with the reviewed labels. One nuance on BREITHORN-15: "a little" understates the deviations of COLI22 and POUI22.

Difficulties

  • No weather: BREITHORN has no place in stations.txt, and there is no data/BREITHORN/weather/ folder (checked after the user's question). The station is indoors.
  • The neighbouring cameras were first read only through their labels in the spanning context; the stations.py comparison on 13–14/10 was added after the user's question.
  • One overview.py --rigid --follow output was mangled by my cut of the columns; I read the per-sensor summary instead.

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; labeling/auto/BREITHORN.jsonl (ids, and descriptions of 9, 16, 17, 21, 24, 25); work/existing.jsonl; work/spanning.jsonl; precedents (3 queries: NORDEND-112, JUNGFRAU-50, JUNGFRAU-89, NEBRASKA-10, BREITHORN-17/24/25/27). Tools run: inventory, overview (with --rigid and residuals), metrics (4 h and 1 h), summary, ranges (--rolling, --outside), days, stations (all stations, around the gap), thumbnails (zone and sensors), plot. Images: 13/10 00:00 and 04:00; 19/10 08:00; 21/10 20:00; 22/10 08:00 (whole); zone sheets at 13/10 08:00–16:00, 13/10 20:00, 14/10 00:00–16:00, 15/10 08:00–16:00, 16/10 12:00–16:00, 17/10 08:00–12:00, 18/10 08:00–12:00, 19/10 12:00–20:00, 20/10 08:00–12:00, 21/10 12:00.