Files
aimsight-labeling/auto/analyses/NIEDERHORN_2026-10-21T1200_2026-11-01T0000/analysis.md
T

18 KiB
Raw Blame History

Labels: NIEDERHORN, 2026-10-21T12:00Z → 2026-11-01T00:00Z (UTC)

Request: verify the 42 existing labels of the range (work/existing.jsonl, all reviewed by the user on 2026-11-17), keep / correct / remove, and add what they miss. DETECTORS: no. Result: 39 kept, 3 corrected, 0 removed, 7 added (49 labels in labels.jsonl). The 12 spanning labels hold within the range.

Data: hourly cadence, 250 instants, 5 references (PWD_REF08_C09 to PWD_REF12_C09), 30 targets, camera pose at every instant, no image quality metrics (export columns peak to snr empty: quality.py not applicable), system metrics present. No weather: data/NIEDERHORN/weather/ does not exist (the export is named after a Toronto project; no Swiss position), so rain or fog are read from the images and the curves only.

Method: inventory.py on the range; overview.py (table no) then --rigid on 26/10; days.py --night 3-9 --bounds; metrics.py; summary.py on the moments; my scripts in scripts/: scan.py (per instant: rejections outside the Target issue regimes, corr deviations against a rolling median, labels covering), spikes.py (one-measurement spikes, both neighbours on the same side, at instants no label covers), steps.py (median step of the targets hour by hour, to test the sunrise starts), series.py (one sensor), build.py (writes labels.jsonl: the kept labels copied as they are minus id). validate.py: 0 errors (2 overlaps already present: NIEDERHORN-127/128, NIEDERHORN-132/134). After the changes, spikes.py at 4 noises finds no uncovered spike of a target.

Verdict per old label

Old id Verdict Evidence
NIEDERHORN-98 kept e06 rejected 13-14, e09 13, REF08-REF11 error above 5 MAD at 13:00; median step of the targets jumps at 13:00 (not at 12:00: REF09 already rises at 12:00, without effect on the targets)
NIEDERHORN-99 kept pan, tilt, rotation, scale > 5 noises 13:00-22:00 (days.py), temperature up to 17:00; curves back by 23:00
NIEDERHORN-100 kept 21T22: the 12 named targets rejected, B12_017-020 and e13 jump in corr_y
NIEDERHORN-101 kept same signature as 98; REF10 error still high at 14:00, end 15:00 right
NIEDERHORN-102 kept days.py: pose cycle 13:00-19:00, scale slow; curves back at 22:00
NIEDERHORN-103 kept e02 rejected alone at 22T16
NIEDERHORN-104 kept no measurement at 23T00; metrics present, voltage normal. Remark: the description does not say that the 00:00 image is also missing (images 22T23 → 23T01)
NIEDERHORN-105 corrected bounds 23T11→15 → 23T09→16: B07_016 also spikes at 10:00 (corr_x up, corr_y down, about 8 noises, a leaf touches the spot on the thumbnail) and its corr_x is still off at 15:00, back at 16:00
NIEDERHORN-106 kept e01, e02 rejected at 12:00, e06 13-14, e09 13; references up 13-15
NIEDERHORN-107 kept days.py pose cycle 13:00-21:00; back 22:00
NIEDERHORN-108 kept 23T19: rejections and under-threshold jumps as listed
NIEDERHORN-109 kept e06 13-14, e09 13, references above 5 MAD 13-14
NIEDERHORN-110 kept pose 13:00-21:00, scale slow during the night
NIEDERHORN-111 kept the six named targets rejected 17-19, the others hold
NIEDERHORN-112 kept rejections 25T04-16 (e06 04-16, B16_026 06-14, B07_017, e13, e15, then e10, e09, B12_020); first touched 04:00, last 16:00; e06's evening shift before 03:00 is its usual thermal return (same level every night)
NIEDERHORN-113 kept pose 13:00-23:00
NIEDERHORN-114 corrected bounds 26T03→05 → 26T01→05: B07_016 also spikes at 02:00 (corr_y up, about 7 noises, error up), one calm measurement before the 04:00 spike: same intermittent episode
NIEDERHORN-115 kept B07_016 rejected at 13:00 alone
NIEDERHORN-116 kept e06 rejected and REF10 error up at 14:00. Remark: with --rigid (base 12:00), at 14:00 references and targets move together in the image (common shift, scale up), references' residuals small: the pose jump is the camera warming, now labelled apart (new thermal label); 116's facts stay true and its "daybreak light" covers the e06 rejection and the reference error, as on the sunny mornings where Visibility and Thermal coexist
NIEDERHORN-117 kept no measurement, no metrics, no image at 26T15
NIEDERHORN-118 kept e02 rejected at 26T22 (e01 too, its daily regime, NIEDERHORN-10); B07_016 has a small corr_y rise at that measurement (about 4 noises), not named
NIEDERHORN-119 corrected bounds 27T07→10 → 27T06→10: B07_016 corr_y already up at 07:00 (4-5 noises, error almost double), leaves next to the spot on the 07:00 thumbnail
NIEDERHORN-120 kept 27T12 rejections as listed. Remark: B07_016's corr_y also rises at 12:00 (and a little at 11:00), not named
NIEDERHORN-121 kept e06 12-14, e09 12-13; REF09 error high until 15:00
NIEDERHORN-122 kept pose 13:00-21:00
NIEDERHORN-123 kept no measurement at 28T12; image and metrics present
NIEDERHORN-124 kept 28T14 rejections and under-threshold jumps as listed
NIEDERHORN-125 kept late cycle: pan up from 15:00, tilt drop 19-21, scale 16-22
NIEDERHORN-126 kept REF11 rejected at 28T19, e02, e06 up then back (step 28T19 → 20)
NIEDERHORN-127 kept rejections 29T01-15 (e06 01-13 and 15, B16_026, B07_017, e13, e15, e10, e09, B12_020, e02); the copied target list holds on this day
NIEDERHORN-128 kept references above 5 MAD only at 13:00
NIEDERHORN-129 kept pose 13:00-17:00, scale slow; remark: the pan keeps going down until 22:00 (below its night level), the text says "until about 18:00"
NIEDERHORN-130 kept 29T21 and 22 rejections as listed. Remark: under the threshold, B08_025 (strongly), B08_024, B15_019 at 21:00 and B15_020 at 22:00 jump without being named
NIEDERHORN-131 kept step at 30T02: e14, e15 up in x and y, e10 down, e09 down in y, e06 down in x then back over the morning; flat before (29T23-30T01 stable), offset until the end
NIEDERHORN-132 kept B16_026 rejected at every measurement from 30T02
NIEDERHORN-133 kept B07_016 rejected 04:00, spike 06:00
NIEDERHORN-134 kept e10 rejected 30T04-06, 08-11 … 31T21-01T00
NIEDERHORN-135 kept e09 rejected 13:00; remark: references still a little high at 14:00 (as on 29/10, where 128 also ends at 14:00)
NIEDERHORN-136 kept pose 13:00-21:00
NIEDERHORN-137 kept 31T11 rejections and jumps as listed
NIEDERHORN-138 kept REF11 rejected at 31T13, e06, e09 rejected, references up 13-14
NIEDERHORN-139 kept pose 14:00-23:00

Spanning labels (not in labels.jsonl): NIEDERHORN-2 to -8 (e03, e04, e05, e07, e08, e11, e12 rejected at every or almost every measurement of the range), -9 (B15_017: night rejections until 25T04, 27T04-12, last 28T00, then only trains), -10 (e01 rejected every day), -11 (B08_024 rare daytime rejections), -70 (B12_019 nights), -79 (B07_018 nights): all hold within the range.

1. My labels

Only the corrected and the added labels; the 39 kept labels are in the table above.

2026-10-22T20:00Z → 2026-10-23T01:00Z (added)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: NIEDERHORN-83, -87, -30, -49 (same target, same plant, reviewed): one label per burst, cause from the thumbnail, "according to the expert"
  • Family: Obstruction (one target, object in front), not Visibility (the neighbours, B07_017 the witness, do not move)
  • Measured: B07_016 corr_y 1.24 (20:00) → 1.56 (21:00) → 1.08 (22:00) → 1.58 (23:00) → 1.19 (01:00), own noise about 0.057; corr_x down 0.14 at 21:00; error 0.19, 0.17, 0.19 at 20, 21, 23 (usual 0.08-0.12). 00:00 missing for all (NIEDERHORN-104). 21:00 falls inside NIEDERHORN-102 (thermal): labelled apart
  • Images: thumbnails 20:00, 21:00, 23:00: blurred leaves sweep across and next to the spot
  • Confidence: high

2026-10-23T04:00Z → 2026-10-23T06:00Z (added)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction
  • Measured: B07_016 at 05:00 corr_x −0.135 (neighbours −0.51, −0.36), corr_y 0.90 (1.48, 1.10): spike of 5 noises in x; error barely up
  • Images: thumbnail 05:00: leaves just at the lower right of the spot, not on it: cause deduced
  • Confidence: medium (a single measurement, leaves next to the spot only)

2026-10-23T09:00Z → 2026-10-23T16:00Z (NIEDERHORN-105 corrected)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: NIEDERHORN-87 (a long intermittent episode on the same target, reviewed)
  • Family: Obstruction
  • Measured: B07_016 10:00 corr_x −0.065 / corr_y 0.78 (11:00 −0.42 / 1.36): 8 noises; 12-13 rejected (error 0.90, 0.96); 14:00 +0.14 / 0.98; 15:00 corr_x −0.24 (about −0.40 on the other days at that hour, −0.42 at 16:00). 11:00 single calm measurement: not a pause of several, same episode. 09:00 a little off in x (3.8 noises, under my threshold): not counted
  • Images: thumbnail 10:00: a leaf touches the spot; 09:00: leaf next to it
  • Confidence: high

2026-10-25T16:00Z → 2026-10-25T18:00Z (added)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction; inside NIEDERHORN-113 (thermal), labelled apart as the rules ask
  • Measured: B07_016 17:00 corr_x −0.18 (16:00 −0.45, 18:00 −0.40): about 5 noises against both neighbours; corr_y down a little; error 0.15 (0.04-0.06 around)
  • Images: thumbnail 17:00: a blurred leaf sweeps just under the spot
  • Confidence: high

2026-10-26T01:00Z → 2026-10-26T05:00Z (NIEDERHORN-114 corrected)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction
  • Measured: B07_016 02:00 corr_y 1.54 (01:00 1.05, 03:00 1.06): 7 noises, error 0.17; 04:00 corr_x +0.05 / corr_y 0.43: about 10 noises, error 0.17; 05:00 back
  • Images: thumbnails 02:00 and 03:00: leaves at the lower right of the spot; the 04:00 thumbnail of the original label: leaf tips on the spot
  • Confidence: high

2026-10-26T06:00Z → 2026-10-26T08:00Z (added)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction. Two calm measurements (05, 06) since the 04:00 spike: treated as a pause, a separate label
  • Measured: B07_016 07:00 corr_y 1.49 (1.09, 1.10): 6 noises; corr_x −0.55 (−0.40, −0.40); error 0.155
  • Images: thumbnail 07:00: leaves right next to the spot
  • Confidence: high

2026-10-26T12:00Z → 2026-10-26T16:00Z (added)

  • Phenomenon of the catalogue: The camera moves through a thermal effect
  • Precedents: the daily cycles of the range (NIEDERHORN-99 … -139); WILDSTRUBEL-22 for the form. Rule "thermal movement always labelled, even on an overcast day"; no label existed for 26/10
  • Family: Thermal displacement, not Visibility: with overview.py --rigid 2026-10-26T12:00Z, at 14:00 the overall movement is a common translation (ty about 0.8 px), a rotation and a scale rise, the references' residuals stay small (≤ 0.13 px), the median residual of the targets small (0.06 px): references and targets move together, the camera moves; the error of the references rises only moderately (REF10 0.16 → 0.37). The Visibility part (e06 rejected, REF10 error) stays in NIEDERHORN-116
  • Measured: housing temperature 26.0 (12:00) → 28.0 → 30.5 (14:00), 31 at 16:00, slowly down afterwards; tilt −0.0468 → −0.0432 → −0.0271 (14:00, +15 noises) → −0.0456 (16:00); rotation −0.0276 → −0.0353 (14:00, −7 noises) → −0.0260; scale 1.00005 → 1.00015 → 1.00016, slow return overnight (secondary axis); pan −0.0007 → +0.0021 (2-3 noises) → 0.0004 at 17:00. corr at 14:00: B08_024 x −10, B07_018 x −6, B08_025 x −5 noises (rolling), same direction as on the sunny days; median of the targets in corr_x −0.05 at 14:00, back at 16:00; corr_y hardly moves. 15:00 missing (NIEDERHORN-117): the cycle is visible on both sides, the gap does not cut it
  • Images: 14:00 whole image: the sun glows through thin clouds at the top of the field, with flare and reflections on the rails
  • Confidence: medium (one measurement of movement, the 15:00 point missing; but the signature is that of the other days)

2026-10-26T16:00Z → 2026-10-26T18:00Z (added)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction; separated from NIEDERHORN-115 by 14:00, the 15:00 gap and 16:00
  • Measured: B07_016 17:00 corr_y 1.44 (1.13, 1.12): 5.5 noises; corr_x −0.53 (−0.42, −0.43); error 0.127 (0.085-0.108)
  • Images: thumbnail 17:00: a blurred leaf sweeps next to the spot
  • Confidence: medium (moderate spike)

2026-10-27T06:00Z → 2026-10-27T10:00Z (NIEDERHORN-119 corrected)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction
  • Measured: B07_016 07:00 corr_y 1.41 (06:00 1.12): 4-5 noises, error 0.18 (0.10-0.12 around); 08:00 rejected (0.39); 09:00 corr_x −0.02 / corr_y 0.76; 10:00 back. 11:00 and 12:00 rise again in corr_y (1.50, 1.73; error 0.16, 0.27): 11:00 under my threshold, 12:00 is the instant of the train of NIEDERHORN-120: left there
  • Images: thumbnail 07:00: leaves next to the spot; 08:00 (original label): leaves on the spot
  • Confidence: high

2026-10-28T03:00Z → 2026-10-28T05:00Z (added)

  • Phenomenon of the catalogue: A plant shaken by the wind in front of a target
  • Precedents: as above
  • Family: Obstruction
  • Measured: B07_016 04:00 corr_y 1.39 (1.12, 1.11): about 5 noises; corr_x −0.47 (−0.35, −0.35); error 0.099 (0.077-0.079)
  • Images: thumbnails 03:00, 04:00, 05:00: leaves appear next to the spot at 04:00 only
  • Confidence: high

2. What I left without a label

  • The night drift of the camera: every night the pan rises slowly and the scale drops while the housing cools (strongest on 26-27/10, temperature at its lowest of the range at 27T11, pan up by about 0.01°, scale down to 0.9998); the median corr_x of the targets moves by about the noise of single targets (+5 median noises at 27T06-07 in days.py, against a reference night itself drifting). Not labelled, as in the reviewed labels of the range (the thermal labels close in the evening).
  • The pan below its night level after the cycle of 29/10 (down to −0.0097° at 22:00, back by 30T07): no corr effect; inside the slow night return.
  • B07_016 small rises under 4 noises: 23T09, 27T11, 26T22 (inside NIEDERHORN-118), 27T12 (inside NIEDERHORN-120): named in the verdict table.
  • PWD_REF09_C09 at 12:00 on sunny days: its error rises an hour before the others (first light); steps.py shows no step of the targets at 12:00 (median step 0.5-1.5 noises against 3 at 13:00): the sunrise labels keep their 12:00 starts. Its corr_x spike at 22T12 (4.5 noises) is a reference's corr: nothing to read into it.
  • B07_018, e06 corr_x shifts flagged at 23:00-02:00 by the rolling median: their daily thermal return (same night level each night), not events.
  • At the start of the range: B07_016 spikes at 21T11 (corr_x up, corr_y down), before the range: an event 21T10 → 21T12 belonging to the previous batch, not labelled here.

3. The documentation

  • Helped: labeling-rules.md §1 ("thermal movement always labelled, even on an overcast day"; "what falls under a long label is labelled apart"), §2 (start/end on the touched points), §3 ("an intermittent episode is cut at a pause of several measurements"); families.md (Thermal or Visibility boundary: references moving apart versus moving with the targets); tools/GUIDE.md (--rigid to tell the camera from a reference). Precedents NIEDERHORN-83/-87 fixed the form of the plant labels.
  • Missing: "a pause of several measurements" has no number; I cut at two calm measurements (by analogy with the No data rule) and joined across a single one. NIEDERHORN-87 (one label over a whole windy day) shows the other reading is also validated: the rule would deserve a number.
  • Contradicted or misleading: several reviewed labels carry copied descriptions (NIEDERHORN-98/101/106/109/121/128/135 identical, "image of 21.10 at 13:00" cited on other days; 112/127 identical lists): they hold on these days, but the sunrise ends differ for similar reference errors (128 and 135 end at 14:00, 98-109 at 15:00). NIEDERHORN-104 omits the image, which labeling-rules.md §5 asks of a No data. No weather for this station (non-Swiss): the instructions say to read it at the start; nothing to read.

Difficulties

  • No pandas in the environment: my scripts use tools/scripts/data.py (standard library).
  • precedents.py --exclude NIEDERHORN <start> <end> excludes only the range: the NIEDERHORN precedents of other dates come first (useful here).
  • The Write tool refused .md files in this run; analysis.md and report.md were written through Bash.

Sources read

agents/agents/analyst.md, instructions/analyst.md, work/prompt.txt, request.json, work/existing.jsonl, work/spanning.jsonl; tools/GUIDE.md, tools/DATA.md; labeling/families.md, labeling/labeling-rules.md, labeling/reading-rules.md, labeling/stations.md; labeling/auto/NIEDERHORN.jsonl (labels of the range and its edges); precedents (plant, overcast thermal); data/NIEDERHORN/ through inventory.py, overview.py (and --rigid on 26/10), days.py --bounds, metrics.py, summary.py, thumbnails.py (B07_016 and B07_017 on 22-23/10, 25-28/10), plot.py; whole image 2026-10-26T14:00Z; the other stations' exports only to count measurements at 23T00, 26T15, 28T12.