Files
aimsight-labeling/auto/analyses/SNCF01_2026-10-14T0645_2026-10-22T1030/scripts/build_labels.py
T
maximeandClaude Opus 5.5 ddb17c2189 auto: labels requalified under the 5-noise rules (only what stands out from the noise): 32 analyses refined, the 6 paused periods verified, the long labels requalified with the fixes the period analysts found; 1153 -> 961 labels
- 32 analyses: each label kept, rewritten or removed by its analyst (section Requalification of the report; mm magnitudes where the targets have them)
- SNCF01 x3, WILDSTRUBEL 2026-09-01 -> 2026-10-06 x3: verified from their existing labels under the new rules
- long labels (no analysis): auto/analyses/long-labels/<STATION>/; NORDEND-169 and -195 run to 07/11, so NORDEND-283 and -284 (their continuation) are removed
- reading-rules.md, stations.md: the citations of removed or rewritten labels point to their successor, or to [expert]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-17 23:58:00 +01:00

119 lines
7.4 KiB
Python

"""Build labels.jsonl from work/existing.jsonl: keep, correct, remove, add (verification of SNCF01, 14/10 -> 22/10)."""
import json
import pathlib
HERE = pathlib.Path(__file__).resolve().parent.parent
existing = [json.loads(l) for l in (HERE / "work" / "existing.jsonl").read_text().splitlines() if l.strip()]
REMOVE = {"SNCF01-55", "SNCF01-58", "SNCF01-62", "SNCF01-65"}
CORRECT = {
"SNCF01-38": {
"start": "2026-10-14T09:30Z",
"end": "2026-10-14T16:30Z",
"description": (
"By day (14/10), in the sun (hard shadows on the wall on the image of 13:00): the corrected curves of the low "
"targets fluctuate much more than at night, from one measurement to the next, with one-measurement jumps that "
"come back at once, mostly CV.4, CV.2, F4.3, CV.6, F4.2, BG.2, F4.5 and F4.6, and CV.2 and CV.4 are rejected "
"at times; the night before and the evening after are calm. Deduced cause: heat shimmer from the air heated "
"above the tracks (a fast noise over the scene, by sunny day only), not visible as such on the images; some "
"rejections of CV.2 and CV.4, low targets, may also come from passages, not seen. Do not trust the small "
"movements of these curves by day; this is not a movement of the structure."
),
},
"SNCF01-40": {
"description": (
"One single measurement: CV.2, CV.4 and CV.8 are rejected, their curves are missing, the recognition error of "
"the reference BG.3.1 jumps strongly while staying under its threshold, and CV.6, F4.2 and F4.6 make a corr "
"jump that comes back at the next measurement. Cause: a passage in front of the bottom of the scene and of "
"BG.3.1, deduced from the signature (several low targets and a reference hit together for one measurement; "
"no image at that minute, the station takes one per hour). The structure is not moving: point to ignore, "
"nothing to do."
),
},
"SNCF01-51": {
"description": (
"One single measurement: the CV, F4 and P targets are rejected and their curves are missing; the references "
"BG.3.1 and R1 are rejected too, the compensation retains none of them, and the targets that stayed under "
"their threshold, BG.2, BG.3 and F4.5, jump strongly in corr_x and corr_y for the time of the measurement, as "
"does the reference R2. Everything comes back at the next measurement. Cause: a train passing in front of the "
"targets and the references, seen blurred across the whole lower scene on the 16:00 image. The structure is "
"not moving: point to ignore, nothing to do."
),
},
"SNCF01-57": {
"start": "2026-10-18T07:45Z",
},
"SNCF01-60": {
"family": "Displacement",
"description": (
"On 19/10, an overcast day (images of 08:00 and 13:00, no sun on the wall), the camera moves in the morning "
"and does not come back: its pan and tilt drop and its scale rises, then stay at their new level in the "
"afternoon and at night, until the outage of 20/10. At the same time the targets move apart from each other "
"in the image (residual relative to the overall movement): corr_x of P5, P6 and P7 rises and that of P2 drops "
"clearly, and corr_y of P5 rises, through the afternoon and the night, while the residual of the compensation "
"rises from late morning and the recognition error of the reference R1 from the late afternoon. No return before the outage; "
"after it, on 21/10, P2 and P5 are still near their new levels. Not the daily thermal cycle: no sun, the pose "
"does not come back, no temperature is available. Cause not established (a camera or reference movement that "
"the compensation does not follow, or a movement of the scene). Movement of the structure not established: "
"to monitor; part of the month-long drift labelled separately."
),
},
"SNCF01-64": {
"description": (
"One single measurement: the CV, F4 and P targets are rejected and their curves are missing; the references "
"BG.3.1 and R1 are rejected too, the compensation retains none of them, and the targets that stayed under "
"their threshold, BG.2, BG.3, F4.5 and F4.6, jump strongly in corr_x and corr_y for the time of the "
"measurement, as does the reference R2. Everything comes back at the next measurement. Cause: a passage in "
"front of the targets and the references, deduced from the signature, the same as that of the train seen on "
"the 16:00 image of 16/10 (no image at that minute). It falls inside the episode of the reference R1 of that "
"morning, labelled separately. The structure is not moving: point to ignore, nothing to do."
),
},
"SNCF01-68": {
"description": (
"Daytime thermal cycle (21/10, sunny, between two measurement outages): in late morning the pan of the camera "
"drops and its scale rises, then the camera comes back in the afternoon and its tilt rises in the evening. The "
"compensation absorbs most of it, but not all: in the late afternoon and the evening, with the tilt rising, "
"corr_x of P5 and P7 dips clearly below its level of the morning, "
"then they come back at the isolated measurement after the gap. Deduced cause: the camera and its mount "
"move with the heat of the day (camera pose); the structure, a masonry wall, is not supposed to move like "
"that. No temperature is available to confirm it. Normal, nothing to do."
),
},
}
ADDED = [
{
"start": "2026-10-20T04:45Z",
"end": "2026-10-20T08:00Z",
"family": "Visibility",
"description": (
"At night on 20/10, before dawn, the reference R1 is less well seen: its recognition error rises a notch at once and "
"stays high until mid-morning, and its measured position slides alone in the image (the targets do not move "
"relative to the overall movement). The compensation follows it: corr_y of P7, P6 and P5 rises together "
"clearly for a few hours, while the reference R2 moves the other way, then everything comes back as the "
"position of R1 returns, around eight. The train passage of that morning, labelled separately, falls inside. "
"Nothing in front of R1 on the night images of 05:00, 06:00 and 07:00 (its spot whole, same place); the "
"night is one of poor visibility for several faint targets (CV.6 rejected all night, F4.2 at times). Cause "
"deduced: the reference less well seen in the night of poor visibility, not an obstacle. Do not trust the "
"small shift of these curves over the range; this is not a movement of the structure."
),
},
]
out = []
for lab in existing:
lid = lab["id"]
if lid in REMOVE:
continue
new = {k: v for k, v in lab.items() if k != "id"}
if lid in CORRECT:
new.update(CORRECT[lid])
new.pop("review", None)
out.append(new)
out.extend(ADDED)
out.sort(key=lambda d: (d["start"], d["end"]))
(HERE / "labels.jsonl").write_text("".join(json.dumps(d, ensure_ascii=False) + "\n" for d in out))
print(len(out), "labels;", len(REMOVE), "removed;", len(CORRECT), "corrected;", len(ADDED), "added")