auto: the analysts verified the labels of 32 periods (11 stations, up to 14 days each): 710 labels -> 835, 587 unchanged; one analysis folder per period with its report (section Verification)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
maximeandClaude Opus 5.5 committed 2026-11-17 15:13:44 +01:00
1 parent df979301b5
commit 2a5d87659b
279 files changed
+10758 -710

No files matched your search

@@ -0,0 +1,43 @@
"""Every 3 h over the range: median of BJ_D corr_x - BJ_G corr_x, median error_value of BJ_C/BJ_D/BJ_G and of BL_D/PBL_D/PBJ_C;
then per day the correlation, measurement by measurement, of the spread with the BJ error_value."""
import sys, statistics as st
from collections import defaultdict
sys.path.insert(0, "tools/scripts")
import data
m = data.load("data/JUNGFRAU")
a, b = data.parse_time("2026-10-10T16:15Z"), data.parse_time("2026-10-21T10:30Z")
by = defaultdict(dict)
for s, ms in m.items():
for x in data.filter_window(ms, a, b):
by[x["t"].replace(second=0, microsecond=0)][s] = x
bins = defaultdict(lambda: defaultdict(list))
days = defaultdict(list)
for t, d in sorted(by.items()):
k = t.replace(hour=t.hour // 3 * 3, minute=0)
ev = st.mean(d[s]["error_value"] for s in ("BJ_C", "BJ_D", "BJ_G"))
if d["BJ_D"]["corr_x"] is not None and d["BJ_G"]["corr_x"] is not None:
sp = d["BJ_D"]["corr_x"] - d["BJ_G"]["corr_x"]
bins[k]["spread"].append(sp)
days[t.date()].append((ev, sp))
bins[k]["evBJ"].append(ev)
bins[k]["evO"].append(st.mean(d[s]["error_value"] for s in ("BL_D", "PBL_D", "PBJ_C")))
for k in sorted(bins):
v = bins[k]
print(data.iso(k), "spread %.2f" % st.median(v["spread"]) if v["spread"] else "spread -",
"evBJ %.3f" % st.median(v["evBJ"]), "evOther %.3f" % st.median(v["evO"]))
for d, p in sorted(days.items()):
if len(p) > 10:
print(d, "n", len(p), "r(evBJ, spread) %.2f" % st.correlation([x for x, _ in p], [y for _, y in p]))
# step to step: does the spread jump with the BJ error_value from one measurement to the next?
ts = sorted(by)
dev, dsp = [], []
for t0, t1 in zip(ts, ts[1:]):
if (t1 - t0).seconds != 900:
continue
p = [by[t][s] for t in (t0, t1) for s in ("BJ_D", "BJ_G")]
if any(x["corr_x"] is None for x in p):
continue
e = [st.mean(by[t][s]["error_value"] for s in ("BJ_C", "BJ_D", "BJ_G")) for t in (t0, t1)]
dev.append(e[1] - e[0])
dsp.append((p[2]["corr_x"] - p[3]["corr_x"]) - (p[0]["corr_x"] - p[1]["corr_x"]))
print("steps", len(dev), "r(d evBJ, d spread) %.2f" % st.correlation(dev, dsp), "slope %.1f" % st.linear_regression(dev, dsp).slope)
@@ -0,0 +1,13 @@
"""Per measurement between two instants: BJ_D - BJ_G corr_x, BJ_C corr_y, mean error_value of BJ_C/BJ_D/BJ_G, of BL_D/PBL_D/PBJ_C."""
import sys, statistics as st
sys.path.insert(0, "tools/scripts")
import data
m = data.load("data/JUNGFRAU")
a, b = data.parse_time(sys.argv[1]), data.parse_time(sys.argv[2])
w = {s: {x["t"]: x for x in data.filter_window(ms, a, b)} for s, ms in m.items()}
for t in sorted(w["BJ_D"]):
g = lambda s, c: w[s][t][c]
sp = g("BJ_D", "corr_x") - g("BJ_G", "corr_x") if g("BJ_G", "corr_x") is not None and g("BJ_D", "corr_x") is not None else float("nan")
print(data.iso(t), "spread %.2f" % sp, "BJ_C_y %.2f" % (g("BJ_C", "corr_y") or float("nan")),
"evBJ %.3f" % st.mean(g(s, "error_value") for s in ("BJ_C", "BJ_D", "BJ_G")),
"evO %.3f" % st.mean(g(s, "error_value") for s in ("BL_D", "PBL_D", "PBJ_C")))
@@ -0,0 +1,29 @@
"""Builds labels.jsonl from work/existing.jsonl: kept labels unchanged (review kept, id dropped), corrected ones rewritten (review dropped)."""
import json
D = "labeling/auto/analyses/JUNGFRAU_2026-10-10T1615_2026-10-21T1030/"
NEW = {
"JUNGFRAU-35": {"start": "2026-10-11T16:15Z", "end": "2026-10-12T07:30Z", "family": "Thermal displacement",
"description": "Cooling of the evening and the night: from late afternoon the camera tilt rises again while the corr_y of the bridge targets stays at its trough, then from nightfall the corr_y of the bridge targets rises until morning, the camera tilt too; references stable, very stable error_value. Deduced from the curves and the camera housing temperature, which drops from late afternoon: the bridge rises as it cools, and the camera fixed to it compensates it poorly. Normal, nothing to do."},
"JUNGFRAU-47": {"start": "2026-10-14T23:00Z", "end": "2026-10-15T08:00Z", "family": "Visibility",
"description": "Two references lost, the camera is blind: the error_value of REF_G then REF_D rises strongly, the reprojection error rises first; REF_G is rejected, then REF_D too (the camera is no longer registered at all for two hours), then REF_G comes back alone with REF_CT (reprojection at zero, two references) before REF_D comes back. All the targets jump in steps on corr_x and corr_y at each switch, then return. Once accepted again, REF_D is still seen poorly, in jolts, until early morning, and the targets (BJ_D, BJ_G, PBJ_C above all) stay a little shifted until it becomes sharp again. The cause of the poor visibility of the references is not established (very humid night at the nearby MeteoSwiss station, no rain). Every measurement of the range is invalid; if it repeats, check the state of the reference prisms."},
"JUNGFRAU-53": {"start": "2026-10-16T10:30Z", "end": "2026-10-16T16:45Z", "family": "Thermal displacement",
"description": "Warming of the day: when the measurements resume, the corr_y of the bridge targets sinks until late afternoon, the corr_x rises; the camera housing temperature rises clearly in the early afternoon. Deduced from the curves and the temperature. Normal, nothing to do."},
"JUNGFRAU-58": {"start": "2026-10-18T10:00Z", "end": "2026-10-19T10:15Z", "family": "Visibility",
"description": "Visibility drops over the whole scene during a long rain: the error_value of almost all the targets and of the references REF_G and REF_CT rises clearly; PM_C is rejected almost all the time and disappears from the curves, PM_D and PBL_C are at times; the corrections become agitated, with common jumps and a shift of BJ_G, BJ_D, PBJ_C that are not movements. Rain measured at the nearby MeteoSwiss station from late morning to the early hours of 19/10, air close to saturation; seen on the 14:00 image: the road is wet and the scene slightly veiled; on the 21:00 image the background is veiled in grey and the spots have halos, nothing in front of the targets. The error_values stay high until the morning cutoff, after which everything returns. Do not trust the curves over the range."},
"JUNGFRAU-59": {"start": "2026-10-19T08:45Z", "end": "2026-10-19T10:15Z", "family": "No data",
"description": "No measurement for any sensor for more than an hour: all the curves are missing; the station nevertheless takes its image at the first missing measurement and synchronizes within the hour, normal voltage around it. The same gap falls at the same moment on the other stations measuring at that time (CLARIDEN, FLORIDA, LENZSPITZE, NIEDERHORN, NORDEND, SNCF01): a common cause outside the station (transmission or reception of the data), not a failure of the station; it is not established. The curves are missing over the range; nothing to do on the station."},
"JUNGFRAU-61": {"start": "2026-10-19T16:15Z", "end": "2026-10-20T11:00Z", "family": "Visibility",
"description": "The three large spots nearest the camera, BJ_C, BJ_G and BJ_D, are seen less well from late afternoon and all night: their error_value rises clearly, and measurement by measurement, in step with it, the corr_x of BJ_G goes down and that of BJ_D rises (the two targets at the edges move apart in opposite directions) and the corr_y of all three goes down, then they come back together when it drops; the error_value of the other targets rises a little, the references stay calm, the camera pose and the housing temperature stay almost flat. The night after the rain is very humid at the nearby MeteoSwiss station; the 00:00 image is slightly veiled in grey, nothing in front of the spots. Cause deduced: a poorer view of these near spots in the humid air, not a thermal movement (the temperature does not change). Everything returns in the late morning of 20/10 as their error_value comes down. This is not a movement; do not trust the curves of BJ_C, BJ_G and BJ_D over the range."},
"JUNGFRAU-64": {"start": "2026-10-20T14:15Z", "end": "2026-10-21T09:15Z", "family": "Visibility",
"description": "Visibility drops over the whole scene, mostly at night, during the rain: the error_value of almost all the targets and of the three references rises, very clearly for REF_D (peak in the evening) and REF_G (in jolts all night), without any reference being discarded; PM_C is rejected all evening then at times in the early morning. Seen on the 20:00 image: the background is veiled in gray, like a mist lit by the infrared, and the spots of REF_G and PM_C are smaller than on a clear night. The raw position of REF_G moves away from that of the other references, the camera pose computed from them is distorted and the corrections of many targets move slightly (corr_x of BL_C and PM_D, corr_y of the BJ targets); the thermal cycle of the evening is added and makes the case hard to read. Everything returns in the morning of 21/10. Cause: the rain, measured at the nearby MeteoSwiss station from late afternoon of 20/10 to the morning of 21/10 with air close to saturation, consistent with the grey veil of the night images; the structure does not move, data of the range less reliable, nothing to do."},
}
out = []
for l in open(D + "work/existing.jsonl"):
d = json.loads(l)
i = d.pop("id")
out.append(NEW.get(i, d))
out.sort(key=lambda d: (d["start"], d["end"]))
with open(D + "labels.jsonl", "w") as f:
for d in out:
f.write(json.dumps(d, ensure_ascii=False) + "\n")
print(len(out), "labels,", sum("review" in d for d in out), "kept with review")