16 KiB
16 KiB
Automatic labels: NORDEND, 2026-11-04T21:12Z → 2026-11-07T14:12Z (UTC)
Analysis requested from the Data page (hint: "the detectors do not see the disturbances in corr-x and corr-y"). No label of labeling/auto/NORDEND.jsonl touches the range: the last ones end on 01/11 at 00:00, and the batch from 01 to 04/11 is not labeled. Nothing is therefore extended, nor contradicted. Working figures in work/, those of the report at the root of the folder.
1. My labels
2026-11-04T21:15Z → 2026-11-05T06:30Z
- Map phenomenon: The camera moves through a thermal effect
- Family: Thermal displacement. The pose follows the housing temperature and comes back with it; it is always labeled.
- Measured: night of 04 to 05/11, housing temperature from 22.8 to 19.2 °C; pan rising (0.028 → 0.039°), tilt dropping (−0.141 → −0.152°), scale dropping, without rest until the turning point of 05/11 at 06:30 (pan 0.0389, then 0.0379 at 06:45, tilt −0.1508 then −0.1493).
corr_xof the right-edge targets a little above the day (scripts/cycle_jour.py). The cycle started on 04/11, before the batch, with no neighboring label: it starts at the batch edge. - Images: none useful (night, web, nothing that touches the pose).
- Confidence: medium. The start of this cycle is outside the batch.
2026-11-04T21:15Z → 2026-11-07T14:00Z
- Map phenomenon: A target that degrades, and sometimes recovers (CTR01, continuation of NORDEND 16.10 19:00)
- Family: Target issue, like the reviewed label NORDEND-195 of the October batch (rule: same target, same family).
- Measured: CTR01 rejected at 124 measurements out of 260; median error from 0.42 to 0.38 per day (threshold 0.4), as since 16/10 (
scripts/par_jour.py, from 20/09 to 12/11). Rejections almost all night from 04 to 05/11, then every morning until early afternoon (05/11 07:00-14:45, 06/11 06:45-14:45, 07/11 06:30-14:00). Outside the compensation (pattern_type=ref): its rejections do nothing to the other targets. - Images: thumbnails of 04/11 23:00, 05/11 10:00 and 22:00, 06/11 11:00, 07/11 04:00 and 12:00 against PR17. Small spot at night, almost invisible by day when PR17 stays sharp.
- Confidence: medium. Old and constant regime; the cause remains not established.
2026-11-05T00:45Z → 2026-11-05T01:30Z
- Map phenomenon: A spider web in front of the lens (a thread on a reference), with A reference discarded: the base changes
- Family: Obstruction. A seen thread hides a reference (boundary "A hindered reference"; same family as NORDEND 18.08 22:15, thread on CTR03).
- Measured: CTR03 at 0.295 and 0.356 (00:30, 00:45), rejected at 01:00 (0.404) and 01:15 (0.553), 0.280 at 01:30. At both rejections, 0 reference retained (
inlierempty for CTR02 and CTR04), reprojection of CTR02 and CTR04 at zero,sigma0empty; 17 targets out of 18 more than 4 noises incorr_y(median −0.073 and −0.079), 11 incorr_x(+0.015). PR17 and PR18 at 28 and 26 noises. Nothing at 00:45 nor at 01:30 (scripts/pics_ctr03.py). CTR02 and CTR04 stay under their threshold. - Images: zone around the references at 01:00: a big blurred bright point of the thread on the spot of CTR03. Whole image of 05/11 23:00 for the web.
- Confidence: strong.
2026-11-05T03:00Z → 2026-11-05T09:45Z
- Map phenomenon: A spider web in front of the lens (a thread on a spot)
- Family: Obstruction. A single target, a thread seen in front of it at night.
- Measured: error of PR09 at 0.29 in median over the morning (0.10 to 0.12 on the other days), rejected at 03:15, 05:45, 06:15, 07:15, 07:45, 08:15, 08:30, 08:45 and 09:30; PR10 and PR11 at their level. The
corrof PR09 does not move outside the rejections of CTR03 (ranges.py --rolling 8). Bounds: the measurement before the first rejection, the measurement after the last. - Images:
--fixedthumbnails of PR09 and PR10: at 06:00, the halo of the thread covers the background of PR09; at 08:00 (day), nothing visible; the night of 04/11 23:00, nothing. - Confidence: medium. The thread is seen at night; by day, the cause is deduced.
2026-11-05T03:15Z → 2026-11-05T03:45Z
- Map phenomenon: A spider web in front of the lens; A reference discarded: the base changes
- Family: Obstruction. Same reference, same signature, wandering thread seen at 01:00 and 04:00.
- Measured: CTR03 rejected at 03:30 only (0.534); 0 reference retained; 17 targets out of 18 jump in
corr_y(median −0.081), return at 03:45. PR09 is rejected at 03:15, the start bound: it is the other event (PR09 label), which overlaps this one. - Images: no image at 03:30 (one per hour); at 04:00, the thread is between CTR03 and CTR02.
- Confidence: medium (cause deduced).
2026-11-05T06:30Z → 2026-11-06T07:00Z
- Map phenomenon: The camera moves through a thermal effect
- Family: Thermal displacement.
- Measured: morning turning point at 06:30 (pan and tilt leave their night level at 06:45). Housing temperature from 19.2 to 29.7 °C (max at 14:00). Pan lowest 0.0099° around 14:15 (−44 noises of the night), tilt highest around 11:30-12:00, rotation lowest at 14:00, scale highest at 14:45 (
days.py --night 3-5 --bounds). Slow return in the evening and at night, without rest until the turning point of 06/11 at 07:00. Day minus night, in noises:corr_xPR18 −7, PR16 −5, PR13, PR14, PR15, PR17 −3;corr_yPR18 and PR17 +6, PR16 +5. Jolts: 09:45 (8 targets a little low incorr_y, in the middle of the fast tilt movement) and 12:00. At 12:00, the raw medianyof all sensors, references included, takes +0.43 px for one measurement, the rotation and the tilt jump: 12 targets more than 4 noises incorr_y(median −0.020), return at 12:15. Placed in the cycle label (rule of the pose agitation). - Images: none needed. The image of 06/11 12:00 shows an overcast day, without obstacle in front of the group.
- Confidence: strong for the cycle, medium for the choice of keeping the jolt of 12:00 in the cycle.
2026-11-05T14:45Z → 2026-11-05T15:15Z
- Map phenomenon: A spider web in front of the lens (a thread on a spot)
- Family: Obstruction. Same target, same cause as in the morning; separated from the morning by more than five hours without rejection.
- Measured: PR09 rejected at 15:00 (0.410), neighbors at their level; PR09 still right at the threshold in the afternoon (high error range 14:15-19:30).
- Images: thumbnail of 15:00: nothing visible (day).
- Confidence: low to medium (cause deduced).
2026-11-05T23:00Z → 2026-11-06T10:00Z
- Map phenomenon: A spider web in front of the lens (placed on a reference, it makes it move away); A reference discarded: the base changes
- Family: Obstruction. Thread seen on CTR03.
- Measured: the error of CTR03 goes from 0.12 (19:30) to 0.21 (19:45) then 0.32 (20:00) and stays between 0.25 and 0.40. Rejected at 13 measurements, from 23:15 to 09:45. At each, 0 reference retained, reprojection of CTR02 and CTR04 at zero, 15 to 18 targets out of 18 more than 4 noises in
corr_y(median −0.065 to −0.080). Between the rejections, nothing clear: at most 3 targets off level. Start at the measurement before the first rejection, and not at 19:45. From 19:45 to 23:00, CTR03 accepted shifts the targets only under their noise: against the targets, CTR03 moves away by about 0.01 px (overview.py --rigid), andcib_corr_yrises by 2 noises at most. - Images: thumbnails and zone at 21:00, 23:00, 06/11 06:00: a wide bright blurred band of the thread on CTR03; CTR02 and CTR04 sharp. Whole image of 23:00 (
fig2). - Confidence: strong.
2026-11-06T07:00Z → 2026-11-07T06:30Z
- Map phenomenon: The camera moves through a thermal effect
- Family: Thermal displacement.
- Measured: turning point at 07:00 (pan 0.0383 then 0.0374 at 07:30; the poses of 06:00, 06:45, 07:15 on two references are left out of the reading). Housing up to 27.2 °C. Pan lowest around 13:45 (−39 noises), tilt highest around 13:00, scale highest at 13:00. Slow return, incomplete on the morning of 07/11: pan lower by 11 noises, tilt higher by 16 noises than the night before; the night is warmer (housing around 22 °C against 19 °C). Day minus night:
corr_xPR18 −10, PR16 −8, PR15 −6, PR13 −5, PR17 and PR14 −4, PR04 +7;corr_yPR18 +8, PR16, PR17, PR15, PR08 +7. - Images: 06/11 12:00, overcast day, nothing in front of the group.
- Confidence: strong.
2026-11-06T09:15Z → 2026-11-06T09:45Z
- Map phenomenon: A spider web in front of the lens (a thread on a spot)
- Family: Obstruction. A single target; the thread seen on it the night before.
- Measured: error of PR10 from 0.24 to 0.26 at 09:00-09:15, rejected at 09:30 (0.419), 0.095 at 09:45. Nothing on its neighbors. This label falls under that of CTR03 (rejected at 09:45, the end bound): two events, two sensors, overlap intended ("what falls under a long label is labeled apart").
- Images: thumbnail of 05/11 22:00: halo of the thread on PR10. On 06/11 by day, nothing.
- Confidence: low to medium.
2026-11-06T21:15Z → 2026-11-07T11:00Z
- Map phenomenon: A spider web in front of the lens; A reference discarded: the base changes
- Family: Obstruction.
- Measured: CTR03 between 0.18 and 0.30 all of 06/11 by day (without rejection or effect on the curves), then rejected at 22 measurements from 21:30 to 10:45, tighter in the morning (09:30 to 10:15 in a row). At each, 0 reference retained, 15 to 17 targets in
corr_y(median −0.058 to −0.075). After 10:45, CTR03 stays under the threshold (0.28 to 0.40) until 12:15, without effect on the curves. It falls back to 0.108 at 12:30. - Images: thumbnail of 07/11 04:00: the bright band on CTR03; 07/11 10:00 (day): nothing visible; 07/11 20:00 (after the batch): the thread moved to the right of CTR03.
- Confidence: strong.
2026-11-07T06:30Z → 2026-11-07T14:00Z
- Map phenomenon: The camera moves through a thermal effect
- Family: Thermal displacement.
- Measured: turning point at 06:30 (pan 0.0296 then 0.0286, tilt −0.1412 then −0.1382). Housing from 22 to 26.3 °C, less hot than the days before. Pan lowest around 13:00, tilt highest around 10:15, rotation lowest around 09:45, scale rising. Not back at the end of the batch. Day minus night:
corr_xPR18 −7, PR13 −6, PR16 −4;corr_yPR06 +17, PR16 +8, PR15 +6. The reference night is low for PR06 (−0.02 the night of 06 to 07/11). - Images: 07/11 14:00, the web against the light on the group (
fig5); nothing else. - Confidence: medium. The cycle is cut by the end of the batch, and the reference night is mixed with the rejections of CTR03.
2026-11-07T11:15Z → 2026-11-07T11:45Z
- Map phenomenon: A spider web in front of the lens (a thread on a spot), cause deduced
- Family: Obstruction.
- Measured: error of PR04 rising since 09:30 (0.15 to 0.19), rejected at 11:30 (0.733), 0.129 at 11:45; neighbors at their level;
corrof PR04 without a range off level apart from this point. - Images: thumbnails of PR04 at 11:00 and 12:00: whole spot, background a little brighter, nothing clear. Image of 14:00: the web crosses the group.
- Confidence: low (cause deduced from a thread seen near the group that day, not on PR04).
2026-11-07T13:45Z → 2026-11-07T14:00Z
- Map phenomenon: A spider web in front of the lens (a thread on a spot)
- Family: Obstruction.
- Measured: PR07 rejected at 14:00 (0.401), last measurement of the batch;
endat the batch edge, the next measurement (14:15) being outside the selection. Over 6 measurements (12:00-14:00), its error is higher than usual. - Images: thumbnail and image of 14:00: the background of PR07 lit by a bright band of the thread that descends through the group (
fig5). - Confidence: medium.
2. What I left without a label
- CTR02 right at the threshold on 05/11, from 01:45 to 16:15 (error between 0.08 and 0.38, without rejection): the thread of the web passes over it (zone at 04:00). No target moves at these instants: at most 1 to 3 targets more than 4 noises from their local level (
pics_ctr03.py). An error that rises without effect oncorris not labeled. - CTR03 accepted but veiled (05/11 19:45-23:00, 06/11 by day, 07/11 11:00-12:15): barely shifted against the targets, without effect on the
corrbeyond their noise. - Error rises of the targets without rejection or jump: PR01 and PR02 on 05/11 (night and morning), PR11 during the nights of CTR03, PR12 and PR17 on 07/11 at the end of the morning (PR17 up to 0.29 for a threshold of 0.4).
ranges.py --rolling 8finds no own deviation ofcorrfor them. - The web itself, present the whole batch: apart from the named sensors, it does not veil the scene. The error of the other targets stays at its level; it is not a Visibility.
- The dirty glass (NORDEND 25.10 05:45, closed on 01/11): no step nor change of
corrlevel linked in this batch; not taken up. - The thermal movement is labeled every day, as D-117 requires.
3. The documentation
- What served: the sheet "A spider web in front of the lens", especially "placed on a reference, it makes it move away: all targets on a step" (NORDEND 18.08 22:15), and the boundary "A hindered reference" of
families.md, which give the family. The sheet "A reference dropped: the base changes" for the mechanism (reprojection at zero). The expert's rule "three references are needed". The sheet of the NORDEND thermal cycle, without rest at night, and the morning turning point rule. - What I missed:
- A word on the comb: a reference right at the threshold for hours, rejected one measurement in two or three. The sheet says "step" (a reference hidden and rejected in a row). Here, they are one-measurement peaks, repeated all night, to be cut into one episode per night rather than one event per rejection. I followed the rule "one label per parking or per episode". The threshold of the "pause of several measurements" that cuts an episode is not quantified: I cut at the day without rejection (06/11) and kept together two-hour pauses at night.
- How to bound an episode whose nuisance (CTR03 right at the threshold) starts hours before the first touched point. I bounded on the curves (first rejection), as the bound rule asks.
- What seemed contradicted: nothing.
Difficulties
- The instruction asks to read the whole batch, but the batch from 01 to 04/11 has no label. The thermal cycle and the regime of CTR01 therefore start at the batch edge, with no neighboring label to extend.
days.py --bounds: the reference night is mixed with the peaks of CTR03 (the measurements of the targets there remain accepted). The night level of thecorris distorted by it:corr_y07/11 "96 measurements beyond the threshold". I reread the cycles with my own script, which leaves out the instants where CTR03 is rejected (scripts/cycle_jour.py).
Sources read
.claude/skills/lire-les-donnees/SKILL.mdlabeling/label-map.md: summary, sections 2 to 4, sheets B (l. 119-290: passage, headlight, vehicle, animal, web, plant, obstacle on the spot, dirty glass, rain, visibility, part of the scene, reference that loses its brightness, turbulence), C, D and E (l. 293-434)labeling/families.md,labeling/labeling-rules.md(in full)labeling/knowledge/:measurement-chain.md,displacement-columns.md,gaps.md,thermal-cycle.md,camera-images.md,quality-metrics.md,expert-annotated-cases.mdlabeling/auto/NORDEND.jsonl(the labels of the October batch that touch CTR01, CTR03, PR09, the web and the thermal cycle)- Data:
data/NORDEND/throughinventory.py,overview.py(with--rigid),metrics.py,days.py --bounds,ranges.py,summary.py,plot.py,thumbnails.py, the whole images of 05/11 18:00 and 23:00, 06/11 12:00, 07/11 14:00; the raw CSVs through my scriptsscripts/pics_ctr03.py,scripts/par_jour.py,scripts/cycle_jour.py,scripts/bornes.py