Files
aimsight-labeling/auto/analyses/NORDEND_2026-11-05T0600_2026-11-05T1200/analysis.md
T

53 lines
8.1 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Labels: NORDEND, 2026-11-05T06:00Z → 2026-11-05T12:00Z (UTC)
The range is already covered by four labels of `labeling/auto/NORDEND.jsonl`, which I keep and do not redo:
- NORDEND-229 (Thermal displacement, 04/11 21:15 → 05/11 06:30): the end of the night cycle, up to the morning turning point;
- NORDEND-234 (Thermal displacement, 05/11 06:30 → 06/11 07:00): the thermal cycle of the day;
- NORDEND-230 (Target issue, CTR01, 04/11 21:15 → 07/11 14:00): CTR01 rejected at 23 of the 25 measurements of the range;
- NORDEND-232 (Obstruction, PR09, 05/11 03:00 → 09:45): the rejections of PR09 (06:15, 07:15, 07:45, 08:15 to 08:45, 09:30).
I add a single label, the camera jolt at 12:00.
## 1. My labels
### 2026-11-05T11:45Z → 2026-11-05T12:15Z
- **Map phenomenon**: A camera step that the compensation does not erase.
- **Precedents**: NORDEND-80 (01/09 15:30, Displacement, reviewed): same shape, a one-measurement step of the pose, several targets that jump together and come back. I took up its family, its one-measurement bounds on either side and its conclusion. SNCF01-42 (15/10 05:15): a step that stays, also classed Displacement. MICHIGAN-2: a one-measurement deviation, classed Thermal because the housing temperature makes a brief drop there, which is not the case here (temperature in regular rise, without a jolt over the hour).
- **Family**: Displacement. It is not the thermal agitation, because the pose does not make a larger step in the direction of the morning drift: it jumps (rotation, tilt), then comes back right away at the next measurement. `families.md`, "what is not labeled": a camera step that leaves something on the `corr` is labeled, and concludes "the camera moves". It is not Visibility (no rise of the `error_value` specific to the instant, sharp image) nor Obstruction (no reference lost, nothing in front). NORDEND-234 already describes the jolt in its text; the rule "what falls under a long label is labeled apart" gives it its own label.
- **Measured**:
- camera pose (`overview.py`, 15 min table): rotation −0.0096° at 11:45, −0.0054° at 12:00, −0.0106° at 12:15; tilt −0.1209°, −0.1161°, −0.1229°; sigma0 from 0.00209 to 0.00181 then 0.00211;
- raw positions: median of `y` at +0.77, +1.21, +0.74 px; median of `x` at −0.15, −0.25, −0.12 px; the three references with the targets (CTR04 in `y` up to +1.28);
- `--rigid` (base 06:00): fitted rotation from −0.0092° to −0.0031° then −0.0089°; residual of CTR04 at +0.032 / +0.029 px, above its morning level; `cib_max` at 0.060 px against about 0.04 before and after;
- `ranges.py --rolling 4 --instants 3`: 8 sensors off level at 12:00, and only at that instant. In `corr_y`: PR01 −7, PR07 −6, PR10 −6, PR02 −5, PR04 −5 noises; in `corr_x`: CTR04 +6. The `corr` of the references also move (CTR02 in `corr_y` −6, CTR03 +6);
- `summary.py`, `corr_y`: PR01 0.162 → 0.123 → 0.161; PR10 0.188 → 0.148 → 0.184; PR04 0.139 → 0.103 → 0.128; PR07 0.102 → 0.072 → 0.102; PR02 0.170 → 0.135 → 0.158; everything comes back at 12:15;
- what does not move: no reference lost (3 retained at the 25 instants), no rejection outside CTR01 (usual regime). The `error_value` of PR01, CTR02 and PR09 is high at 12:00 (0.23 to 0.25), but these three sensors go like this from a low level to a high level all morning (10:15 to 10:45, 11:15), without a pose jolt or effect on the `corr`: it is not the cause. The housing temperature rises regularly (about 25.5 °C at 11:00, 27.8 °C at 12:00, hourly median).
- **Images**: the whole image of 12:00 shows a sunny day, the left of the scene in the sun, the targets in the shade, nothing in front of them, only a blurred bright streak in the sky, far from the targets. Thumbnails of PR09, PR10, PR12, PR01, CTR03 and PR17 at 06:00, 09:00, 10:00 and 12:00: sharp spots by day, the same at 10:00 and at 12:00.
- **Confidence**: medium. The shape is clear and passes the threshold of 5 noises on 8 sensors, but the deviation is small (a few hundredths of a px). The cause is not seen. A second jolt of the same shape follows at 12:30 (rotation at −0.0034°, tilt at −0.1309°), outside the range and not read.
## 2. What I left without a label
- **09:45, one-measurement dip in `corr_y` of PR09, PR10, PR12 (−5 to −7 noises), weaker on PR03, PR05, PR13, PR02, PR01; return at 10:00.** The pose makes no jolt there: it is the moment when the morning thermal drift goes fastest (tilt +0.0055° in one step, against +0.002 to +0.003 around), and it continues in the same direction without coming back. The `error_value` drops instead of rising. According to the rule "the agitation from one measurement to the next is written in the thermal label", it stays in NORDEND-234, whose description already says it ("the pose shakes from one measurement to the next: at 09:45"). It is the most fragile boundary of my work: a reviewer could make it a second one-measurement Displacement.
- **The drift of the right edge** (`corr_x` of PR18, PR16, PR15, PR13, PR14 dropping, `corr_y` of PR17, PR18, PR16 rising) **and the slow rise of the `corr_x` of PR04 and the `corr_y` of PR08** follow the pose from 06:30 to noon: it is the thermal cycle, NORDEND-234. Its description does not name PR04 or PR08 for this day; they are not separate events.
- **CTR01**: rejected almost everywhere, covered by NORDEND-230.
- **PR09**: the rejections until 09:30 are covered by NORDEND-232. Its end at 09:45 is perhaps one step too early: PR09 is still low in `corr_y` at 10:00 (0.126 against about 0.15 to 0.17), i.e. about 3 to 4 of its noises, and back at 10:15. Its dip of 09:45 is shared with PR10 and PR12 (above). I do not change it: the deviation of 10:00 stays at the edge of its noise.
- **The back-and-forth of the `error_value` of CTR02 (reference), PR01 and PR09** between a low level and a high level, under the threshold except the rejections of PR09: it does not touch the `corr` at those instants (10:15, 10:30, 10:45, 11:15). It is not labeled.
## 3. The documentation
- **What served**: the sheet "A step of the camera that the compensation does not erase" and the precedent NORDEND-80, for the family and the conclusion; the boundary of `families.md` on the camera step that leaves a remainder in the `corr`; rules 1 ("what falls under a long label is labeled apart") and 5 (agitation in the thermal label), to separate 12:00 from 09:45; the NORDEND context of section 4 (rock face, camera cycle every day, web).
- **What I missed**: a clear rule between "agitation of the pose, in the thermal label" and "one-measurement camera step, labeled apart", when the jolt falls in the middle of a cycle. I decided by the shape of the pose: a step that comes back at the next measurement is a jolt, a larger step in the direction of the drift is agitation. This criterion is mine, not the map's.
- **What seemed contradicted**: NORDEND-234 puts the jolt of 12:00 in the thermal cycle, without its own label. For me, it requires one (above). The rest of NORDEND-234 is right on the range: start at 06:30, at the turning point of the tilt and the pan, verified on the pose table.
- **Detectors**: not read (DETECTORS: no).
## Difficulties
- `labeling/precedents.py` failed at first: Ollama did not answer. The container then started by itself ("health: starting"), and the search worked on rerun.
## Sources read
- `instructions/analyst.md`, `tools/GUIDE.md`
- `labeling/label-map.md`: summary, sections 2 to 4, sheets "The camera moves through a thermal effect", "The camera drifts away when the temperature of its housing drops, without a daily cycle", "A step of the camera that the compensation does not erase"
- `labeling/families.md`, `labeling/labeling-rules.md` (in full)
- `labeling/knowledge/measurement-chain.md`, `displacement-columns.md`, `gaps.md`
- `labeling/auto/NORDEND.jsonl` (labels from 04 to 07/11 and Target issue of CTR01)
- precedents: two `precedents.py` queries (NORDEND-80, SNCF01-42, MICHIGAN-2, MICHIGAN-8, LENZSPITZE-45 and other thermal cycles)
- tools: `inventory.py`, `overview.py` (with and without `--rigid 2026-11-05T06:00Z`), `ranges.py --rolling 4 --instants 3`, `summary.py` (`corr_y`, `error_value`), `metrics.py`, `plot.py`, `thumbnails.py`, image of 12:00