Files
aimsight-labeling/auto/analyses/BREITHORN_2026-10-03T0000_2026-10-13T0000/analysis.md
T

92 lines
11 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: BREITHORN, 2026-10-03T00:00Z → 2026-10-13T00:00Z (UTC)
Task: verify the 8 existing labels of the range (`work/existing.jsonl`, all reviewed on 2026-11-17) and the spanning label BREITHORN-2. Cadence 4 h (00, 04, 08, 12, 16, 20 UTC), 61 instants, no missing step, a pose at every instant, no image-quality metrics (old camera). No weather: BREITHORN has no place in `~/aimsight-data/extraction/stations.txt` (the line has the name only), so I fetched no MeteoSwiss data. The station is indoors, in a factory (`stations.md`).
## Verdict per old id
| id | verdict | evidence |
|---|---|---|
| BREITHORN-1 | kept | REFPOU22, COLH20''' and POUI22 rejected from 03T04 to 04T16 (10 steps, inventory); everything back at 04T20; the vehicle is at the bottom of the 04T08 image; the corr_x of REFH22 goes from about −0.009 to −0.03/−0.046 and comes back at 04T20. Starts at the first measurement of the data (edge). |
| BREITHORN-2 (spanning) | correct inside the range | drift of corr_x: COLJ19 0.076 → 0.170, POUI19 0.110 → 0.287, COLI21 −0.001 → 0.081, COLH20 0.029 → 0.070, POUI20 0.036 → 0.077, COLI20 −0.003 → 0.040; COLI24 corr_y −0.033 → −0.138; POUI19 corr_y flat. Tilt 0.0094 → 0.0197, pan and rotation going down, housing temperature 31.5 → 30. Nothing to change. |
| BREITHORN-3 | kept | 04T16: every sensor out of level (rolling ranges: REFH22 corr_y +44 noises, refs −24, targets −9 to −26); REFH22 error_value 0.566 (threshold 0.7), POUI24 rejected; back at 04T20. |
| BREITHORN-4 | kept | COLI22, COLI24, POUI22 rejected 06T20 → 07T04; nothing else out of level; back at 07T08. |
| BREITHORN-5 | kept | 07T12: REFI'21 rejected (error 0.99), 3 references retained; COLJ19, POUI19 corr_y +9 noises, COLI20, COLI22, POUI24 +6 to +8; back at 07T16. |
| BREITHORN-6 | kept | COLI24 rejected 07T20–08T00, COLI22 and POUI22 08T00; POUI22 corr_y +0.199 at 07T20 (+18 noises), POUI24 corr_y +0.409 at 08T00 (+59 noises); back at 08T04. |
| BREITHORN-7 | kept | REFH22, COLI22, COLI24, POUI22 rejected 09T00 → 09T04; POUI24 corr_x +0.17, corr_y +0.27/+0.35 (+40/+52 noises); back at 09T08. The other targets shift a little (REFG44N, REFI'21 corr_y −0.013, COLI21 −0.033, under 3 noises): "without making the other targets jump" holds at the 5-noise threshold. |
| BREITHORN-8 | kept | COLI22, COLI24, POUI22 rejected at 09T20 only; nothing else out of level. |
| BREITHORN-9 | kept | 10T04: COLI22, COLI24, POUI22, POUI24 rejected; REFH22 error 0.475, corr_y +70 noises; every other sensor −8 to −27 noises; back at 10T08. |
## 1. My labels
The 8 kept labels are in `labels.jsonl` unchanged (with their review field). Here are the 4 added ones.
### 2026-10-04T00:00Z → 2026-10-04T12:00Z
- **Phenomenon of the catalogue**: the camera moves through a thermal effect.
- **Precedents**: BREITHORN-14, -21, -24, -17 (reviewed): the same morning signature at this station (rotation and scale deviate at the temperature trough, COLH20 and REFPOU22 up in corr_y, REFG44N/REFI'21 down, back at noon), labelled Thermal displacement on the expert's word. They set the family, the bounds (from the last calm measurement to the noon return) and the wording.
- **Family**: Thermal displacement. Not Visibility: the error_value of COLH20 stays at its level and nothing is in front of it. Not Obstruction: the vehicle of BREITHORN-1 hides only COLH20''', POUI22 and REFPOU22.
- **Measured**: compared with the previous evening (`scripts/thermal_table.py`, `work/thermal_table.txt`), COLH20 corr_y is +0.049 at 04T04 and +0.046 at 04T08, then +0.011 at 04T12. COLI21 is +0.017 at 08. The rotation is −2.9e-3° and −2.5e-3° (`days.py --bounds`: −6.3 noises, 04T04 → 04T08); the scale is −2.5e-5 and −5.1e-5. The housing temperature goes 32.5 → 30.5 → 29.5 °C, then back to 31. REFPOU22 is rejected (3 references), so the usual reference signature cannot be read.
- **Images**: 04T08: the parked vehicle is in the foreground at the bottom of the scene; the spots in the upper part, including COLH20, are clear.
- **Confidence**: medium. The rotation and COLH20 match the signature, but with only 3 references and the vehicle present the case is less clean than 5 and 6 Oct.
### 2026-10-05T00:00Z → 2026-10-05T12:00Z
- **Phenomenon of the catalogue**: the camera moves through a thermal effect.
- **Precedents**: as above; BREITHORN-24 (28 Oct) has the same amplitude and the same two-measurement shape.
- **Family**: Thermal displacement. The error_value of the references stays far below threshold (REFG44N 0.09 → 0.125, REFPOU22 0.057 → 0.066), with no rejection and nothing on the images.
- **Measured**: compared with the previous evening:
- at 05T04 / 05T08, REFPOU22 corr_y is +0.049 / +0.046 (+5.5 noises with a rolling level), COLH20 +0.040 / +0.062, REFG44N −0.027 / −0.012 (−5.3 noises at 04), REFI'21 −0.021 / −0.008;
- the scale is −5.3e-5 / −7.7e-5 (−6.5 noises at 08) and the rotation −1.4e-3 / −2.6e-3 (−5.5 noises at 08);
- at 05T12 everything is back (+5, +13, +3, +5 ×1e-3);
- the housing temperature falls to its daily trough (30.0 / 29.5 °C).
The corr_y of COLJ19 and POUI19 rise at 05T08 and keep part of it: that belongs to the drift of BREITHORN-2.
- **Images**: 05T08 and 05T16 are identical in content: spots clear, nothing in front.
- **Confidence**: high.
### 2026-10-05T20:00Z → 2026-10-06T12:00Z
- **Phenomenon of the catalogue**: the camera moves through a thermal effect.
- **Precedents**: as above; BREITHORN-17 and -25 also start at night.
- **Family**: Thermal displacement (same reasons).
- **Measured**: compared with the previous evening:
- at 06T00 / 04 / 08, REFPOU22 corr_y is +0.029 / +0.040 / +0.045 (+5.0 noises at 08), COLH20 +0.040 / +0.035 / +0.045, REFG44N −0.022 / −0.014 / −0.018, REFI'21 −0.019 / −0.011 / −0.014;
- the scale is −2.9e-5 / −2.8e-5 / −6.1e-5 (−5.2 noises at 08);
- at 06T12 the corr are back (+9, +15, −1, −1); the scale is still −3.1e-5 there and back at 16. That is the slow relaxation of a secondary axis, which does not push the end back;
- at 05T20 all is calm (+3, +13, −1, +1);
- the housing temperature is 31.5 → 30.0 → 31.0 °C.
- **Images**: 06T08: spots clear, a small light at the far left, away from any spot.
- **Confidence**: high.
### 2026-10-11T20:00Z → 2026-10-12T16:00Z
- **Phenomenon of the catalogue**: the camera moves through a thermal effect.
- **Precedents**: BREITHORN-21, a weaker case accepted in review.
- **Family**: Thermal displacement.
- **Measured**: compared with the previous evening:
- at 12T00 / 04 / 08, REFPOU22 corr_y is +0.021 / +0.025 / +0.025, COLH20 +0.010 / +0.028 / +0.019, REFG44N −0.014 / −0.007 / −0.007;
- COLI21 is −0.011 / −0.026 / −0.016 and COLH20''' −0.014 / −0.023 / −0.013;
- POUI24 is −5.9 noises at 12T04 (rolling level), and POUI19 −5.2 noises at 12T08 (corr_y −0.172 / −0.179 against about −0.12 to −0.13 before). POUI19 is still −0.146 at 12T12 and back at 12T16 (−0.100), hence the end;
- the scale is −2.7 / −3.8 / −3.3 e-5 and the rotation −0.4 / −1.3 / −1.3 e-3;
- the housing temperature is 29.5 → 28.75 °C, the lowest of the range, then 30–31.
The corr of single targets pass 5 noises, so the "no cycle" rule does not apply.
- **Images**: 12T04: identical to the calm images, nothing in front.
- **Confidence**: medium. The pattern is coherent, but it is the weakest of the four, and only two targets pass 5 noises.
## 2. What I left without a label
- **3 Oct, morning**: COLH20 corr_y is higher at 03T08 (about +0.028 against the 03T04/03T12 values) and the scale is −3e-5 (2.7 noises), with the housing temperature at 30.25. There is no previous evening to compare with, REFPOU22 is rejected from 03T04 (the base changes), and nothing passes 5 noises. Not labelled.
- **11 Oct, morning**: REFPOU22 +0.013 / +0.018, COLH20 +0.011 / +0.019 at 04/08, scale −3.8e-5 (3.2 noises). No sensor and no pose axis reaches 5 noises: no thermal label (labeling-rules §2/§1, 5-noise rule).
- **7 Oct night and 8 Oct**: the scale is low all day and the rotation steps down from 08T08 (−8 noises, no return). That is the progressive tilt of BREITHORN-2 (spanning), not a daily cycle; COLH20/REFPOU22 do not show the morning signature.
- **12T16 and 12T20**: the camera pose is exactly the same at both instants, and so is the corr of the four references, while the targets differ a little (within their noise). It looks like a registration reused from the previous measurement. The corr show nothing abnormal, so there is no label; noted as a data oddity.
- **The drift itself** (corr_x of COLJ19, POUI19, COLI21…) is BREITHORN-2, out of my scope.
## 3. The documentation
- The cases that decided it were the reviewed BREITHORN thermal labels (14, 17, 21, 24, 25, 27) and NEBRASKA-10. Together they give a very specific signature for this indoor station: rotation and scale deviate at the housing-temperature trough, COLH20 and REFPOU22 rise in corr_y, REFG44N and REFI'21 go down, everything is back at noon. Searching for it over the whole range (labeling-rules §1, "a signature found once is searched over the whole batch") found 4 occurrences that the existing labels of the range missed. The previous analysts of this range seem to have looked only at rejections.
- The 5-noise rule ("no thermal label on a day when neither the pose nor the corr deviate by more than 5 noises") sits uneasily with this precedent. On a 4 h cadence the noise is the step between consecutive measurements, and those steps already include the cycle. Several reviewed labels (BREITHORN-21: REFPOU22 about 3.5 noises) fall below it. I applied the rule per sensor and per pose axis: 11 Oct was left out, 12 Oct kept because POUI19 and POUI24 pass it. A note in `labeling-rules.md` on how to measure the noise at a slow cadence would help.
- `days.py --bounds` on the median of the targets did not show these cycles, because the targets move in opposite directions and cancel. The per-sensor table of my script and `ranges.py --rolling 6 --instants 2` showed them.
- No weather for BREITHORN (no place in `stations.txt`); for an indoor station it matters little.
## Difficulties
- My first run of `thermal_table.py` matched no instants: the measurement times carry seconds (`00:00:03.37`), while the pose and temperature are on the minute. I fixed it by rounding to the minute.
- `validate.py` flags the overlap of BREITHORN-1 and BREITHORN-3 (two Obstruction labels). It was already there and was reviewed by the user, so I kept it as it is.
## Sources read
`instructions/analyst.md`, `tools/GUIDE.md`, `tools/DATA.md`, `labeling/families.md`, `labeling/labeling-rules.md`, `labeling/reading-rules.md`, `labeling/stations.md`, `labeling/auto/BREITHORN.jsonl` (ids, and the texts of 10, 11, 14, 17, 21, 24), `work/existing.jsonl`, `work/spanning.jsonl`, the precedents (`precedents.py`, thermal morning query). Data: `inventory.py`, `summary.py` (corr_x, corr_y, error_value at 4 h, `--mark`), `metrics.py` (4 h, 2 → 14 Oct), `days.py --night 12-23 --bounds`, `ranges.py --rolling 6 --instants 2`, `plot.py`, `scripts/thermal_table.py` (my own, 3 → 30 Oct), images 04T08, 05T08, 05T16, 06T08, 12T04. `tools/scripts/data.py` was read for its loader API only.