62 lines
12 KiB
Markdown
62 lines
12 KiB
Markdown
# Automatic labels: LENZSPITZE, 2026-11-01T10:30Z → 2026-11-01T11:40Z (UTC)
|
||
|
||
Range chosen by the user in the Data page; their hint: "Around 11:00, the track targets move a little; a train?". No label of `labeling/auto/LENZSPITZE.jsonl` touches the range (the last one, LENZSPITZE-64, ends on 01.11 at 00:00). 5 min rate, plus one measurement at 11:01 (these off-rate measurements, at minute 01, often return on this station); 16 instants, no missing step, no rejection, the 9 references retained everywhere, one camera pose at each instant (`inventory.py`).
|
||
|
||
## 1. My labels
|
||
|
||
### 2026-11-01T10:30Z → 2026-11-01T11:40Z
|
||
- **Map phenomenon**: The camera moves through a thermal effect
|
||
- **Family**: Thermal displacement. It is always labeled (D-117). Here the pose follows the warming of the housing and the `corr_y` come back with it in the afternoon. It is not Displacement: the return is seen over the day, and the same cycle returns every sunny day on the station.
|
||
- **Measured**: `metrics.py` from 31.10 to 02.11: housing temperature (`temperature_deg_c`, the only one filled) at 26.5 °C at 10h, 27.3 °C at 11h and 29.7 °C at 12h, higher at 13h, then dropping; r(cam_pan, temperature) = −0.81 over 49 h. `overview.py` over the day: start of the pose around 06:50 (scale and tilt), then a second phase from 11:05 (tilt), then 11:15 (pan dropping, rotation rising), return around 14:15-15:00. In the batch: tilt from 0.0139° to 0.0074° between 11:00 and 11:05, pan from 0.0116° to 0.0047° between 11:10 and 11:15, rotation from 0.002° to 0.013° between 11:15 and 11:40; scale at 0.99968-0.99973. Raw positions: `x` and `y` of all references and of the median of the targets move together (−0.5 px in `x` at 11:15). `summary.py --step 30min` from 03:00 to 16:00: the `corr_y` of 1018, 1019, 1016, 1008 and 1020 drops at 07:00, stays low until 10:30, then rises from 11:00 toward its night level, reached around 13:00. `--rigid` with base 10:30: the targets move as a block (`tx`, `ty`), except the own deviations described under the label of 8012. Bounds: the cycle touches both edges of the batch, with no neighboring label, hence 10:30 and 11:40 (rules, section 2).
|
||
- **Images**: 10:01, 11:00 and 12:01: full sun on the track, sharp shadows.
|
||
- **Confidence**: medium. The shape of the cycle is sure; the tilt step at 11:05 and that of the pan at 11:15 are more abrupt than an ordinary cycle and fall at the same instants as the train and as 8012. I read them as the cycle, because the pose keeps them and comes back in the afternoon.
|
||
|
||
### 2026-11-01T10:55Z → 2026-11-01T11:05Z
|
||
- **Map phenomenon**: A passage hides the targets for the time of one measurement (variant: the flux stays whole, the `error_value` rises and a second pattern appears)
|
||
- **Family**: Obstruction. The passage is seen on the image, and it touched a group of sensors, not the whole scene through a diffuse nuisance.
|
||
- **Measured**: at 11:00, the `error_value` of 35 sensors out of 41 exceeds their median by more than 5 MAD, references included (8004 the highest, under its threshold); no rejection (`overview.py --min 1`). `quality.py --table` at 1 min, with the calm reference of 30-31.10: at 11:00, median sharpness at 0.62 (0.26 to 0.48 on 8010, 8004, 8003, 1016, 1012, 101B, 1014; 0.86 to 0.89 on 8014 and 8017) and median `dominance` at 2.2 (up to 6 to 7.7 on 8010, 8003, 8004, 1016, 1014, 1012); the flux stays whole (median 1.08), except 1006 at 0.34 at 11:00 and 0.43 at 11:01, back to 1.03 at 11:05. At 11:01, the image is even less sharp (0.83); at 11:05, it is back. `ranges.py --rolling 3 --instants`: 15 sensors off level at 11:00 (`corr_y` of 1022 at −10 noises, 202H at +6, 1009, 1013, 8019 at −5, 1001, 1005, 1020, 8010, 102H, 1017; `corr_x` of 101B, 1012, 1013, 202H, 8019); 5 at 11:01 (`corr_x` of 1008, 1013, 1012, 8013, `corr_y` of 1020); these jumps come back at 11:05 (`summary.py --step 1min`, `corr_y` of 1022, 202H, 1005, 1001 back from 11:01 or 11:05). The step of the `corr_y` of 1018, 1019, 1016, 1008 and 1020 that starts at the same instant and stays belongs to the return of the thermal cycle.
|
||
- **Images**: on the image of 11:00, a train passes on the track to the left of the instrumented track, in transparency and motion blur. On the fixed-frame thumbnails, it covers 8003 and 8010 and passes at the bottom of the references 8011, 8012 and 8013; it is absent at 10:01 and at 12:01 (`thumbnails.py --zone 1300,750,1000,550`, `thumbnails.py --fixed` on 8012, 8011, 8013, 8010, 8003, 1017).
|
||
- **Confidence**: strong. The train is seen at the minute of the measurement, and the image quality gives its signature on the sensors it covers.
|
||
|
||
### 2026-11-01T11:05Z → 2026-11-01T11:40Z
|
||
- **Map phenomenon**: close to "The sun in the field or behind a spot" (variant: the background of a reference in the sun makes a small common jump of several targets, under the threshold); not seen here
|
||
- **Family**: Visibility. It is the light on a reference that changes, with nothing in front. According to the boundaries, a reference that moves away from the others and from the targets, with a sharp edge, is Visibility and not Thermal displacement. It is not Displacement: the signature (flux, `dominance`) is that of a spot whose lighting changes, and it returns on other days at the same hour.
|
||
- **Measured**: `capteur_jours.py 8012` (my script): flux of 8012 from 21,000 to 31,000 between 11:05 and 11:15, `dominance` from 0.59 to 0.34, raw `x` from 2043.51 to 2043.36 px, `corr_x` from 0.01 to 0.33, then everything stays so until 11:45. `overview.py --rigid 2026-11-01T10:30Z`: the residual `dx` of 8012 goes from 0.07 (11:10) to 0.34 px (11:15) and 0.41 px (11:20), to slowly decrease down to 0.16 px at 12:30; the other references stay under 0.1 px. `sigma0` rises from 0.0014 to 0.0030 at 11:15 and stays high. The residual of the targets 8003 (−0.08 px), 8010 (+0.11 px) and 1017 (+0.10 px) changes at the same instant. `corr_x` (residuals of `overview.py`): first deviation at 11:10 (8003 at +8 noises, 8004 at +6), then step at 11:15. 8003 drops by about −0.12, 1011 and 1001 by about −0.08, 1003, 1006, 1021, 201B, 7013H and 8004 by −0.05 to −0.08. 8010 and 1017 rise by about +0.08 and 1002 by about +0.05. The step holds until 11:40. Same pattern on 22.10 (11:10), 27.10 (11:15) and 02.11 (11:20); not on 03.11.
|
||
- **Images**: fixed-frame thumbnails of 8012 at 10:01, 11:00 and 12:01, and on 31.10 at 12:00: nothing in front, nothing that changes visibly at this size. Whole images of 02.11 and 03.11 at 11:01: sun both days. No image between 11:00 and 12:01.
|
||
- **Confidence**: medium. The fact is sure (the reference changes brightness and shifts, the base follows). The cause is only probable, since 03.11, also sunny, does not show it.
|
||
|
||
### 2026-11-01T11:15Z → 2026-11-01T11:25Z
|
||
- **Map phenomenon**: Atmospheric turbulence, or a partial passage like that of 11:00; I could not decide
|
||
- **Family**: Visibility, from the image-quality signature (sharpness down, `dominance` up, flux whole: the image shakes, conventions of `lire-les-donnees`). Choosing Obstruction would have required an image of the passage.
|
||
- **Measured**: `quality.py --table sharpness` at 5 min: median sharpness at 0.81 at 11:20, 0.79 at 11:30 and 0.73 at 11:45, against 0.89 to 0.93 between these measurements; 8014, 8017 and 1021 to 1024 are almost not touched, as at 11:00. At 11:20, the pose makes a one-measurement round trip (tilt at 0.0149°, against 0.0104° and 0.0087° around; common `x` at −0.52 against −0.38 and −0.34). `ranges.py --rolling 3 --instants 1`: 6 sensors between 3 and 6 noises at 11:20 (`corr_x` of 7013H, 1024, 102H, 1010, 92H; `corr_y` of 1017), only one at 11:25; none at 11:30.
|
||
- **Images**: none at 11:20 (hourly images only).
|
||
- **Confidence**: low for the cause, strong for the fact (a one-measurement jump).
|
||
|
||
## 2. What I left without a label
|
||
|
||
- 11:30 and 11:35: the image is even less sharp (11:30) and the reference 8013 at times loses part of its light, but `ranges.py` shows no sensor off level on `corr_x` or `corr_y`. No effect on the curves.
|
||
- 10:30 to 10:50: `quality.py --outside` flags 8010 (and, weakly, 101B, 1005, 1007), because of the `dominance` of 8010, high all morning (3 to 4.6 times its calm reference) without change; its `corr_x` stays within 3 noises. It is a usual state at this hour, not an event of the batch.
|
||
- The `error_value` of the reference 8017 above its median from 10:30 to 10:40, and the `corr_y` of 8017 and 1024 a little low at the start of the batch: it is the slow evolution of the morning cycle, already under the thermal label.
|
||
- The thermal movement of the camera is labeled (D-117), even where the compensation absorbs it; so nothing thermal is left aside.
|
||
|
||
## 3. The documentation
|
||
|
||
- **What served me**: the LENZSPITZE variant in "A pass-by hides the targets…" (flux whole, `dominance` up, a second pattern); the turbulence sheet (flux whole, image less sharp); "The sun… behind a spot" (the background of a reference in the sun makes a small common jump); the LENZSPITZE line of section 4; the rule of `families.md` that puts a reference moving away with a sharp edge in Visibility; the bound rules at the edge of the batch.
|
||
- **What I missed**: on a station whose images are taken only hourly, nothing says how to separate an intermittent turbulence from a partial passage when the same sensors lose their sharpness. The quality signature is the same, and the map gives no spatial criterion (e.g. the sensors to the right of the field intact). The map also does not know a reference that *gains* light at once: 8012 brightens at the same hour on several sunny days, and not all.
|
||
- **Apparent contradiction**: a passage is given as "one measurement, sometimes two" that comes back at the next. Here, the passage falls at the turning point of the thermal cycle (the `corr_y` of several targets makes a step that stays). The `corr_y` over the whole day was needed to see that this step is the return from the morning dip, and not a trace of the train.
|
||
- No indication from the expert in conversation, apart from the hint of the request (a train around 11:00), confirmed by the image.
|
||
|
||
## Difficulties
|
||
|
||
- No clear category for a one-measurement jump whose cause hesitates between turbulence and partial passage, without an image; I put it in Visibility by the signature, with the reservation written.
|
||
- The crosses of `thumbnails.py` are slightly offset from the spots on LENZSPITZE (reconstructed frame, `displacement-columns.md`): to keep in mind so as not to read a spot "out of its place".
|
||
|
||
## Sources read
|
||
|
||
- `.claude/skills/lire-les-donnees/SKILL.md`
|
||
- `labeling/label-map.md`: summary, sections 2 to 4, sheets "A pass-by hides the targets for the length of one measurement", "A target left under the threshold latches onto a headlight, a reflection or another spot", "A vehicle or machine that stays in front of the targets", "Atmospheric turbulence", "The sun in the field or behind a spot", "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`
|
||
- `labeling/knowledge/measurement-chain.md`, `displacement-columns.md`, `gaps.md`, `quality-metrics.md`
|
||
- `labeling/auto/LENZSPITZE.jsonl` (neighboring labels and habits of the station)
|
||
- Data: `data/LENZSPITZE/` through `inventory.py`, `overview.py` (batch, day, `--rigid`), `metrics.py`, `quality.py` (`--common`, `--table`, `--figure`, `--labels --outside`), `ranges.py`, `summary.py`, `plot.py`, `thumbnails.py`, the raw CSV of 01.11 (instants), my script `scripts/capteur_jours.py`; images of 01.11 at 10:01, 11:00 and 12:01, of 02.11 and 03.11 at 11:01
|