Files
aimsight-labeling/auto/analyses/FLORIDA_2026-10-11T0945_2026-10-22T0315/analysis.md
T

45 lines
4.2 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: FLORIDA, 2026-10-11T09:45Z → 2026-10-22T03:15Z (UTC)
Verification of the 23 existing labels: 23 kept, 0 corrected, 0 removed, 5 added. The work was cut short by the usage limit: the checks listed under "Not checked" were not done.
Weather: there is no `weather/` folder for FLORIDA, so no weather is available.
## Verdict per old id
- FLORIDA-23 kept: r1_r is rejected at 10:00 and there is no pose at that instant (inventory).
- FLORIDA-24, 25, 26, 28, 31, 33, 36, 38, 44 kept. In the hourly pose (metrics.py), pan, tilt and scale follow the housing temperature. They start and come back within one or two measurements of the bounds (pan near -0.033 at night and -0.05 to -0.056 at the peak; scale from 1.0005 to 1.0015).
- FLORIDA-41 kept, with a doubt. In the hourly pose, the tilt starts falling and the scale rising from 08:00, before the gap at 09:00 (tilt 0.0098 at 07:00, 0.0057 at 08:00, 0.0026 at 10:00). The cycle may therefore start around 07:45 rather than 10:45. This was not checked at the native step.
- FLORIDA-27, 29, 32, 34, 35, 40, 42 kept: their bounds match the gaps listed by the inventory exactly. The claims about other stations were not re-checked.
- FLORIDA-30 kept (not re-checked at the native step).
- FLORIDA-37 and FLORIDA-39 kept: they cover all the rejections of p1_r, r1_r, Sellette 4, Sellette 5 and Pont Ernen on the 18th and 19th, and the missing poses.
- FLORIDA-43 kept: Sellette 5 is rejected at 05:30 and 05:45 and is back at 06:00.
- FLORIDA-45 kept: the last rejections are r1_r at 18:15 and Pont Ernen at 18:00. From 18:30 the pose is back, at the cold level. After that, Sellette 4's corr_x stays a little lower (about 0.075 against 0.11 the night before), which is one or two noise widths: not labeled.
- Spanning FLORIDA-1 and FLORIDA-14: consistent with the rejections of Sellette 6 and Massif 2 in the range (inventory).
## 1. My labels (added)
### 2026-10-12T00:30Z → 01:00Z, 01:45Z → 03:00Z, 03:45Z → 04:30Z, 06:15Z → 06:45Z, 22:15Z → 22:45Z
- **Phenomenon of the catalogue**: closest is "a target latches onto another spot", here applied to a reference; no exact match.
- **Precedents**: JUNGFRAU-7 and FLORIDA-51 (a reference poorly seen at night distorts the pose: Visibility). Not followed, because here p1_r's error_value does not rise.
- **Family**: Obstruction, because the cause is a local point beside one spot, not the whole scene. The nature of the point is not established: medium confidence.
- **Measured**: the jumps are at 00:45, 02:00, 02:45, 04:00 and 06:30 on the 12th, then 22:30. Sellette 4 corr_x rises by about 0.15 to 0.2, which is 5 to 9 noise widths. In the rigid fit on the targets (`overview.py --rigid` from 00:30), p1_r moves by about -0.1 to -0.19 px in dx while the targets stay still. p1_r's error_value stays flat (0.38 to 0.46).
- **Images**: the fixed-frame thumbnails at 01:00, 02:00, 03:00 and 04:00 show a small bright point at the upper left of p1_r at 01:00 and 03:00 (no jump), absent at 02:00 and 04:00 (jump). Copied to `p1_r_point.png`.
- **Confidence**: medium.
## 2. What I left without a label
- Isolated deviations of 3 to 4 noise widths (14/10 21:00, 21/10 21:15–22:45, and a few others): too small.
- The night of 21–22: the pose sits at a colder level that matches the temperature (tilt +0.02 and scale 1.0002 at 16–17 °C, the same relation as on the other nights). The corrected curves are flat apart from the small offset of Sellette 4 noted above.
## 3. The documentation
- The `--rigid` mode of `overview.py` was decisive: it showed that the reference moves, not the targets.
- Outside the range: FLORIDA-46 says "the camera does not move", but at 03:30 on the 22nd the pose jumps and Sellette 5 jumps for one measurement too.
## Not checked (cut short)
- FLORIDA-41 start at the native step.
- FLORIDA-30 at the native step.
- The other stations for the No data labels.
## Difficulties
- The usage limit cut the verification short.
## Sources read
instructions/analyst.md, tools/GUIDE.md, tools/DATA.md, labeling/families.md, labeling-rules.md, reading-rules.md, stations.md, labeling/auto/FLORIDA.jsonl (labels around the range), precedents.py, inventory, overview (with rigid), metrics, days, ranges, scripts/native.py, scripts/uncovered.py, thumbnails of 12/10.