Files
aimsight-labeling/auto/analyses/BREITHORN_2026-10-03T0000_2026-11-01T0000/analysis.md
T
2026-11-18 10:51:31 +01:00

3.0 KiB
Raw Blame History

Requalification of the long labels: BREITHORN

Measured with tools/scripts/ranges.py --columns corr_x,corr_y (k 5, merge 12h) on each label's range, with level = median of the range and noise = the sensor's own noise (work/ranges_.txt, work/ranges_all_.txt). Parallax: scripts/parallax.py and scripts/parallax_fit.py (work/parallax.txt, work/parallax_fit.txt).

  • BREITHORN-2: corrected (parallax). The drift goes beyond the threshold: on corr_x, POUI19 reaches +11.1 noises (12/10 12:00, about +0.3 mm) and COLJ19 +7.7 noises. The description now gives the parallax as the deduced cause and names POUI22 (opposite in corr_y).
  • BREITHORN-13: corrected (parallax, plus the period analysts' correction). The drift goes beyond the threshold: on corr_x, COLH20 reaches +12.6 noises (about +0.8 mm), COLJ19 +8.2 and POUI19 +7.3. COLI22, COLH20''' and POUI24 are named, as are COLI24 and POUI22 (opposite in corr_y).

Parallax (asked by the user, 2026-11-18)

targets.csv gives the targets' distances: COLJ19 8.9 m and POUI19 9.4 m; COLI20, POUI20, COLH20 and COLH20''' 20–23 m; POUH21 and COLI21 30 m; COLI22 and POUI22 40 m; COLI24 and POUI24 60 m. The references are at 31–50 m (mean of 1/d ≈ 1/40 m). If the camera shifts sideways by T, a point moves by f·T/d in the image. The similarity fitted on the references removes f·T/d_ref, so each target keeps a remainder ∝ p = 1/d − 1/d_ref:

  • a large remainder on the near targets;
  • none at the references' distance;
  • a remainder of the opposite sign beyond it.

Change of corr from the start to the end of each range (median of the first 6 accepted measurements against the last 6), regressed on p:

  • 03/10 → 13/10: corr_x r = +0.82, corr_y r = +0.53.
  • 13/10 → 16/10 (pause and slight return, no label): corr_x r = −0.86, the same parallax with the opposite sign.
  • 16/10 → 01/11: corr_x r = +0.78, corr_y r = +0.27.

Fitted model: a lateral shift (tx, ty) plus a shift along the axis (radial from the image centre), each × p. It explains 70 %, 67 % and 61 % of the drift of the targets in the three ranges (the lateral shift alone: 60 %, 66 %, 54 %). With f ≈ d / mm_per_px ≈ 2700 px, the camera would shift laterally by about 0.7 mm (x) and 0.4 mm (y) up to 13/10, and about 0.9 mm and 0.5 mm from 17/10 on. This is a rough model.

What it explains: COLJ19 and POUI19 drifting the most, the 20 m group drifting less, POUH21 (30 m, ≈ d_ref) not drifting, the far targets turning the other way, and the sign flip during the pause.

What it does not explain (remainders of several noises):

  • COLI21 (30 m): +0.06/+0.04 px up to 13/10, +0.075/+0.11 px from 17/10, about 7–10 noises.
  • COLH20''' in corr_x: the opposite sign of what the model gives.
  • The corr_y of COLI24 and POUI22: the right sign, about four to five times stronger than the model.
  • COLI20 in corr_y after 17/10.

These stay "not established" in the descriptions. A remainder of the rotation that depends on the position in the image is possible, but I did not test it.