Files
aimsight-labeling/reading-rules.md
T

12 KiB

Reading rules

What the data of a station mean: where each cause enters the measurement, what it touches, and the rules and false friends that the expert and the reviewed labels have established. What the files and columns are: tools/DATA.md. How to place a label: labeling-rules.md; the families and their boundaries: families.md, whose rules are not repeated here.

Cases are cited by label id (NORDEND-80, in auto/<STATION>.jsonl; precedents.py finds others); [expert]: the expert's word, from a label that is not in auto/. Maintained by hand.

1. How a measurement is made

  1. The scene: targets on the structure, references around it, assumed fixed.
  2. The view: the air between the camera and the targets, and what can pass or settle in front.
  3. The image: the lens, then the capture (IRON, or DIFF which combines several captures).
  4. The pattern measurement: each sensor is found in the image: x, y, error_value; on recent cameras, the image quality. Above error_threshold, the measurement is rejected.
  5. The compensation: a 2D similarity (translation, rotation, scale) fitted on the references (global_ref) gives the camera pose (pan, tilt, rotation, scale, sigma0) and, per reference, reproj_err and inlier.
  6. The corrected curve: corr_x, corr_y, the position of the target once the estimated camera movement is removed.

A displayed value is the end of this chain: find the stage that gave way.

2. Where each cause enters, and what it touches

Stage Cause What it touches Time Witnesses
View (air) fog, rain, condensation: reduced visibility a homogeneous drop in flux; error_value rises, rejections; raw positions without a step hours quality (flux), images
View (air) heat haze (turbulence) a fast noise; only when the structure really heats up; not under a bridge the hot hours temperature, error_value by day versus night
View (obstacle) object, leaves, snow on a target: occultation; a passage one or a few targets; on a recent camera, their flux cut, the neighbours intact (LENZSPITZE-9); a passage: one measurement from one measurement to weeks quality, images
Target problem target (badly oriented, dirty, glued) one target, lastingly days and more error_value of the target alone
Lens thermal deformation of the optics (sun on the station) the whole image deforms, which a similarity cannot represent short: follows the sun on the station housing temperature
Camera and mount expansion of the mount; wind, jolt, shock raw positions of all sensors, references included, together; the compensation absorbs it (corr flat) or badly: corr follows the inflections of the camera movement specific to the mount camera pose, housing temperature
References reference hidden or less well seen the transformation is redone without it: points shifted at once, often grouped in one part of the image, the shift growing with the distance to the references; back when it returns: reversible the time of the nuisance error_value, reproj_err of the references, sigma0, camera pose
References a reference moved the corr of the targets shifted lastingly lasting reproj_err of the reference
Structure thermal expansion of the structure corr, an amplitude specific to each target, references stable the longest: slow heating, late cooling no air temperature (section 3)
Structure real movement corr of one or several targets, which lasts; references and pose stable, good image quality lasting all of the above, at calm

A jump that returns to normal is not serious: the return is what makes it benign. Saying "jump" is not enough; the question is its cause, quite often a reference hidden or less well seen.

3. Thermal effects

  • Three movement effects: the lens, the mount, the structure, told apart by their heating and cooling times; the structure is the slowest. Heat haze is apart: a noise, not a movement.
  • The cases cross: a camera on a bridge has two structures, the one that carries it (the mount) and the one it monitors.
  • In waves: expansion, then contraction; often stronger in Y than in X, depending on the orientation.
  • The kind of structure says whether a thermal movement is possible: a bridge yes; a rock face, too massive, or a road, no: the movement is then that of the camera (stations.md). [expert, WILDSTRUBEL-102]
  • The camera can move through a thermal effect at a different rhythm from the structure, not having the same heating time. [expert]
  • The temperature that counts is that of the camera housing, not that of the air around, even sheltered. The temperatures available are those of the compute board (sensor_*_deg_c), inside the housing, heated by the sun and by the computing, as hourly means: a witness for the lens and the mount, not for the structure nor the air. temperature_deg_c, on a few stations: unknown sensor. [expert, BREITHORN-14]
  • A change of regime is understood through the temperature: look at it. [expert]
  • The direction of a thermal movement is read on the housing temperature, not on the hour; when the scale of the camera follows it, the effect is on the camera. [NORDEND-155, CLARIDEN-13]

4. Gaps

  • A gap is not necessarily an outage: acquisition, transmission, ingestion and display are distinct layers.
    • At the end of a file, empty values can be only a synchronization delay (last_sync): a delay only if the other stations do not have their data yet either.
    • At the start of a file, empty values precede the first measurement: not an outage.
    • A gap surrounded by a normal voltage is not an energy problem.
  • A gap is read with the other stations and the metrics (the boundary "A gap" of families.md). [JUNGFRAU-52, NORDEND-50, WILDSTRUBEL-201, CLARIDEN-19]
  • A power cut leaves a signature: the voltage below the knee (about 12.8 V) just before or just after the gap; the temperature drops during the gap more than at the same hours on nights without a gap, which alone does not settle it (a common gap can do the same, an outage that comes back after sunrise shows nothing).
  • A station on mains that comes back with its battery drawn down has lost the mains.
  • A NaN or inf value in an export is an absence of measurement.

5. Reading rules

  • What moves in the image is read before the cause, on the raw positions against the overall movement (overview.py --rigid): the targets do not move relative to each other, it is the camera [BREITHORN-2]; targets move alone, a movement is not excluded [CLARIDEN-14]; nothing moves and the curves jump, it is the base or the configuration [WILDSTRUBEL-136, CLARIDEN-38].
  • A jump is judged against the neighbours: if it brings the target back toward them, it is a return to normal, and the artifact was before; if it takes it away from them, it is to be treated quite differently. [expert]
  • A level is also judged against the past of the target, over weeks: a jump that brings it back to an old level is a return. [NORDEND-94, NORDEND-169]
  • A high error_value is not a problem in itself: its variations count. [expert]
  • A very stable error_value says the targets stay well visible: while they move, it rather suggests a poor compensation than an obstruction of the view. [expert]
  • A rejected measurement is removed from the display, not a measurement just under the threshold: a target left under the threshold keeps its jump. [expert, NIEDERHORN-13]
  • At least three references are needed; dropping a single one is enough to distort everything: a kind of triangulation. The position of the targets rests on the network of references: references badly seen, targets badly anchored. [expert]
  • The reprojection error says how many references hold: at zero, only two remain, a very bad sign, not a perfect measurement; absent, only one, and no measurement is possible. Except on a station built with two references, where it is always at zero. [JUNGFRAU-22, FLORIDA-4, expert]
  • The pose of the camera is reliable only if the references hold: a reference rejected, dropped or sliding against the targets distorts it; a single hindered target does not. [FLORIDA-58]
  • A drift specific to a few targets that follows the camera, on the same dates, is the scene deforming with it, not a second movement; it is separate only if it has its own dates. [NEBRASKA-6, NEBRASKA-5]
  • Neighbouring cameras are read together: the same movement, the same loss or the same gap on the cameras of one building have a common cause; what touches only one is on its line of sight. [MICHIGAN-3, BREITHORN-20]

6. False friends

  • Jumps of many targets, all in the same direction, are not a movement when they fall at the reference switches: raw positions without a step, daily cycle that continues. [expert]
  • A target that slides slowly away from its neighbours is not moving when its error_value rises with the sliding and a step brings it back toward them. [expert]
  • A rise over several days is not necessarily a real movement: it may follow a lasting drop in temperature. [expert]
  • Corrections stuck at zero are not a failure: a twin of a reference. [expert]
  • No signal does not mean no obstruction: only the images show the web. [expert]
  • A web is not the spider: the web veils in part for hours or days; the spider hides everything for the length of one measurement. [expert]
  • References that do not move in the rain do not contradict the rain: higher up, they suffer it less. [NIEDERHORN-12]
  • The step of a real movement is not its amplitude when the camera and the references have moved too. [expert]
  • A swarm of insects is not a train: same signature; only the image tells them apart. [NIEDERHORN-66]
  • In a tunnel, a temperature drop at the same moment does not make a thermal effect: the bump of a group comes with the error_value and returns before the scale of the camera. [CLARIDEN-12, CLARIDEN-13]
  • The return of the error_value is not the end of an obstruction: after a mark on the glass, the curves stay shifted. [NORDEND-169]
  • A spot on a thumbnail that follows the measurement is not necessarily there: the frame follows a measurement perhaps latched onto the neighbour; only a fixed frame shows it. [FLORIDA-43]

7. Non-negotiable rules

  • Judge nothing on a binary inlier: use continuous weights, residuals, influence.
  • Neither sigma0 alone nor the number of references is ever a sufficient criterion.
  • The image quality of a target is not the quality of the reference network; a small residual is not a well-constrained transformation.
  • A common-mode movement is not a structural movement; coherent references are not physically fixed references: the camera cannot prove that a network moving as a block is fixed.

8. Lessons from cases outside the labels

  • When everything moves together, the measurements alone do not separate the structure from the camera (Barcelona, a bridge whose vault subsided: the references, the camera and the structure moved; the expert called it a real movement, confirmed by other monitoring systems of the site). An unstable compensation then distorts the amplitude of a real movement: the movement is reported, its amplitude is not established. [expert]
  • Two prisms too close interfere: a 24 h cycle correlated between the two, or isolated peaks on one target at a time, too large for the structure, with no obstruction on the images; a neighbour much brighter than the others. On a recent camera, dominance shows it. [client report]
  • Vegetation grows over days and weeks: a partial, progressive occultation, not reversible from one measurement to the next like a passage; the hardest obstruction according to the expert. [client report]
  • References masked (two of four, by vegetation): a 24 h cycle on pan, roll and tilt, badly compensated, is then not a movement of the camera but a base too weak for the transformation. [client report]
  • A target of low contrast: by day, square waves on one axis, back at night, error_value rising by day without reaching the threshold. [client report]
  • Snow: noise and apparent movements on all sensors and the camera, a brief loss of data, the signal quickly back. [client report]