Technical Field Note #13 · November 2026

The wrong calibration file: how we caught a stereo calibration made for a different sensor mode

Field Note #13 · Diagnosing a stereo calibration made for the wrong sensor mode (ODL-EGO S1 capture rig)

An evaluation package we shipped earlier included a manufacturer stereo calibration file that, on inspection, belonged to a 16:9 recording mode, while the footage was recorded in the rig's 4:3 1600×1200 mode. A data reviewer flagged it, correctly. This note shows how the mismatch shows up and how it was confirmed. Seven file-level checks expose it without any data. On 112,538 ChArUco corners the wrong file gave a 2.40 px median epipolar distance against 0.33 px for the rig's 1600×1200-mode file. On the same real bimanual clip the reviewer saw, the error dropped from 2.82 px to 0.73 px, and the share of matches under 1 px rose from 10% to 61%. Every delivery now names the calibration file, its recording mode and a validation run on that delivery's own frames.

Published November 2026⏱ 12 min readStatus Preliminary · automated analysisOpen data GitHub · Hugging Face · Kaggle
2.40 → 0.33 pxEpipolar distance on board corners, wrong vs right filemedian, 112,538 cornersVerified (automated)
2.82 → 0.73 px (3.9× lower)Epipolar distance on the real work clip11,360 feature matches, 16 framesVerified (automated)
10% → 61%Matches under 1 pxsame matches, both filesVerified (automated)
60.2 vs 65.08 mmBaseline difference between files8.1% depth-scale differenceVerified (file contents)

Why this matters

A calibration file can be correct and still wrong for your footage

Stereo cameras often support several recording modes (resolution, aspect ratio, crop, binning). A calibration is valid only for the mode it was computed in. A file from another mode loads without errors, rectifies without errors and produces depth without errors. It is simply wrong, by several pixels of vertical misalignment and by several percent of metric scale.

We shipped exactly this mistake in an earlier evaluation package. This note documents it as a reusable check: what to look at in the file, what the residuals look like, and how to confirm the right file on the footage itself.

  • Rule we now follow: a calibration file is delivered only together with its recording mode and a residual check on the same delivery's frames.

Setup

Two manufacturer files, one rig, one recording mode

Rig: R8 stereo+IMU, recording mode 1600×1200 per eye at 30 fps, fisheye lenses. Two calibration files were supplied with the hardware.

File A: a stereo JSON (pinhole model with 5 distortion terms, 60.2 mm baseline, no resolution field). This was the file bundled in the earlier package. File B: a Kalibr camchain (equidistant fisheye model, 800×600, 65.08 mm baseline).

Test data: (1) ChArUco board recordings of 2026-10-26 in the 1600×1200 mode, every second sampled frame, 112,538 matched corners; (2) the 120 s real bimanual work clip from the earlier package (2026-10-21), 16 frames spread across the clip, 11,360 SIFT matches after a loose outlier filter that uses no calibration.

Check 1

Seven checks that need no data

Before computing anything, the file itself gives the mismatch away. File A states no resolution, its principal point is 217 px from the centre of a 1600×1200 image and doubles to a ~2004×1044 (16:9) image, and its focal length is 34% longer than File B's. File B states 800×600, exactly half of the recording mode, and its principal point lands 15 px from the centre after scaling.

Table 1. Seven checks that need no data
Check (no data needed)File AFile BWhat it tells you
Resolution stated in the filenot stated800×600 (exactly ½ of 1600×1200)A file without its resolution cannot be trusted for a mode
Principal point (px)(1002, 522)(807, 613) after ×2Should sit near (800, 600) on a 1600×1200 sensor
Distance of principal point from 1600×1200 centre217 px15 px>100 px is a red flag
Image size implied by principal point × 2~2004 x 1044~1615 x 1226A implies a ~16:9 image; the footage is 4:3
Focal length (px)646482A 34% longer focal length means a different crop/scale
Lens modelpinhole + 5 termsequidistant fisheyeA wide-angle lens needs a fisheye model at the edges
Baseline (mm)60.265.08Depth scales with baseline: an 8% difference becomes an 8% depth error

Check 2

Residuals on targets with known geometry

On 112,538 board corners, File A gave a median epipolar distance of 2.40 px (90th percentile 7.89 px) and File B 0.33 px (1.04 px). With intrinsics held fixed, File A also failed to reproduce the board itself: 1.63 px median reprojection against 0.28 px.

The tell-tale pattern is that File A's error is large already at the image centre (2.1 px) and peaks at 400–550 px from it, while File B rises smoothly from 0.25 px toward the fisheye periphery.

Bar chart of median epipolar distance by radius for two calibration files
Figure 1. Median epipolar distance by radius from the image centre, ChArUco recordings, both files.
Table 2. Residuals on targets with known geometry
TestDataFile A (16:9 mode)File B (1600×1200 mode)
Epipolar distance, medianChArUco recordings, 112,538 corners2.40 px0.33 px
Epipolar distance, 90th percentilesame7.89 px1.04 px
Left reprojection, median (intrinsics fixed)same1.63 px0.28 px
Epipolar distance, medianReal bimanual clip, 11,360 feature matches2.82 px0.73 px
Matches under 1 pxsame10%61%

Check 3

Confirming on the real footage, without any target

A buyer usually has no target in the footage. Feature matches between left and right frames are enough: with the right calibration, matched points must lie on the epipolar line. On the real work clip, File A placed the median match 2.82 px off its line, File B 0.73 px; 61% of matches were within 1 px with File B against 10% with File A.

File B is not perfect at the extreme corners of the fisheye image, where both files show large values (Field Note #11 measured the same periphery limit on targets). The comparison that matters is the central field, where the work happens.

Two overlapping histograms of epipolar distance, centred near 0.7 and 2.8 pixels
Figure 2. Epipolar distance of the same 11,360 matches under each file (log scale).
Two heat maps of epipolar distance per image cell for two calibration files
Figure 3. Median epipolar distance per 200 px cell on the real work clip. Blank cells had fewer than 25 matches.

Interpretation

What changed in our process

Verified: File B is consistent with the 1600×1200 recording mode on targets and on real footage; File A is not. Interpretation: File A was produced in a 16:9 mode of the same camera family and should never have been bundled. Its 60.2 mm baseline would also have scaled all depths by about 8% if used.

Process changes: each delivery names the calibration file, the recording mode it belongs to and the device it was made on; includes the seven file checks; and includes an epipolar check on its own frames, so a mismatch cannot pass silently again.

  • Check the resolution field and principal point before using any calibration file.
  • Run a feature-match epipolar check on a few frames of every delivery; a median above ~1 px in the image centre is a stop signal for this rig.
  • Never infer the recording mode from the file name or the supplier's description alone.

SCOPE AND LIMITS

Limitations

What this note does not show, stated up front.

  • Single R8 stereo+IMU unit, single indoor office, single day (2026-10-26); results are not a population estimate for other units.
  • All measurements are automated (OpenCV detection, triangulation, fitting). No independent human re-measurement was performed: AI/algorithm-only preliminary analysis.
  • The factory calibration file is third-party material; it is used for evaluation but is not redistributed. Raw videos are not published (they contain people and a private office).
  • Feature matches on real footage include some false matches; medians and shares under 1 px are reported, not means.
  • The real-clip test uses 16 frames of one 120 s clip; the board test uses one day of recordings.
  • The origin of File A (which mode of which unit) is inferred from its parameters; the manufacturer has not confirmed it.

FAQ

Questions data buyers ask

Q01How do I know whether a stereo calibration matches my recording mode?
Check that the file states the same resolution (or an exact scale of it), that the principal point lies near the image centre, and that left/right feature matches fall within about 1 px of their epipolar lines on your own frames.
Q02What happened in this case?
A calibration made for a 16:9 mode was applied to 1600×1200 footage. On the same real work clip the median epipolar error was 2.82 px with that file and 0.73 px with the rig's 1600×1200-mode calibration.
Q03Can you check calibration without a calibration target in the footage?
Yes. Feature matches between the left and right images are enough to measure epipolar error; a target is needed only to measure metric scale.
Q04Does a wrong calibration file also affect depth scale?
Yes. Here the two files differ in baseline by 8.1% (60.2 vs 65.08 mm), which scales every depth by the same amount.
Q05Is the corrected calibration perfect?
No. It is consistent in the central field and degrades toward the fisheye corners, as measured in Field Notes #09 and #11.

OPEN DATA AND CODE

Reproduce and inspect the evidence

Every number in this note can be checked: derived tables, analysis code, figures and a runnable notebook are public.

GitHubAnalysis code and reproduction stepsorigin-data-lab/stereo-calibration-sensor-mode-mismatch-diagnosisView repository Hugging FaceDataset card, tables and figuresorigin-data-lab/stereo-calibration-sensor-mode-mismatch-diagnosisOpen dataset KaggleData explorer and runnable notebookorigindatalab/odl-stereo-calibration-mismatchExplore on Kaggle
Run the notebook on Kaggle

Licence: CC BY 4.0 (data and figures) · MIT (code)

Cite this note

Origin Data Lab (2026). Technical Field Note #13: Diagnosing a stereo calibration made for the wrong sensor mode. https://origindatalab.io/technical-field-notes/stereo-calibration-sensor-mode-mismatch-diagnosis.html

CONTINUE READING

Related field notes

FIELD NOTE #12ChArUco target design for wide-angle egocentric cameras: measured detection limitsFIELD NOTE #11Independent validation of a fisheye stereo factory calibrationFIELD NOTE #10Timestamp integrity and camera-IMU alignment of a stereo+IMU egocentric rig

Need real-world data measured to this standard?

Origin Data Lab plans and runs egocentric, stereo+IMU and field video capture with documented QA, for physical AI and robotics teams.

Request sample dataContact usData services