Real fisheye stereo photographs

The browser's Fisheye stereo demo uses real indoor photographs, not rendered images. Three pairs (000, 014, 027) come from OpenCV extra's testdata/cv/cameracalibration/fisheye/calib-3_stereo_from_JY dataset.

Attribution and source records

Credit: OpenCV contributors / calib-3_stereo_from_JY. Original image bytes, the calibration test source (including its upstream license notice), and the board-coordinate file are preserved in source/. These third-party materials retain their upstream terms; Calibrex does not relicense them under its own Apache license. OpenCV extra supplies no separate image-specific license file. See OpenCV contribution terms and the original test data.

sources.json binds each original file to an immutable upstream URL and SHA-256. calibration.json transcribes the distinct left and right intrinsics, distortion, and right-from-left transform from TEST_F(fisheyeTest, stereoCalibrate) in the preserved OpenCV test source.

Processing and limits

Run PYTHONPATH=src python tools/build_real_stereo_demo.py from the repository root. The generator resizes the original 1280×800 images to 640×400, transforms intrinsics with the same pixel-centre convention, rectifies through Calibrex's native camera implementations, and runs its image-only ZNCC patch matcher. RGB is preserved for display; matching uses grayscale. Board corner coordinates and known object geometry do not guide matching. There is no rendered scene.

The baseline is approximately 99.4 mm, following the reference translation and upstream board-coordinate scale (0.0244, interpreted as metres). Displayed Z uses that calibrated scale: Z = f × B / disparity. It is depth along the rectified left camera's Z axis, not ray distance.

There is no independent depth truth or depth-accuracy score. Vertical mismatch is a correspondence diagnostic, not a distance error. Repetitive checkerboard patterns can match the wrong square; occlusions and weak texture also produce errors. Integer disparities and outliers are retained. No uniqueness or left/right consistency filter is applied. Colours clip to 0.3–3.0 m without changing the displayed point measurements.

These photographs belong to the calibration dataset used to obtain the supplied reference parameters. This is a saved demonstration, not a held-out benchmark or a live calibration run. KB4 uses the upstream calibration. MEI and Double Sphere are newly fit to the measured board corners, and their real-image rectification/depth demos use captures 000, 014 and 027. Switching models preserves the selected photographed pair, so all nine combinations use the same three raw image pairs. Each model retains its own fitted transform and baseline. Match sets are recomputed per model and need not identify the same scene points; their counts are not an accuracy ranking. The displayed fit-excluded reprojection value is an aggregate fit diagnostic, not a metric computed for the selected photograph. They are not alternative labels on the KB4 projection.

Run PYTHONPATH=src python tools/fit_real_fisheye_models.py to regenerate these model calibrations. The native projectors fit 27 captures jointly with board poses and the stereo transform. Every fifth capture is excluded from the final fit; those poses are estimated from left corners only before evaluating right reprojection. Full fitted board poses and split IDs are retained in calibration_mei.json and calibration_ds.json, so reported residuals can be recomputed from the original 2D/3D observations without rerunning optimization.

The supplied all-capture KB4 calibration initializes intrinsics and board poses. Consequently these fit-excluded diagnostics are not a fully independent benchmark. No metric depth truth or general lens-accuracy claim is made. All models use a common output focal length for the comparison; each uses its own fitted stereo transform and baseline. Raw photographs remain identical.

The generated receipt binds the generator, native implementation, original photographs and calibration to their SHA-256 values. Each frame is fetched only when selected in the browser.

Before converting rectified floating-point intensities to shared display/matching 8-bit pixels, the producer rounds to eight decimal places and clips to 0–255. This removes sub-1e-8 numerical differences near integral intensities across NumPy/SciPy paths, which otherwise change gradient thresholds and seeded patch selection. Camera maps and disparity/depth equations remain unchanged. Tests retain exact pixel and correspondence comparisons against the saved receipts.