Explore the real-data demos¶
Open the browser gallery and select a method tab. Eleven panels now cover the original five sensor-pair demos and six added explorations.
| Panel | What to try | Data and interpretation |
|---|---|---|
| LiDAR–LiDAR | Orbit the point clouds and slide the comparison | Saved NTU VIRAL alignment |
| IMU–LiDAR | Play or stop the deskew recording | Real MID360 sweep |
| Camera–IMU | Compare measured motion windows, explore rotation × time and inspect residuals | Measured Hilti camera/gyro series; saved baseline, fixed gyro bias |
| Camera–LiDAR | Change six offsets and compare the projections | Real KITTI returns and photographs; manual experiment |
| Fisheye stereo | Keep the photographed pair while switching models; orbit 3D and inspect matches | Real images; supplied KB4 and newly fitted MEI/DS; no depth truth |
| Method comparison | Inspect the shared-capture residual matrix and switch four native board methods | Shared ACFR fit/holdout split; separate targetless A2D2 evidence |
| Robot hand–eye | Switch eight methods and inspect motion closure | Measured ETHZ poses; closure is not absolute accuracy |
| GNSS tracks | Zoom receiver tracks and change a candidate lever arm | Measured comma2k19 fixes; fused path is not truth |
| Wheels / INS | Select each wheel or the body-axis view | Measured wheel speeds and fused camera FRD velocity |
| Radar tracks | Select a snapshot and change candidate yaw | Real processed CAN targets; reported relative speed, not raw Doppler |
| Live calibration | Run, step, stop or reset a six-parameter fit | Actual browser updates on measured ACFR centres; holdout excluded from fitting |
In Live calibration, switch Observation condition between full coverage, two measured captures, and four explicitly injected fit outliers. Compare Least squares with Huber, then run or single-step the fit. The rank warning blocks underconstrained geometry; the unchanged holdout curve shows how outlier bias differs from numerical convergence. These are controlled changes to a real board dataset, not simulated sensor recordings or a literal vehicle-motion test.
In Camera–IMU, the time slider changes which IMU samples are averaged over the fixed camera intervals, while the rotation slider injects one camera-frame rotation around the saved estimate. The displayed offset convention stays consistent across the range. Choose all 59 observations or either of two non-overlapping 20-observation windows selected by camera-rate RMS. Their side-by-side heatmaps share a colour scale; switching windows preserves the injected parameters. Higher RMS does not guarantee calibration accuracy. The 3D residual heatmap, time profile, selected-axis curves and numerical interval inspection update together. The near-minimum contour is not a statistical confidence region, and the grid minimum is not asserted to be calibration truth. See the clock experiment guide for integration, sign convention, source records and interpretation.
The source guide describes coordinate conventions, licenses, reproducible producers and the limits of each view. Each tab clearly distinguishes saved results, manual geometry and actual live optimization. The backlog records the scope delivered for the seven requested additions.
In Fisheye stereo, click a sparse depth dot or use the match slider and previous/next buttons. Both rectified display crops update together at a responsive magnification, with pixel coordinates, an 11×11 neighbourhood outline and the saved correspondence at the cross. The 41×41 context helps inspect repeated patterns. Saved epipolar guides remain visible; this does not recompute ZNCC or certify the match. Missing border context is padded rather than copied from the other camera. Estimated depth has no independent ground-truth validation.
The stereo 3D view back-projects each saved match into the rectified left
camera using X = (u − (width−1)/2) Z / f, Y = (v − (height−1)/2) Z / f and
the existing Z = fB/d. Coordinates are metres, X right, Y down, Z forward;
right-image vertical mismatch remains a separate diagnostic. The metric grid
lies on the camera Y=0 plane; it is a coordinate guide, not an inferred floor. Drag or use orbit
sliders to rotate the orthographic view. Click a point to select its image
match; the match slider and previous/next buttons provide keyboard access.
The default display includes Z ≤ 1 m. Choose Z ≤ 3 m or All estimated depths to include farther estimates; the count always reports points outside the display range. This is a display filter, not an outlier rejection or quality claim. Focus selected match reveals a selected point outside the range and centres the view on it. Reset 3D view fits the current depth range. Colours clip at 0.3–3 m, while coordinates keep their full values. The view shows sparse estimated matches, not a dense reconstruction or independently validated depth.
All KB4 / MEI / Double Sphere models now support pairs 000, 014 and 027.
Changing a model keeps the photographed pair selected; the saved rectification,
match set, 3D view and close-ups update together. Raw photographs stay identical
for that pair. Match counts and model-specific match sets do not establish a
model accuracy ranking; there is no independent depth truth. The fit-level
excluded-capture reprojection diagnostic is shared across that model’s pairs.
Enable Compare all three models side by side to load the current photograph for KB4, MEI and Double Sphere together. Each card shows the saved rectified left image, accepted-match count and median absolute vertical mismatch. Inspect selects that model in the linked image/3D inspector while keeping the photograph.
Histograms share 0–3 px horizontal and 0–100% vertical scales, with equal 0.25 px bins and a denominator equal to that model’s own accepted-match count. Matches at ≥3 px are counted separately. The models generally accept different points; these plots do not rank independent depth accuracy. Raw photographs and output focal length are checked for consistency. Images and guides are saved results. The comparison loads only when enabled, clears old cards when changing pairs, and offers a separate retry if any model or image cannot load.
In Method comparison, the ACFR matrix shows each of four native methods on the same eight held-out board captures, with one shared millimetre colour scale. Click a cell or use arrow keys to select its method and capture in the 3D plot. The observation slider and method selector preserve the selected capture; orbit and zoom stay aligned across methods. Matrix rows use the same 32-fit / 8-holdout split, checked against capture IDs and camera centres in the saved source.
Cells show the norm of transformed-LiDAR minus camera centre, with signed XYZ components in the selected-cell readout. Row RMS is recomputed from displayed coordinates rounded to 1e-7 m; the side panel retains the saved native evaluation metric, formatted for display. This is a shared held-out consistency diagnostic, not independent extrinsic accuracy. A2D2 targetless objectives remain a separate dataset and metric. On narrow screens, scroll the matrix horizontally.