Skip to content

Measured browser demos

These browser panels use real recorded observations, not generated sensor streams. The gallery now has eleven selectable demos. Data loads only when a panel is selected. Drag a 3D plot, zoom, or use the observation slider to inspect its numerical coordinates.

Camera–LiDAR method comparison

The ACFR poses.csv contains 40 real checkerboard captures. Four native methods use the same 32 fit and 8 held-out captures: plane-only, point+plane, Horn points, and line+plane. The plot connects measured camera centres with transformed LiDAR centres. A common held-out centre RMSE is displayed alongside the method's own result; it does not replace each method's different objective or observability checks. There is no accepted independent extrinsic truth.

The line+plane boundary extractor resolves corner symmetry using the fit-only plane rotation; it never uses holdout corner residuals to choose the rotation. Its shared split is checked by the producer.

Pandey mutual-information and Levinson–Thrun edge results are shown separately from existing real A2D2 reports. They retain their FAIL / WARN grades and pre-registration caveats. Different datasets and objectives are not ranked against board-centre RMSE. A2D2's camera-registered points do not establish independent extrinsic accuracy.

Robot hand–eye

Eight native AX=XB methods use the same ETHZ robot-arm motion split. Relative motions are constructed from disjoint absolute-pose blocks before fitting or evaluation. The two point sets are the origins of A X and X B for each held-out motion equation; they are not two global trajectories. Rotation/translation closure measures consistency, not absolute accuracy. Source poses and aligned timestamps are preserved in the original archive.

GNSS tracks and candidate lever geometry

Actual QCOM fixes, u-blox fixes and fused camera poses from the same comma2k19 minute are transformed into a local East/North tangent plane. Height is excluded because altitude conventions may differ. Fusion incorporates GNSS, so the fused trajectory is not independent truth. Receiver sample rates and timestamps differ; equal plotted indices are not treated as simultaneous correspondences.

The forward/right lever sliders draw a candidate, geometrically shifted body path using recorded orientations. They do not estimate a lever arm or assign calibration accuracy. Original GNSS and pose data stay unchanged.

Wheels / INS / vehicle axes

All four wheel-speed channels come from actual CAN wheel encoders. They are compared with the recorded fused camera forward velocity interpolated onto wheel timestamps. A separate view displays forward/right/down velocity axes. Differences can combine clock, wheel scale, turns and camera-to-vehicle alignment; this plot does not identify these effects separately.

Upstream utils/camera.py defines the recorded Hamilton wxyz quaternion as ECEF-from-device. Its transpose maps ECEF velocities into camera FRD; the producer follows that implementation. The prose README describes the direction inconsistently, so the upstream helper and an independent axis regression are preserved as evidence. Camera FRD is not asserted to equal the vehicle frame.

Radar tracks

The four snapshots contain real CAN-reported forward range, left distance and relative speed. Colour represents that relative-speed channel. The yaw slider rotates the target coordinates as a manual extrinsic experiment. These are processed Radar tracks, not raw Doppler samples, a native ego-velocity fit, or independently calibrated Radar extrinsics.

Live calibration

The browser runs an actual six-parameter point least-squares fit on ACFR measured centres. The starting transform is the native fit-only Horn result with a deliberate rotation/translation perturbation. Finite-difference Gauss–Newton updates, damping, a loss-reducing line search and termination checks run locally; the UI displays each actual iteration and residual history.

Only the 32 fit observations enter the updates. The 8 holdout observations are evaluated separately. Tests perturb held-out targets and confirm that the fitted parameters stay unchanged. Run, single-step, stop and reset are explicit actions; switching panels stops the loop. No interpolated replay is labelled as an optimizer iteration. This small browser example is separate from the full native calibration pipeline; closure is not an independent accuracy claim.

Observation coverage and injected outliers

The live panel has three reproducible conditions. Full measured coverage uses all 32 original fit captures. Limited coverage keeps only the two distinct fit centres with the smallest separation; the other 30 fit captures are excluded and the original eight holdouts remain unchanged. This models limited geometric coverage, not a literal straight-driving or IMU motion experiment. Two point correspondences leave rotation about their joining line unobservable. The undamped, centred rotational information block detects rank 5/6, and the browser blocks updates instead of letting damping produce a misleading fitted pose.

Inject 4 outliers copies the target array and shifts every eighth fit target by (+0.60, -0.40, +0.50) m. The four changed targets are marked with red crosses. Original source measurements and all eight holdouts stay intact. Both least squares and Huber (0.05 m vector-residual threshold) start from the same perturbed native reference, which was computed from the clean 32-capture fit. This is a controlled robustness experiment, not a blind outlier benchmark.

Huber weights use the 3D residual norm for each observation; a loss-reducing line search evaluates the same robust objective. The chart continues to show raw centre RMSE for fit and holdout, so the raw fit curve need not decrease under Huber. Numerical convergence, geometric rank and held-out residuals are distinct diagnostics. None is an absolute accuracy or covariance estimate. Resetting or changing either control stops a running fit and creates a fresh history.

Sources, attribution and reproducibility

  • ACFR: pinned measured VLP observations, Apache-2.0. Original CSV SHA-256: 024bc6ed9009652761e9c0df49b106d325a10c88d41a80e9b36ea55fd567e110.
  • ETHZ ASL: pinned real robot-arm archive, DOI 10.3929/ethz-c-000788527, In Copyright – Non-Commercial Use Permitted. Upstream sample code is BSD-3-Clause; that does not relicense dataset rights.
  • comma.ai: comma2k19 example segment, Harald Schafer, Eder Santana, Andrew Haden and Riccardo Biasini (2018). Original repository MIT notice is preserved in source/comma/LICENSE. source/comma/receipt.json binds every NumPy file and helper to its pinned URL and SHA-256. No pickled arrays are loaded.
  • A2D2: displayed targetless report metrics retain their existing dataset conditions and source reports. Raw A2D2 images or point clouds are not copied into this bundle; see the existing report provenance.

Rebuild with PYTHONPATH=src:tools python tools/build_evidence_demos.py. evidence_demo_index.json binds all original observations, producer, native methods and display assets to SHA-256. The source archive uses no generated Doppler or prepared synthetic GNSS paths.