Preparing LiDAR–wheel residual inputs once per fit¶
Complete prepared-input LiDAR–wheel evaluation falls from 2.389020 to 0.491396 seconds (4.861703x). The whole fresh comparison worker falls from 5.353725 to 3.397488 seconds (1.575789x). The larger ratio applies to the calibration API; imports and measurement instrumentation dominate the remaining process time. These are ratios of three-run medians, with every default fit, holdout, uncertainty estimate and known-bad control retained.
Input and scope¶
All 2,762 original KITTI OXTS poses of 2011_09_30_drive_0018_sync generate
a sensor trajectory through a declared body-to-sensor transform: XYZ Euler
angles [1, -2, 3] degrees and translation [0.8, -0.3, 0.9] m. The unchanged
two-frame motion construction yields 2,760 intervals in 29 ten-second
blocks. The common original time origin is removed from both sensor and proxy
timestamps, so finite-difference clock steps are resolvable. Samples and motion
intervals are not thinned. All original vf (column 8) and wu (column 22)
values supply the wheel proxy. These are the same columns used by the existing
KITTI wheel-proxy adapter; no synthetic wheel measurements are added.
KITTI has no wheel odometry. This is a generated sensor-motion execution control against an OXTS proxy, without bag reading or raw LiDAR registration. It measures prepared calibration/evaluation performance and numerical parity, without independent accuracy evidence or a GPU claim. The unchanged policy verdict is warn: the 20 ms clock control is not detected, and roll remains unobservable. The reported clock value is not independently validated.
Every third block is held out, retaining 1,824 training and 936 holdout motions before the unchanged 2 m/s gate. There are nine complete solves: one main fit and all eight default block-jackknife fits. Every solve retains all 301 yaw-clock scan candidates and the original six-parameter robust least-squares tolerances, bounds and 200-evaluation maximum. The holdout base, three rotation controls, speed-scale control, clock control and two residual medians remain in the evaluation.
Change and work¶
Each fit prepares the already gated motion velocity/rate arrays and interval midpoints once. An eight-state LRU reuses wheel interpolation and its velocity noise model for exact binary clock states. Signed zero and every numerical derivative step remain distinct. Rotation, matrix products and all six residual rows per motion are recomputed in their original order for every request. There is no residual cache or cache shared between fits. Derived wheel classes retain their per-request interpolation callbacks.
A conservative 64 MiB bound accounts for prepared arrays, all eight clock states and working memory at 1,024 bytes per motion. Larger inputs and empty sets retain the original public residual path. Public signatures, defaults, noise models and core schemas are unchanged.
| Work per complete evaluation | Prior | Prepared |
|---|---|---|
| Main plus jackknife solves | 9 | 9 |
| Residual requests, including holdout checks | 411 | 411 |
| Least-squares results | 9 | 9 |
| Motion selection calls | 420 | 17 |
| Motions visited by selection | 646,689 | 22,968 |
| Wheel interpolation calls, including scans | 3,120 | 2,836 |
| Wheel sample queries, including scans | 4,829,055 | 4,389,516 |
The audit independently reconstructs every fit's original and gated motion indices, the full block split, all clock scan candidates and the eight-state LRU from the complete parameter history. The selected control exercises 13–15 distinct clock states per fit. The main fit retains all 1,743 gated training motions; the other intervals remain represented by the original holdout and gating policy.
Measurement and validation¶
Three fresh frozen pairs alternate execution order with one BLAS thread on an unreserved host. Whole-worker time includes imports, actual Git identity checks, input loading/construction, complete evaluation and trace instrumentation, official artifact saving, trace compression, imported-source hashes and measurement validation. The calibration API boundary starts before prepared input loading and ends after the official result is saved. Control generation, source freezing, raw-input compression and parent audits are outside both boundaries.
| Pair (execution order) | Prior whole s | Prepared whole s |
|---|---|---|
| 1 (prior, prepared) | 5.215874 | 3.444593 |
| 2 (prepared, prior) | 5.353725 | 3.397488 |
| 3 (prior, prepared) | 6.049091 | 3.342003 |
Baseline: 88d6fba240ed38a744bccd023111303f84b7be5b.
Measured source: 479d15b79bf4258bfc8f94175ed63b77d6eccd0a.
Only calibrex/solvers/lidar_wheel_solver.py differs in the complete core
source manifests. All six official artifacts, every requested residual and
each optimizer result match exactly, including full residual/Jacobian byte
digests, all parameter values, active masks, termination fields, costs and
evaluation counters. Only creation time and independently checked actual
source Git identities are excluded from equality. Input and protocol identities
stay in the comparison.
The earlier absolute-Unix-time cohort is excluded because the finite clock
steps were smaller than timestamp spacing. It exercised only two clock states
per fit. Its timings and byte hashes are retained in
qa/cohort-selection.json; both cohorts have identical core source manifests.
The relative-time cohort above is the sole selected speed claim. No measured
runtime was changed after that cohort.
Regression tests cover every-bit residual parity with all six parameters, exact and signed-zero clock keys, eviction, changed inputs between fits, memory fallback and derived-wheel callbacks. A complete default evaluation also compares every optimizer result and analytic covariance with the scalar path. Benchmark and archive tests reject altered histories after rebound hashes, omitted controls, unresolved clocks, changed selection/Jacobian sizes, wrong actual Git IDs, timing/counter errors, raw-input mutations and changed generator source beyond CRLF/LF encoding. See the byte-indexed archive and validation receipts.
All 59 distinct targeted Windows cases pass: three existing solver/CLI, five preparation, 20 benchmark/replay, 16 archive and 15 registry cases. The 44 solver/CLI, preparation, benchmark and archive cases also pass on Linux with different NumPy/SciPy versions. Strict mypy checks all 378 core modules; Ruff, strict MkDocs and unchanged exports of all 129 public schemas plus their ledger pass. The initial Windows archive launch's missing temporary parent and the corrected passing run are retained with their logs.
Original packets, source files, proxy samples, times and recorded numerical
payloads remain exact. Only offline motion regeneration permits platform math
roundoff: rtol=0, 2e-7 m/s for velocity and 1e-12 rad/s for angular rate.
These tolerances never apply to recorded results, residuals or Jacobians.
KITTI data and generated derivatives retain CC BY-NC-SA 3.0, attributed to Andreas Geiger, Philip Lenz, Christoph Stiller and Raquel Urtasun: KITTI, data license. Source code retains Apache-2.0. This control is outside accuracy claims and leaderboard cohorts.
Reproduce¶
python -m tools.benchmark_lidar_wheel_preparation \
--audit-only docs/assets/lidar_wheel_evaluation_preparation
For remeasurement, extract only regular files from raw-oxts.tar.gz into a
fresh directory and rename download-provenance.json to provenance.json:
python tools/benchmark_lidar_wheel_preparation.py \
--baseline-commit 88d6fba240ed38a744bccd023111303f84b7be5b \
--problem docs/assets/lidar_wheel_evaluation_preparation/problem.json \
--oxts-dir outputs/wheel-remeasurement-oxts \
--output-dir outputs/wheel-remeasurement --repetitions 3
The pinned generator documents the complete transform and common time origin.
To generate a fresh control from the original packets, run
tools/prepare_lidar_wheel_control.py --oxts-dir PATH --output NEW.json.