Reusing and batching GNSS–LiDAR lever-arm evaluation¶
The complete prepared-input GNSS–LiDAR evaluation falls from 13.871298 to 4.896686 seconds (2.832793x). The phase from prepared input loading through official result saving falls from 11.028193 to 2.085993 seconds (5.286783x). These are ratios of three-run medians on a public-trajectory execution control, with unchanged default fitting, uncertainty, holdout and known-bad controls.
Scope and input¶
The reference is all 2,762 original KITTI OXTS poses from
2011_09_30_drive_0018_sync. The control declares a body-to-LiDAR transform,
an antenna lever arm, a 12 ms GNSS clock shift and seeded GNSS/odometry noise.
It retains the full natural drive in 29 non-overlapping ten-second windows,
using the segmenter's default minimum of 20 poses. Every third window is held
out: 19 training windows and ten holdout windows. No window, pair, robust
iteration, optimizer evaluation, jackknife fit or policy control is removed.
The GNSS measurements and odometry are generated from this public trajectory. This measures prepared-input calibration/evaluation performance and numerical parity. It does not measure bag reading, raw LiDAR registration, GPU execution or independent calibration accuracy. Segment scan counts describe the declared generated pose epochs. The unchanged result remains inconclusive for the lever arm on this mostly planar drive; the observable clock component and the unobservable translation axes retain the same evidence and policy decisions.
Implementation¶
Each solve owns prepared GNSS displacement pairs, noise terms and ordered array indices. An exact binary-clock LRU holds 64 interpolation states and a second LRU holds 256 normalized data residuals. Parameter values and signed zeros are never rounded. Returned cached arrays are read-only; each IRLS pass still recomputes robust weights, variance and anchors. Each jackknife solve owns its own context.
For ordered, disjoint windows, one SciPy SLERP and three NumPy interpolation calls handle all the epochs of a trial clock state. Every query remains inside its original window. An out-of-window observability difference follows the original per-window path and preserves its error behavior. Overlapping, unordered or ambiguous windows retain per-window interpolation. Independent 3x3 Procrustes SVDs are batched; each window's covariance and final matrix products keep their original shapes and reduction order. The offset-dependent pose-noise blend and lever-dependent displacement noise remain in the model.
A conservative 64 MiB estimate covers both LRUs, prepared arrays, the combined interpolator and working arrays. Larger inputs retain the original scalar evaluator. Public APIs, defaults and all core schemas are unchanged.
| Work per complete evaluation | Prior | Reused/batched |
|---|---|---|
| Main plus jackknife solves | 9 | 9 |
| Residual requests, including held-out checks | 1,542 | 1,542 |
| Least-squares results | 36 | 36 |
| Per-window interpolation calls | 26,029 | 138 |
| Combined interpolation calls | 0 | 600 |
| Interpolated epochs, including repeated states | 2,368,419 | 935,169 |
| Procrustes SVD dispatches | 26,000 | 1,258 |
| Window alignments evaluated after residual reuse | 26,000 | 19,532 |
The audit independently replays both LRUs from every recorded parameter state, then checks interpolation batches, every queried epoch and SVD dispatches against the observed counters. Repeated mathematical work is reused; the requested residual sets, optimizer budgets and evaluation evidence are identical.
Measurements and validation¶
Each variant ran in three fresh frozen processes, alternating execution order, with one BLAS thread. The host was not reserved. Whole-process timing includes imports, actual Git checks, input construction, complete evaluation and trace instrumentation, artifact saving, trace compression and imported-source hashes. Control generation, freezing, raw-input compression and external audits are untimed.
| Pair (execution order) | Prior whole s | Reused whole s |
|---|---|---|
| 1 (prior, reused) | 13.770146 | 4.896686 |
| 2 (reused, prior) | 14.099818 | 4.875520 |
| 3 (prior, reused) | 13.871298 | 4.954276 |
Baseline: 01127d3a8e6f830d99786d183dbef6a311e50c14.
Measured source: 1c91f5a5bec5f6324be5ca10f652bf863bc04b02.
Only gnss_lever_arm_solver.py differs in the complete core source manifests.
All six complete public results, every residual parameter/digest and every
least-squares result match exactly, including residual/Jacobian digests,
termination messages, costs and evaluation counters. Comparison excludes only
creation time and the independently verified actual source Git identities.
Input and protocol identities remain compared.
Regression coverage includes clock modes, noise terms, unequal windows, rank-deficient/reflected alignments, cache bounds and signed zeros, mutable inputs between solves, memory fallback, exact endpoints, interval gaps and out-of-window differences. Complete block evaluation and every optimizer result are compared with the original evaluator. Archive mutation checks cover histories beyond rebound hashes, counters, epochs, timing, raw pins and actual Git IDs. Validation logs and the two repaired endpoint fixture attempts are retained in the portable archive. All 76 distinct Windows cases pass: 13 existing solver/evaluation, 18 reuse, 18 benchmark/replay, 12 archive and 15 registry cases. The 36 reuse/benchmark cases and the complete offline archive audit also pass on Linux with a different NumPy/SciPy version. Strict mypy checks all 378 core modules; Ruff, strict MkDocs and unchanged exports of all 129 public schemas plus their ledger also pass.
Raw packet and source bytes are hash-exact. Only offline geometry regeneration
allows platform math differences: rtol=0, 2e-8 m translation and 1e-13
rotation-matrix elements. Epoch timestamps, window IDs and all recorded
result/residual/Jacobian comparisons remain exact.
The saved generator's original bytes remain hash-pinned. Its installed Python
source is compared after normalizing CRLF/LF line endings on both sides, so an
LF-only Git checkout can audit a control generated on Windows. All other source
differences are rejected. Six additional regressions cover this checkout case
and accepted/rejected newline/source combinations.
KITTI data and its generated control retain CC BY-NC-SA 3.0 attribution to Andreas Geiger, Philip Lenz, Christoph Stiller and Raquel Urtasun; see the KITTI dataset and data license. Code retains Apache-2.0. This performance control is outside accuracy claims and leaderboard cohorts.
Reproduce¶
Audit offline from the repository root:
python -m tools.benchmark_gnss_lever_arm_reuse \
--audit-only docs/assets/gnss_lever_arm_evaluation_reuse
For remeasurement, extract only regular files of raw-oxts.tar.gz into a fresh
directory and rename download-provenance.json to provenance.json. Then:
python tools/benchmark_gnss_lever_arm_reuse.py \
--baseline-commit 01127d3a8e6f830d99786d183dbef6a311e50c14 \
--problem docs/assets/gnss_lever_arm_evaluation_reuse/problem.json \
--oxts-dir outputs/gnss-remeasurement-oxts \
--output-dir outputs/gnss-remeasurement --repetitions 3
The pinned control-generator.py.txt documents every transform, cut and noise
draw. To generate a fresh equivalent control from the original packets, run
tools/prepare_gnss_lever_arm_control.py --oxts-dir PATH --output NEW.json.