Indexed reference lookup for joint Radar calibration¶
The complete prepared-input seven-DoF Radar calibration workflow changes from 49.293604 to 11.866233 s (4.154107x). The solver API changes from 45.601982 to 9.320259 s (4.892780x). All numerical output, five optimizer iterations, eligibility decisions, rank diagnostics and fourteen holdout probes match exactly across all six runs. This is automatic CPU acceleration through the existing joint solver and RMSE APIs.
Workload and scope¶
The input is byte-identical to the preceding profiled-Radar control:
all 2,762 OXTS reference samples of public KITTI raw drive
2011_09_30_drive_0018_sync, with 2,758 generated Radar velocities, deterministic
0.0001 m/s sinusoidal noise and unequal weights. Generation uses a declared
17-degree yaw, [0.42, -0.18, 0.27] m lever arm and 0.023 s offset.
The Radar velocities are generated execution controls. These measurements
establish performance and parity, not independent Radar accuracy, covariance
coverage, LiDAR pipeline speed or a SOTA result. No locked accuracy cohort is
opened or updated.
The joint solver starts from an identity SE(3) and zero offset, and fits all six extrinsic components plus time. Its original 3 m/s initial-residual gate retains 2,099 measurements and rejects the same 659 IDs in every run. The unchanged default settings retain an 80-iteration limit, bounded clock, finite-difference Jacobians, Huber LM, grouped train/holdout split, complete observability diagnostics and fourteen signed calibration probes. Each run converges after five iterations and makes 171,502 interpolation calls. No measurement budget or numerical stopping rule is reduced.
Change¶
Each residual/Jacobian evaluation previously searched the reference from the first timestamp. The solver now prepares an immutable, finite, strictly increasing reference tuple and uses binary search to select the same interval. Lookup per evaluation changes from O(N) to O(log N), with O(N) preparation work beyond the existing sort. Existing interpolation arithmetic evaluates the selected pair, including the left interval at exact knots.
Both the solve and public RMSE path use the index. Duplicate or non-finite reference clocks retain the original scan semantics in the RMSE path; solve input validation keeps its existing rejection rules. The index is owned by one call, with no cross-solve cache. Public signatures, numerical operations, factor weighting, results and public schemas retain their existing contracts.
Complete fresh-process measurements¶
Three alternating pairs use frozen source trees, real Git identities, the same interpreter/dependencies and one BLAS thread. The host was unreserved; all timing variation is retained. Whole-process time includes startup, JSON parsing, construction, initial-residual gating, the complete optimizer, interpolation counting, all diagnostics, serialization and saving. Input generation, downloading, freezing and external auditing are untimed.
| Pair (execution order) | Prior whole s | Indexed whole s | Paired ratio |
|---|---|---|---|
| 1 (prior, indexed) | 45.545767 | 11.563787 | 3.939x |
| 2 (indexed, prior) | 53.370424 | 11.866233 | 4.498x |
| 3 (prior, indexed) | 49.293604 | 16.945178 | 2.909x |
The reported 4.154107x is the ratio of the two three-run
medians and applies to this workload and reference length. No slow run is
discarded or relabeled as a warmup. The baseline is
2204c39ffc6f661e44972189ec55007f9e5ab8f0 and measured runtime is
c90d226b30a02c840876f9d897d522633e75653a; only the joint Radar solver changes in the
complete runtime manifests.
Parity, validation and provenance¶
Every run saves its complete result, optimizer history, weighting contracts, accepted/rejected steps, initial rejected IDs, all rank/condition and family diagnostics, holdout values and probes. Numerical values receive no tolerance or normalization. Source commits are recorded separately in the provenance wrapper. Interpolation call counts also match in all six workers.
Regression coverage includes exact floating-point interpolation bytes at knots/nextafter neighbors, endpoints, epoch-sized times, empty/single-sample references, duplicate/non-finite clock fallback, complete weighted optimizer history and initial-rejection parity, and logarithmic reference access. The related suite passes 91 cases; fourteen dedicated joint-archive checks exercise complete parity and tamper rejection. Strict core typing, Ruff, documentation and public schema/ledger export parity are checked separately. The original profiled-Radar archive keeps its identical audit result.
The complete archive contains original input/generator identities, source capsules/manifests, all six workers, full results, timings and validation receipts. The input generator and measured launcher are separately bound so the previous control can be reused byte-for-byte. Derived numeric data retains KITTI attribution and CC BY-NC-SA 3.0 terms separately from Apache-2.0 source code.
Reproduce¶
Audit without calibration, a raw download or a GPU:
python -m tools.benchmark_radar_reference_indexing \
--audit-only docs/assets/joint_radar_reference_indexing
Remeasure using the exact committed control and its original generator:
python -m tools.benchmark_radar_reference_indexing --solver joint \
--baseline-commit 2204c39ffc6f661e44972189ec55007f9e5ab8f0 \
--control docs/assets/joint_radar_reference_indexing/control.json \
--control-generator docs/assets/joint_radar_reference_indexing/control-generator.py.txt \
--output-dir outputs/joint-radar-new --repetitions 3