Reusing timed Camera-LiDAR evaluation work¶
Complete prepared-input seven-DoF Camera-LiDAR refinement changes from 136.523854 to 33.203555 s (4.111724x). The artifact-backed API changes from 134.120710 to 30.923065 s (4.337239x). All six complete results, candidate objectives, accept/reject decisions and final holdout values match exactly. This is automatic CPU acceleration of extrinsic-plus-clock refinement with ordinary dependencies.
Workload and scope¶
The control reuses all 9,216 selected original KITTI points of the
probabilistic refinement control, with 1,024
points from each of nine scans of 2011_09_30_drive_0018_sync: 114-116, 229-231,
and 344-346. Original OXTS linear/angular velocities come from the same scan
indices. Image means are generated using a declared pinhole camera and extrinsic,
a 0.02 s clock offset, uniformly declared -0.1 to 0 s point offsets, fixed
0.05 px sinusoidal noise and the original anisotropic covariance/confidence.
The initial pose has a declared rotation/translation perturbation and zero offset.
Image correspondences and point times are generated execution controls. This establishes performance and numerical parity, not independent calibration accuracy, feature-extraction speed, raw-sensor pipeline speed or a SOTA result. No locked accuracy cohort is opened. All 9,216 input points remain in both variants; seven frames/7,168 points train and two frames/2,048 points hold out. The original 500-evaluation bound, Cauchy loss, confidence and optical gates, covariance, split, sensitivity test and step minima remain unchanged. Both variants converge after exactly 341 candidate evaluations.
Change¶
Point motion compensation depends on the clock candidate and fixed capture inputs. Extrinsic-only candidates can share it. One solve now owns a fixed set of point entries, each with a two-state clock cache. It prepares each original body-frame point, row exposure time and inverse body/LiDAR transform once, using the same constant-twist or piecewise-trajectory primitives and operation order. Binary offset keys retain signed zeros; policy keys retain units and stamp semantics. No time is rounded or interpolated differently.
Projection also prepares the candidate's two quaternion normalizations once; point conversion, both Hamilton products and translation addition retain their original floating-point order. Public APIs, results and schemas are unchanged. The cache belongs to one solve, holds no cross-solve state, and uses a conservative 128 MiB estimate at 4,096 bytes per input point. Larger inputs retain scalar deskew. Foreign points or different options also retain scalar behavior.
| Work per complete run | Prior | Reused |
|---|---|---|
| Candidate evaluations | 341 | 341 |
| Point projections, including sensitivity and holdout | 2,462,720 | 2,462,720 |
| Actual point motion-compensation computations | 2,462,720 | 556,032 |
Complete fresh-process comparison¶
Three alternating pairs use frozen source trees, actual Git identities, identical dependencies and one BLAS thread. The host is unreserved; all variation is retained. Whole-process timing includes imports, Git prechecks, problem parsing/model validation, construction, optimization, work counting, result packaging/saving and imported-source hashing. Generation, downloading, freezing, source-input compression and external audits are untimed.
| Pair (execution order) | Prior whole s | Reused whole s | Paired ratio |
|---|---|---|---|
| 1 (prior, reused) | 136.523854 | 35.134018 | 3.886x |
| 2 (reused, prior) | 147.510615 | 33.203555 | 4.443x |
| 3 (prior, reused) | 121.157007 | 26.909230 | 4.502x |
The 4.111724x headline is the ratio of three-run medians
for this workload. Baseline: a844dc00354f987da0e103c0dc5bb9610df2eb2a.
Measured runtime: 0fc27960653a895f4d5c907e8ee65a7e87e7685d. Only the continuous-time
Camera-LiDAR solver differs in the complete core manifests. No run is discarded
or relabeled as warmup. An initial failed provenance-format check is retained
separately, with its error, source/input snapshots and incomplete attempt.
It contributes no successful timing or fabricated result. The retry changes
only the launcher: both variants' complete core manifests and input bytes are
identical between the failed and final cohorts.
Validation and provenance¶
The portable audit checks every pinned source/input/helper/schema, each actual
import and full/abbreviated Git identity, all six complete results, every trace
row, the unchanged initial state/options/split, completion, work counts and
positive finite timers. Only the top-level provenance.created_at and
provenance.git_commit are excluded from cross-source numerical comparison;
the latter is independently checked against actual measured source identity.
The implementation and related suite passes 68 cases, and eight provenance regressions cover actual core abbreviations and wrong/missing identities. Regression coverage includes coordinate bits, signed zero, two-state eviction, seconds/nanoseconds, rolling shutter in both directions, constant/piecewise motion, unsupported temporal ranges, memory fallback, fresh capture/option ownership and complete weighted optimizer traces with uncertainty ablations. Sixteen archive checks and fifteen schema-registry checks bring the distinct targeted total to 107. Strict mypy covers all 378 core files; Ruff, strict docs and all 129 public schemas plus their ledger are checked separately.
The complete archive contains the schema-valid problem, exact original generation inputs, generator, complete source capsules/manifests, every result, timings, failure and validation receipts. Derived data retains KITTI attribution and CC BY-NC-SA 3.0 separately from Apache-2.0 source code.
Reproduce¶
Audit without a calibration run or raw download:
python -m tools.benchmark_continuous_camera_lidar_reuse \
--audit-only docs/assets/continuous_camera_lidar_evaluation_reuse
For a fresh comparison, decompress the two source JSON inputs into a new directory (the prepared problem already contains every calibration input):
import gzip
from pathlib import Path
archive = Path('docs/assets/continuous_camera_lidar_evaluation_reuse')
inputs = Path('outputs/timed-camera-remeasurement-inputs')
inputs.mkdir(parents=True, exist_ok=False)
for name in ('source-points.json', 'motion-control.json'):
(inputs / name).write_bytes(gzip.decompress((archive / (name + '.gz')).read_bytes()))
python -m tools.benchmark_continuous_camera_lidar_reuse \
--baseline-commit a844dc00354f987da0e103c0dc5bb9610df2eb2a \
--problem docs/assets/continuous_camera_lidar_evaluation_reuse/problem.json \
--input-provenance docs/assets/continuous_camera_lidar_evaluation_reuse/input-provenance.json \
--control-generator docs/assets/continuous_camera_lidar_evaluation_reuse/control-generator.py.txt \
--source-points outputs/timed-camera-remeasurement-inputs/source-points.json \
--motion-control outputs/timed-camera-remeasurement-inputs/motion-control.json \
--output-dir outputs/timed-camera-new --repetitions 3