Timed Camera-LiDAR evaluation reuse

All six alternating fresh-process runs are retained. The complete prepared-input workflow changes from a median 136.523854 to 33.203555 s (4.111724x), and the artifact-backed API from 134.120710 to 30.923065 s (4.337239x). The host was unreserved; the benchmark report retains every timing and its variation.

The input contains 9,216 selected original KITTI LiDAR points from nine scans and original OXTS velocities of those scans. Pinhole image means, signed per-point times, fixed noise and extrinsic transforms are generated controls. These are performance/parity inputs, not independently measured image matches, clock truth or accuracy evidence. The original 500-evaluation bound and every solver option remain unchanged; all runs converge after 341 candidate evaluations.

All complete results and objective/acceptance traces match exactly, excluding only the two declared source/run provenance fields (git_commit, created_at). There are 2,462,720 point-projection requests per run. Actual motion compensation computations fall from 2,462,720 to 556,032. Point counts, confidence/projection gates, covariance loss, time sensitivity and disjoint holdout remain intact.

Whole-worker timing includes startup/imports, actual Git prechecks, problem/model validation, object construction, the complete solve, counters, result packaging, saving and imported-source hashing. Input generation, source freezing, compression and external audits are untimed. Both variants use one BLAS thread and identical dependencies. Only calibrex/solvers/continuous_time_camera_lidar_solver.py changes in the complete runtime manifests. Each worker retains actual full and abbreviated Git identities as well as imported module hashes.

problem.json is a public-schema-valid complete input. Compressed source points and the original kinematics control preserve the exact generation inputs; their uncompressed hashes are checked against the original generation receipt. The audit verifies original points/IDs/distributions, OXTS velocities/timestamps, declared point times, source capsules, every run, all results and work counts. index.json binds every other byte, including validation receipts and the initial failed provenance check, which is excluded from successful timing statistics. No failed run is relabeled as warmup.

KITTI source data and derived numeric controls are attributed to Andreas Geiger, Philip Lenz, Christoph Stiller and Raquel Urtasun, KITTI, under CC BY-NC-SA 3.0. The official raw-data page and original source URL remain in the motion input's provenance. Source-code capsules and generator code retain the project's Apache-2.0 license.

Audit without calibration, a GPU or another download:

python -m tools.benchmark_continuous_camera_lidar_reuse --audit-only docs/assets/continuous_camera_lidar_evaluation_reuse