Scoped transform reuse for joint LiDAR calibration¶
The complete public Livox joint SLAC pipeline improves from 800.253310
to 231.253153 seconds (3.460508x). Complete
run_calibration calls improve from 796.891655 to 228.921617
seconds (3.481068x). This is automatic CPU acceleration of
the existing native point-to-plane joint factors.
Change¶
Each point factor previously decoded both six-dimensional tangent blocks and composed their initial transforms on every residual evaluation. Thousands of points share those same transforms during each central-difference Jacobian column. The graph now provides a fresh, bounded scratch cache for each linearization or residual pass. Built-in point-to-plane factors reuse only the composed immutable transform. Every point rotation, scalar residual, whitening, finite difference and optimizer decision still executes in its original order.
Keys contain the exact float bytes of both initial frames and both tangent blocks, including signed zero. At most 128 transforms are retained per scope; this is an entry bound, not a process-memory claim. Each scope resets on return or exception and isolates nested and concurrent evaluations. Standalone factors, custom SE3 subclasses and unusual/non-finite numeric inputs keep the scalar path. Generic user callbacks still run on every evaluation. Public signatures, result fields, options and schemas retain their existing definitions. No new dependency is added.
Complete public workload¶
The two PCDs come from the official public Livox Horizon example. Their raw headers
contain 23,921 source and 23,666 target points. Both reader point caps are
disabled. The existing full-point configuration
uses the native_joint_slac backend for this execution comparison, with its
existing 1 m voxel, 1.5 m correspondence gate, 40-iteration ceiling and
convergence tolerance. Only backend and resolved data/output paths differ from
that configuration.
Every run processes 16,884 point-plane observations, 40 joint LM iterations, 41 complete Jacobians, and 65 residual passes, including the candidate objectives, spatial holdout and 24 known-bad probes. There are 13,599 training point factors plus the six-residual train-only pose prior, and 3,285 holdout factors. Each complete Jacobian has shape 13,605 x 12. Inspection, plane preparation, fixed-baseline solving and report saving also remain inside the complete pipeline.
This existing development capture remains degenerate, with rank 9, in both versions. Its status, diagnostics and final output are preserved. These are timing and parity measurements, not independent calibration-accuracy evidence. No SOTA cohort or leaderboard claim is opened.
Measurements¶
Two alternating pairs use fresh processes and frozen source trees with actual
Git identities. One BLAS thread is used on an unreserved host. Whole-worker time
includes cold startup, input loading, the complete pipeline, trace capture,
compression and saving. API time covers the entire run_calibration call,
including array-digest capture and the pipeline's own artifact saving. Downloads,
source freezing and external audit are outside the timed workers. All four
runs are retained.
| Pair (order) | Previous whole s | Reuse whole s | Paired ratio |
|---|---|---|---|
| 1 (prior, current) | 717.827525 | 244.669707 | 2.934x |
| 2 (current, prior) | 882.679095 | 217.836599 | 4.052x |
Reported ratios divide the two-run medians. This small comparison establishes the observed benefit for this capture and host; other graphs and surrounding stages can have different costs.
Parity and reproduction¶
All four full schema-valid calibration results match exactly after removing only run ID, wall-clock dates and verified source identity. The full parameter and evaluation trace also matches: every complete residual and Jacobian array has the same SHA-256 digest of its little-endian float64 bytes. Event counts, shapes, ordered factor IDs, final states and all probe states are checked. Full optimizer histories and diagnostics are retained inside each result.
The measured launcher is preserved verbatim. The post-measurement auditor additionally verifies both duplicate Git labels in run and solid-state metadata before excluding all three recorded source identities. The repair record binds the original launcher, final auditor and unchanged measured-file hashes; no calibration value, input, trace, timing or source capsule was edited.
Unit tests compare actual residual/Jacobian bytes and complete LM results with the scalar path, including observable geometry, outliers, unequal weights, whitening, dense/Schur solvers and fixed blocks. Tests also cover signed zero, distinct initial frames, bounded eviction, changed values, nested/concurrent scopes, errors, direct calls and unusual-input fallback. Windows/Python 3.12 and Linux/Python 3.10 pass the same tests. Archive mutation tests reject rehashed work reductions, altered array hashes, missing events, changed factor/probe states, unfrozen imports and false reported timings.
The complete archive contains
all full results, all event traces, measurements/logs, frozen source capsules,
the exact launcher/config/tool schema, input identities and validation evidence.
Raw PCDs are obtained separately from their original public source; the archive
does not redistribute them or imply a data license from the code license.
The baseline is c0fbf5e2ddbebd589fab9e11b235e9f7ab84d5e6 and the measured runtime is
8e0a7fe645b07bc10badc641e9902e35362c8925. Only the three declared graph modules differ.
python -m tools.benchmark_joint_lidar_transform_reuse \
--baseline-commit c0fbf5e2ddbebd589fab9e11b235e9f7ab84d5e6 \
--output outputs/joint-lidar-transform-reproduction --repetitions 2
python -m tools.benchmark_joint_lidar_transform_reuse \
--audit-only docs/assets/joint_lidar_transform_reuse
Use committed runtime sources and a new output directory. Offline archive audit requires no PCD download, ROS installation or GPU.