Changelog for package mp2p_icp_core

3.0.1 (2026-09-28)

• Disable LTO for GCC < 12, which miscompiled devirtualized calls into MRPT classes (hanging/aborting unit tests on ROS 2 Humble)
• Contributors: Jose Luis Blanco-Claraco

3.0.0 (2026-09-28)

• FilterMLS: fix per-point fields (color, intensity, normal) getting misaligned with their points under TBB
• sm2mm: do not run per-frame filters on IMU-only observations
• Keep per-point fields in sync for VoxelAverage and flatten_to decimation (fixes out-of-bounds read)
• Add FilterRangeBiasCorrection: per-class range and incidence surface-bias model
• FilterDeskew: IMU methods take their initial velocity from 'twist' when defined
• FilterDeskew: fix ignore_accelerometer=true integrating a phantom free fall
• Add DecimateMethod::RotatingIndex, removing the shared per-scan displacement of FirstPoint decimation
• Matcher_Cov2Cov: optional per-point weighting of pairings by range and by beam incidence angle
• Solver_GaussNewton: make the parallel accumulation of H and g deterministic across thread counts
• FilterDecimateVoxels: faster ClosestToAverage/VoxelAverage via PointCloudToVoxelGridAverage; smaller voxel_t for FirstPoint
• Fix pre-existing clang-format-14 violations
• Add unit tests for previously untested code, fix a Matcher_Adaptive crash
• ICP: optional freezePairingsAfterIteration parameter
• Port to MRPT's CMultiMetricMap::mapsList()
• Port to MRPT 3.x (mrpt::opengl -> mrpt::viz, per-component find_package, yaml API changes, camera controller, etc.)
• Contributors: Jose Luis Blanco-Claraco

2.14.1 (2026-09-15)

• Geman-McClure: square the kernel parameter, as its own derivation says (#102)
• Two optional, env-gated solver diagnostics: H spectrum and the Birge ratio (#93)
• Contributors: Jose Luis Blanco-Claraco

2.14.0 (2026-09-05)

• Point matchers (Matcher_Points_DistanceThreshold, Matcher_Points_Blend): fixed a nondeterministic pairing order caused by parallel reduction depending on TBB thread scheduling; the correspondence list is now a function of the input alone, regardless of thread count.
• Solver_GaussNewton: the robust kernel now receives a whitened residual (using the existing pair weight as the whitening factor) instead of a raw metric or Mahalanobis norm, so [robustKernelParam]{.title-ref} means the same number of sigmas for every pair type. Bit-identical for pipelines using default weights.
• GravityPrior: report the fraction of the final rotational information the prior actually contributed, via new [gravity_information_share]{.title-ref} field in [OptimalTF_Result]{.title-ref} and [Results]{.title-ref} (negative when no prior was given); this value is now also persisted across serialization ([Results]{.title-ref} format v1) and reset at the start of each solve.
• New FilterPlanePatches: extracts large planar patches (equation, centroid, area, point count) into [metric_map_t::planes]{.title-ref}, complementing FilterEdgesPlanes' per-point labeling. Deterministic greedy extraction seeded by per-point normals. [metric_map_t]{.title-ref} serialization goes to v6.
• FilterVoxelReduction: made the downsampling order-independent (selection no longer depends on input point order) and tightened bounds on [range_min]{.title-ref} and [normal_agreement_deg]{.title-ref}.
• Contributors: Jose Luis Blanco-Claraco

2.13.1 (2026-08-25)

• Merge pull request #94 from MOLAorg/feat/mm-filter-generators-create-layer mm-filter: create new metric map layers via generators:

• Fix stale doc URLs missing the mp2p_icp_core path segment These sm2mm/mm-filter demo comments and the sm2mm.h docstring still pointed at pre-split paths (apps/sm2mm, apps/mm-filter, top-level demos/), left over from the mp2p_icp_core/mp2p_icp_viz package split. Same class of issue as flagged in review for the new demo added in this branch; fixing the other pre-existing occurrences found by a repo-wide search.

• Address review: fix demo URL, doc precedence, and unresolved formula check

  • Fix stale doc link in the new demo (missing mp2p_icp_core path segment).

  - CreateMetricMapFromDefinition(): docs said "exactly one" of the two definition sources must be given, but the implementation

File truncated at 100 lines see the full file