v0.2.0 — Mapping / SLAM / Perception QA platform (132 commits since v0.1.0)

[0.2.0] - 2026-05-23

This release consolidates roughly half a year of work since v0.1.0: 132 commits
that turn CloudAnalyzer into an end-to-end QA / benchmark / operations layer for
mapping, localization, and perception 3D pipelines.

Added

Map / SLAM evaluation

• ca map-evaluate — MapEval-inspired accuracy / completeness @ threshold
  metrics for reconstructed maps against a GT map; supports MapEval-style
  initial alignment and optional colored error PLY artifacts.
• ca posegraph-validate — lightweight g2o / TUM / key_point_frame sanity
  check for manual loop-closure sessions.
• ca loop-closure-report — before/after/ref map + trajectory + posegraph QA
  in one command, with session-root auto-discovery and a quality-gate exit
  code. Available as a kind: loop_closure check in cloudanalyzer.yaml.
• ca scan-match-debug — scan-to-map ICP/GICP diagnosis with artifact bundle.
• ca slam-debug — SLAM debug report generator with primary and secondary
  diagnosis labels.
• ca mme — reference-free Mean Map Entropy via parallel scipy KDTree
  (contributed by @Taeyoung96 in #6).

Benchmarks

• ca benchmark info / ca benchmark eval — point a SLAM pipeline's map and
  trajectory at a frozen suite (reference + gate) for a one-command regression
  check; ships with a synthetic figure-8 suite under
  benchmarks/slam/synthetic-figure8/.
• ca benchmark init — build a suite.yaml from a local reference map and
  trajectory; works with arbitrary on-disk references.
• Newer College Dataset and KITTI Odometry wrappers
  (scripts/prepare_newer_college_mini.py, scripts/prepare_kitti_mini.py)
  with mini suite scaffolding under benchmarks/slam/.

Cross-representation geometry QA

• ca geometry-evaluate — same Chamfer / AUC / F1 metrics as ca evaluate,
  but first normalizes the source through a representation adapter so
  non-point-cloud inputs can be scored against a reference scan.
  • 3D Gaussian Splatting PLY adapter (opacity-aware filtering).
  • Triangle mesh adapter (OBJ / STL / GLB / GLTF / PLY-with-faces;
    surface-sampled, not vertex-only).
  • --representation auto|point-cloud|gaussian-points|mesh selector; output
    carries a representation block consumed by ca report-pr-comment.

CI / automation

• ca report-pr-comment — turn any CloudAnalyzer summary JSON
  (ca check, ca run-evaluate, ca benchmark eval) into a Markdown PR
  comment with worst-regression-first triage and ↑/↓ deltas against a baseline.
• ca bundle pack / show / unpack / diff — freeze one QA run (summary
  JSON, per-check reports, optional baseline) into a single qa_bundle.zip
  with a metadata header (project, commit, PR, notes). Reopenable without the
  original CI workspace.
• ca history — trend gate metrics across many bundles sorted by
  metadata.created_at; works on a rolling archive populated by
  ca bundle pack over time.
• New reusable workflows: pr-comment.yml, self-qa.yml,
  config-quality-gate.yml, baseline-gate.yml. PR comments idempotently
  update the same marker so re-runs don't stack duplicates.

Perception QA

• ca ground-evaluate — ground / non-ground segmentation QA
  (Precision / Recall / F1 / IoU).
• ca detection-evaluate — 3D bounding box detection QA from JSON box
  sequences with per-IoU mAP gate.
• ca tracking-evaluate — 3D multi-object tracking QA (MOTA, ID switches)
  from JSON box sequences.
• Checked-in public RELLIS-3D seed frame under
  demo_assets/public/rellis3d-frame-000001/ with deterministic
  detection / tracking JSON examples.

Public demo

• Static perception batch report at
  https://rsasaki0109.github.io/CloudAnalyzer/demo/perception/, regenerated
  by scripts/build_perception_demo.py and gated by CI for no-drift.

Changed

• Performance: vectorize NN distance paths with scipy.spatial.cKDTree
  (Phase 16). Both ca.experiments.map_evaluate.nn_thresholds and
  ca.stats.compute_stats previously walked an Open3D KD-tree one point at a
  time from Python, which made real-sized maps impractical. Both now use
  batched cKDTree.query; 100k-point NN finishes in well under a second.
  API and metric values are unchanged; sampling_policy.nn_backend in
  nn_thresholds shifts "open3d_kdtree" → "scipy_ckdtree".
• README: rewritten to point at docs/commands/*.md and docs/ci.md for
  long examples (772 → 430 lines) and brought up to date with all new
  subcommands.

Fixed

• pr-comment.yml: replace unsupported jq test() call with contains() to
  unblock the reusable workflow.
• Pre-existing mypy errors cleaned up so mypy ca/ cloudanalyzer_cli/ stays
  the CI gate.
• build_public_pack made deterministic so the public benchmark pack tests
  pass without external state.

Contributors

• @Taeyoung96 — ca mme (#6).