Skip to content

0.7.1 — 2026-07-19

← Back to the changelog index

Documentation

  • README overhaul + engineering-learnings consolidation. Reordered the README to a user-first structure (pitch → experimental caveats → supported markers & requirements → install → quick start, with benchmark depth moved below and into docs/explanation/performance.md), added the missing supported-family / platform / Python-version and API-reference information, and tightened the prose. Consolidated ~19 dated engineering postmortems/diagnostics (~3,500 lines) into four durable, repo-internal docs/engineering/lessons/ pages (pose covariance; rotation tail & edge refinement; EdLines; recall/quad/ICRA) plus an index and a going-forward convention; removed the distilled raw sources (git history preserves them). No code or shipped-behavior change.
  • README + SOTA-doc pose metrics refreshed for the v0.7.0 high_accuracy default. The render-tag comparison table (README render-tag-comparison snippet, re-used in docs/explanation/performance.md, and docs/engineering/benchmarking/render_tag_sota_20260713.md §1/§2) still showed the pre-refinement high_accuracy row. Refreshed to the shipped edge-refinement-on config on the same AMD EPYC-Milan hardware (competitor + standard rows reproduced identically and are unchanged): rot p99 0.600°→0.249° (1080p) / 1.113°→0.267° (2160p), rot p50 0.057°→0.041°, trans p99 18.6→20.1 mm (1080p) / 39.0→37.0 mm (2160p), latency 13.8→15.2 / 56.4→58.0 ms. high_accuracy now leads the rotation tail too (below OpenCV apriltag's 0.376°); corrected the stale "off by default" note.