[px12811]

Description

After #2750 added confidence-ordered mAR matching, equal-IoU targets still chose the first eligible target instead of COCO's last-target tie rule. A higher-confidence prediction could therefore consume the only target available to a later prediction and undercount recall.

This PR now builds on the merged #2750. It reverses the eligible target indices before selecting the maximum IoU, preserving confidence order, per-class detection limits, and crowd/ignore handling. The shared matcher and other metrics are unchanged.

For targets [0, 0, 10, 10] and [10, 0, 20, 10], with predictions [0, 0, 20, 10] and [0, 0, 10, 10] at scores 0.9 and 0.8, current develop returns AR [0.05, 0.50, 0.50]; this fix and pycocotools return [0.05, 0.55, 0.55].

Motivation and Context

Refs #2701 and follows up on #2750. Conflicts with #2750 have been resolved while retaining its crowd/ignore support and tests.

Changes Made

• Match COCO's equal-IoU target tie rule in the existing confidence-ordered matcher.
• Retain box/mask/OBB confidence-order regression coverage and add equal-IoU cases with ordinary, crowd, and ignored targets.
• Correct the expected AR in fix(metrics): leave crowd and ignored targets out of MeanAverageRecall #2750's overlapping-target test: official COCO gives [0.1, 0.6, 0.6] because its first prediction has equal IoU with both targets.
• Update the matcher docstring and changelog.

Testing

CPU validation with Python 3.13.16 and the NumPy fallback backend:

• New equal-IoU regressions on develop 4922f3a4: 9 failed; all pass with the fix.

• mAR tests: 46 passed. All metrics tests: 691 passed.

• Full suite: 5751 passed, 123 skipped, 94% coverage. Baseline: 5733 passed, 123 skipped; every previously passing test still passes.

• pre-commit run --all-files: passed, including mypy.

• Official pycocotools 2.0.11: 8/8 tie comparisons agree for boxes/masks, with and without crowd targets. The unmodified baseline disagrees on all eight.

• Supplemental seeded comparison: 100 box scenes + 100 mask scenes, all 606 valid area results agree within 1e-7; this randomized comparison excludes crowd/ignore targets.

• Added/updated tests

• Updated docstrings

• Added an Unreleased changelog entry

Additional Notes

No new dependency or public API. Implementation and tests were developed with Codex assistance.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94%. Comparing base (a8b1ef4) to head (2ab2d54).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #2758   +/-   ##
=======================================
  Coverage       94%     94%           
=======================================
  Files           83      83           
  Lines        13644   13644           
=======================================
  Hits         12839   12839           
  Misses         805     805           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

🟢 Approval recommended

The localized matching change has focused regression coverage and no unresolved blocking findings.

No open issues 🎉

What changed in this PR

🟢 Approve — This localized fix aligns mAR matching with COCO’s confidence ordering while preserving per-class detection limits and other metrics’ behavior. No actionable issues remain.

Changes:

• Match predictions by descending confidence, independently at each IoU threshold, with COCO-compatible target tie-breaking.
• Add 14 regression cases covering geometries, ordering, class isolation, thresholds, and empty inputs.
• Update the mAR docstring and changelog.

Static coverage included behavior, compatibility, tests, and documentation; tests were not rerun here. Code: 5/5 (focused implementation); testing: 5/5 (targeted regressions); documentation: 5/5 (accurate scope). Overall risk: 1/5.

File	Description
tests/​metrics/​test_mean_average_recall.py	Adds confidence-order and matching-edge-case regressions.
src/​supervision/​metrics/​mean_average_recall.py	Introduces the mAR-specific matcher and documents its behavior.
docs/​changelog.md	Records the matching correction and unchanged scope.

_{Review effort: Balanced}

[Borda]

Can we rather install this package as a test dependency and run live results