fix(metrics): harden scoring edge cases

[Borda]

This pull request introduces several important improvements and bug fixes to the detection metrics, focusing on stricter class ID validation, aligning mean average precision (mAP) with the COCO 101-point AP standard, improved handling of ignore flags, and enhanced per-class recall reporting. These changes improve metric correctness, robustness, and user transparency, and include expanded tests to verify the new behaviors.

Metrics calculation and correctness:

• MeanAveragePrecision now uses the COCO 101-point AP averaging method, which may change some reported mAP values to be more accurate and consistent with the COCO standard. [1] [2] [3] [4] [5] [6]
• ConfusionMatrix now strictly validates class IDs, rejecting negative or out-of-bounds values instead of wrapping them via integer casting, preventing subtle bugs in confusion matrix indexing. [1] [2] [3]

Ignore flag and COCO compatibility:

• The ignore flag in targets is now preserved and respected in COCO-style evaluation, allowing users to explicitly exclude certain ground truths from scoring. [1] [2] [3] [4]

Mean Average Recall improvements:

• MeanAverageRecallResult.recall_per_class now exposes per-class recall for each max-detection cutoff (e.g., @1, @10, @100), providing more granular recall analysis. The string representation and result shape have been updated accordingly. [1] [2] [3] [4] [5] [6] [7] [8]

Expanded and improved tests:

• New tests ensure that negative and overflowing class IDs are rejected, that the AP calculation matches the COCO standard exactly, and that ignore flags and per-class recall are handled as intended. [1] [2] [3] [4]

These changes improve the reliability and interpretability of detection metrics, and may require users to recalibrate thresholds or update their code if they previously relied on the old behaviors.

---

part of #2408

[codecov]

Codecov Report

❌ Patch coverage is 97.29730% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 87%. Comparing base (6a69197) to head (9b1210f).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #2411   +/-   ##
=======================================
  Coverage       87%     87%           
=======================================
  Files           70      70           
  Lines        10198   10218   +20     
=======================================
+ Hits          8823    8843   +20     
  Misses        1375    1375           

🚀 New features to boost your workflow:

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

[Copilot]

Pull request overview

This PR hardens dataset import/export round-trips (Pascal VOC, COCO, LabelMe) and tightens dataset/classification equality semantics so that equality checks reflect ordered classes and annotation/image content more predictably across IO boundaries.

Changes:

• Make DetectionDataset / ClassificationDataset equality stricter (ordered classes, ordered image_paths, and pixel-wise comparison for in-memory images); add Classifications.__eq__ using NumPy value equality.
• Improve Pascal VOC serialization/parsing: round float coordinates on export, make polygon parsing order-agnostic, and fill bbox-only masks when force_masks=True.
• Enforce COCO category-name uniqueness and improve LabelMe multi-component mask grouping to round-trip as a single detection via group_id.

Assessment (n/5):

• Code quality: 4/5
• Testing: 5/5
• Documentation: 4/5

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/supervision/dataset/core.py	Tightens dataset equality semantics and compares in-memory images by key + pixel values.
src/supervision/classification/core.py	Adds value-based equality for Classifications (NumPy-aware).
src/supervision/dataset/formats/pascal_voc.py	Rounds exported coords, robust polygon parsing, and force_masks rasterization for bbox-only objects.
src/supervision/dataset/formats/coco.py	Rejects duplicate COCO category names to avoid ambiguous mappings.
src/supervision/dataset/formats/labelme.py	Groups multi-component masks via group_id and reconstructs masks from grouped polygons.
tests/dataset/test_core.py	Adds regression tests for stricter dataset and Classifications equality behavior.
tests/dataset/formats/test_pascal_voc.py	Adds/updates tests for rounding, polygon parsing order independence, and bbox-mask filling with force_masks=True.
tests/dataset/formats/test_coco.py	Adds a test ensuring duplicate category names raise a targeted ValueError.
tests/dataset/formats/test_labelme.py	Extends tests to assert grouped export and that multi-component masks round-trip as one detection.
docs/changelog.md	Documents the dataset IO/equality and format-round-trip fixes.

[Copilot]

Pull request overview

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.