#2331 made Precision and F1Score include classes that appear only in
predictions, and added regression tests to both. Recall was not touched, so
the line #2331 replaced is still there and the three metrics disagree about
which classes exist for identical input:
precision.matched_classes -> [0 1] precision_per_class (2, 10)
recall.matched_classes -> [0] recall_per_class (1, 10)
f1.matched_classes -> [0 1]
These read as parallel outputs, so zipping them silently truncates rather
than raising.
Recall for a class with no ground-truth instances is 0.0 rather than
undefined, which is what sklearn reports (it infers labels from the union of
y_true and y_pred) and what #2331 cited as its own standard. MICRO is
unchanged because an absent class contributes no false negatives, and
WEIGHTED is unchanged because its ground-truth support is zero. MACRO does
change, and the changelog says so.
Also of note: recall.py already carried #2331's WEIGHTED zero-support guard,
whose comment refers to 'only false-positive classes'. That state could not
arise in recall.py, because unique_classes came from ground truth alone. The
guard was propagated; the union that gives it meaning was not.
Addresses the review on #2468. Building the class union inside
_compute_recall_for_classes only covers samples that reach it, and samples with
predictions but no targets are skipped earlier in _compute. So matched_classes
could still disagree with Precision and F1Score for list inputs containing a
background image, which is the exact invariant the new test asserts.
Before, for one normal sample plus one background image predicting class 2:
precision.matched_classes -> [0 2]
recall.matched_classes -> [0]
Recall now handles len(targets) == 0 and len(predictions) > 0 the way Precision
does. No recall value changes, since a background image produces no false
negatives; only the tracked class set does.
* test: cover Recall bg-image size-bucket, dup & non-contiguous ids
* docs: strengthen Recall changelog migration note
* docs+perf: Recall doctest example; dedupe-then-union micro-opt
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Re-verify remaining supervision review backlog against develop HEAD; most items were already resolved by an intervening commit, only genuinely-open gaps got new fixes.
Fix float32 precision loss in box_iou_batch for large coordinates (GeoTIFF-scale) by accumulating in float64.
Raise ValueError instead of a strippable assert in EvaluationDataset.load_predictions for unknown image ids.
Add HeatMapAnnotator.reset() to clear accumulated heat for annotator reuse.
Add missing coverage: labelme export basename collisions, _greedy_match matcher, metrics.core ABC/enum contracts, metrics.utils.utils pandas guard; remove a global RNG-seed pollution site in a metrics test.
Document the last two undocumented public exports (calculate_masks_centroids, is_compressed_rle) and add usage examples to 17 previously-example-less public functions/classes (NMS/NMM helpers, draw utils, PolygonZoneAnnotator, mask/polygon converters).
* tests: load_predictions ValueError branch + empty-dataset coverage
* fix: box_iou_batch int-dtype overflow, narrow float32 precision claim
* feat: add reset() to TraceAnnotator/DetectionsSmoother, fix docstrings
* docs: fix temp file leak in coco.py docstring, rename misnamed test
---------
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
- Added `sv.mask_to_roi` as an explicit migration path for exclusive mask bounds
- Fixed COCO, CreateML, and Pascal VOC export validation to reject ambiguous or colliding dataset paths before writing
- Fixed in-memory `DetectionDataset` split and merge behavior
- Fixed `supervision` imports to avoid loading ByteTrack until it is used
- Fixed detection conversion helpers to support coordinate-convention migration while preserving legacy inclusive defaults
- Fixed Azure tag mapping, anchor rounding, and line-zone smoothing to avoid incorrect or ghost detections
- Fixed video processing shutdown handling for timeout and full-queue cases
- Improved downloader, validator, documentation, and regression coverage for the shipped dataset, detection, annotator, image, and video behavior
---------
Co-authored-by: Codex <codex@openai.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
- Reinstated NumPy-safe `Classifications` equality and ordered class-list comparisons in dataset equality.
- Restored greedy matching plus size-bucket scoring for Precision, Recall, F1, and MeanAverageRecall, with regression coverage for the medium-object boundary case.
- Filter size-bucket precision, recall, and F1 against target boxes so predictions no longer claim the bucket.
- Preserve confidence order for bucketed mAR@K scoring and return zero when a bucket has no support.
- Add regression coverage for bucket matching, empty-support mAR, top-K limits, and missing-mask errors.
---------
Co-authored-by: Codex <codex@openai.com>
- Use COCO 101-point AP averaging in the legacy mAP path so perfect and imperfect curves score consistently.
- Validate confusion-matrix class ids before indexing and preserve target ignore flags in the COCO-style evaluator.
- Keep mAR per-class recall for each max-detection cutoff and cover the scoring fixes with focused regressions.
- Return empty mAR scores with the same max-detection axis as non-empty results.
- Add an empty-input regression covering recall score and per-class result shapes.
- Update the public mAR docstring to describe per-image detection limits.
---------
Co-authored-by: Codex <codex@openai.com>
- Fixed crop annotation so overlapping detections sample from the original scene
- Fixed dataset exports to reject basename collisions, including case-insensitive collisions
- Fixed LMM connector mapping to support mirror enum aliases without a hand-maintained dispatch table
- Updated benchmark documentation to install the released inference package with metrics support
---------
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
- Fixed mAP calculation to count predictions on background-only images as false positives
- Fixed all-background mAP inputs to return 0.0 instead of NaN when no ground-truth classes exist
- Updated `from_tensors` documentation to define empty-target background images and their false-positive behavior
---------
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
- Fixed in-memory dict-form `DetectionDataset` image access, iteration, equality, and merge behavior, with deprecation messaging retained
- Fixed mAP to honor `metric_target` for mask and oriented-bounding-box evaluation, including correct IoU routing, area handling, crowd semantics, and missing-content errors
- Fixed `ConfusionMatrix.plot()` when plotting raw counts with default normalization disabled
- Improved mask mAP crowd handling performance and memory usage
- Updated the count-in-zone guide to use current APIs
---------
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Mask pixel-area counting used `np.sum` — `np.array([np.sum(m) for m in
masks])` in `Detections.area` and `np.sum(mask, axis=(1, 2))` in the metrics
`get_mask_size_category`. For boolean masks `np.count_nonzero` (with no axis)
dispatches to NumPy's SIMD popcount over the raw byte buffer, whereas every
axis-reduction form — `np.sum(..., axis=...)` and even `np.count_nonzero(...,
axis=...)` — falls back to a slower generic reduction. So counting per mask
with `np.count_nonzero` is several times faster than the "obvious" vectorized
sum, while producing bit-identical integer counts.
Route both sites through `np.fromiter((np.count_nonzero(m) for m in masks),
dtype=np.int64, count=len(masks))`. `dtype=np.int64` preserves the documented
`Detections.area` mask-branch dtype on every platform (a bare
`np.array([...])` of Python ints would be int32 on Windows).
Measured ~5x on 640x640 masks (e.g. `Detections.area`, N=300: ~24ms -> ~4ms),
faster across densities. `get_mask_size_category` feeds the size-bucketed
F1/Precision/Recall/mAP/mAR metrics, where it is invoked repeatedly per
dataset. Counts are integer-exact (verified over 400 randomized trials plus
empty / all-true / all-false / 1x1 edge cases).
Adds parity tests for `Detections.area` (dense mask) and
`get_mask_size_category` against an `np.sum` reference.
---------
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
- Added compact COCO RLE mask ingestion with a `CompactMask` representation and optional compact mask parsing during inference for substantially lower memory usage on sparse segmentation results.
- Added `Detections.to_compact_masks()` to convert existing dense masks into compact masks while preserving detection and collection metadata.
- Improved compact mask decoding performance with cropped RLE processing, batched decoding on the fast path, vectorized decoding for small images, optimized RLE traversal, and faster delta decoding.
- Improved mask metrics to operate directly on `CompactMask` instances, preserving the compact representation while producing results equivalent to dense masks.
- Fixed mixed-modality inference handling by keeping detections and masks aligned, isolating malformed RLE failures to individual predictions where possible, and falling back safely when decoding cannot be completed.
- Fixed compact mask conversion and parsing to preserve dense-mask pixel content across public parsing and slicing paths, while correctly documenting and applying the intended bbox-cropping behavior for compact COCO RLE masks.
- Improved COCO RLE validation with checks for malformed payloads, invalid dimensions, count overflows, image size limits, count-sum mismatches, bounding-box mismatches, and safe fallback behavior for incompatible mask sizes.
- Added inference benchmarks and documentation demonstrating the memory and inference-time characteristics of compact masks, including guidance on their performance tradeoffs and behavior.
---------
Co-authored-by: Codex <codex@openai.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
The two-pass np.unique deduplication in _match_detection_batch dropped
valid TP assignments when a prediction's best-IoU target was already
claimed by a higher-confidence prediction. The greedy one-pass algorithm
(sort by IoU desc, assign if neither target nor pred already matched)
was already used in _split_detections_by_outcome and
ConfusionMatrix.evaluate_detection_batch but missing from Recall,
F1Score, Precision, MeanAverageRecall, and MeanAveragePrecision.
- Fix all five _match_detection_batch implementations
- Add regression tests reproducing the issue #2378 example in each
affected metric class (IoU matrix [[1.0, 0.667], [0.333, 0.538]])
- Add kind='stable' to np.argsort in _match_detection_batch across all 5 metric
implementations (Recall, Precision, F1Score, MeanAverageRecall, deprecated
MeanAveragePrecision) to ensure deterministic TP assignment when IoU values tie
- Add test_greedy_matching_two_valid_pairs to TestDetectionMetrics covering the
deprecated MeanAveragePrecision._match_detection_batch with the issue #2378 IoU
matrix, closing the one missing regression test flagged by /review
* refactor(metrics): extract _greedy_match helper to eliminate 5x duplication
---------
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
Add postponed annotations to test modules and modernize one test helper annotation for Python 3.9-compatible collection.
---------
Co-authored-by: Codex <codex@openai.com>
- Fixed Precision and F1Score metrics to correctly count false positives on images with no ground-truth objects
- Fixed Precision and F1Score metrics to include prediction-only classes when computing class statistics and confusion-matrix aggregates
- Corrected MICRO and MACRO averaging so false positives from absent classes affect the score as expected
- Preserved WEIGHTED averaging behavior by weighting only classes with ground-truth support
- Added safeguards for all-background evaluation batches, returning stable zero-valued weighted scores when no support exists
- Fixed handling of predictions with `class_id=None` in background-only evaluation paths
- Added regression coverage for background-image false positives, absent-class predictions, averaging modes, and zero-support edge cases
- Updated metric documentation and result metadata to reflect that tracked classes now include prediction-only classes
---------
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
* Supersedes #1967
* fix: COCO-compliant mAR calculation
* Add complex test of mAP
* fix(metrics): cast optional detections fields in mAR metric for mypy
* Add `create_yolo_dataset` utility and refactor tests with fixtures for reusable scenarios
* Refine docstrings for clarity and consistency, adding inline formatting and fixing typos in test helpers and metrics.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: jirka <6035284+Borda@users.noreply.github.com>
* Extend docstrings to test cases for enhanced readability and consistency across testing modules
* Extend test case docstrings across multiple modules for improved clarity and consistency
* Refactor test case docstrings for clarity and consistency in `test_annotators.py`
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* refactor: reorganize project structure and update module imports
* configure setuptools to use 'src' layout and adjust package discovery
* refactor: move tests to `tests` directory and update imports accordingly
* update lint and mypy configurations to align with `src` and `tests` structure
* update test imports to use corrected `tests.helpers` module path
* fix(pre_commit): 🎨 auto format pre-commit hooks