Centroid coordinates must be numerically identical for dense and compact.

All-zero masks should return centroid (0, 0) — same as dense.

All-zero tight crops must still return centroid (0, 0).

Empty CompactMask (0 objects) must return shape (0, 2).

A tiny memory_limit forces chunking yet returns the same matrix.

After merging datasets, class_name must match remapped class_id.

Integration test: from_yolo should produce class_name data.

Basic case: class_name data field is set from classes and class_id.

Empty Detections should not raise an error.

When classes is empty, class_name should not be populated.

COCO categories[].id must be 1-indexed (COCO spec / CVAT requirement).

Multi-segment mask exports segmentation as RLE dict (iscrowd inferred as 1).

COCO polygon segmentation survives the load/export sequence. 1. Write source COCO JSON with polygon segmentation. 2. Load it through Detectio

When masks are NOT decoded (with_masks=False), raw polygon data stored in data['segmentation'] is used as a lossless fallback so as_coco() still e

Single image with two detections of different classes round-trips losslessly.

as_coco -> from_coco is lossless for internal class_ids while the on-disk COCO category ids are 1-indexed (regression for #1181).

Verify that Color.as_hex correctly converts Color objects to hex strings. Scenario: Converting a `Color` object back to a hex string. Ex

Verify that Color.from_hex correctly parses various hex string formats. Scenario: Creating a `Color` object from various hex string formats

compact_masks=True and False must produce pixel-identical final masks.

CompactMask is lossy: True pixels outside xyxy bbox are silently dropped. This test documents that compact and dense paths diverge when a mas

CompactMask + dense mixed dispatch must match dense + dense.

Default (compact_masks=False) keeps dense ndarray masks.

Masks returned from callback are CompactMask after _run_callback.

Test integration scenario with multiple images and varying performance.

Verify that Average Precision is correctly calculated from PR curve points. Scenario: Computing Average Precision (AP) from PR curve

Test that F1 score respects confidence ranking

Test that higher confidence predictions are preferred for matching

Test that predictions are ranked by confidence

Direct from_tensors OBB end-to-end coverage with pre-built tensors.

MetricTarget.MASKS raises ValueError at every public entry point.

metric_target field reflects the value passed to from_detections.

metric_target field reflects the value passed to from_tensors.

Multi-class OBB: TP for exact match, FP+FN for rotation mismatch.

Verify OBB support in ConfusionMatrix. Test scenarios: 1. Perfect OBB overlap (Rotation Match) 2. Rotation Sensitivit

Regression for #1760: thin OBBs with same AABB must not match. Two thin bars (100 px long, 10 px wide) at 45° and -45° share an identical

Test confusion matrix calculation on a YOLO-format dataset. This test verifies that the confusion matrix fix (considering both IoU A

Deprecated confidence= warns and maps value to keypoint_confidence.

Explicit confidence=None is silently ignored — no warning emitted.

Explicit data=None normalizes to empty dict, not None.

Normal keypoint_confidence= path works and emits no deprecation warning.

Invalid constructor parameters raise ValueError.

ValueError raised regardless of kwarg order when both are passed.

contains_holes must agree after CompactMask encode→decode.

contains_multiple_segments must agree after CompactMask encode→decode.

Logger is created with provided name.

Token count in saved YOLO line reflects the is_obb flag at export time.

OBB round-trip via `from_yolo` -> `as_yolo` must preserve the 4 corners.

OBB round-trip with a label-less image must not raise ValueError.

Verify that multiple DetectionDataset objects can be successfully merged. Ensures that multiple `DetectionDataset` objects can be merged int

Test mAR with a roboflow-format dataset loaded from disk. Uses a synthetic YOLO-format dataset loaded via DetectionDataset.from_yolo() t

Logger defaults to INFO level when LOG_LEVEL env var is not set.

Logger is created with default name.

Degenerate boxes (zero-area collapsed or collinear) always score 0. Note: self-comparison of a zero-area box returns 0.0, not 1.0 — this

An OBB whose four corners coincide has zero area — the shoelace formula must not produce NaN or a negative value.

_delta_decode undoes COCO delta: counts[i] += counts[i-2] for i > 2.

_delta_encode applies COCO delta: d[i] = counts[i] - counts[i-2] for i > 2.

_delta_decode(_delta_encode(counts)) == counts for any count list.

Resized CompactMask matches OpenCV-resized dense masks within 1px.

detection > memory_limit emits UserWarning; result shape is still correct.

Internal class_id k must serialize to COCO category_id k + 1.