data["area"] should override computed bbox area even when a mask is present.

All polygons from a multi-component mask must be exported, not just the first.

When detections have no area in data, area should fall back to bbox area.

iscrowd should be stored as int (0 or 1), not as Python bool.

area stored in detections.data should be used instead of bbox area.

iscrowd stored in detections.data should be used when no mask is present.

Detections with no class_id must raise ValueError before +1 arithmetic.

Two-part disjoint mask round-trips through COCO export and import unchanged.

Non-crowd mask export produces one polygon list entry per connected component.

Exporting detections must not shift the source xyxy, and must be repeatable.

Verify that Detections objects are correctly converted to NumPy tensors. Scenario: Converting Detections objects to NumPy tensors.

OBB data in detections.data must be silently ignored when is_obb=False.

`is_obb=True` must serialize the 4 corners from `data['xyxyxyxy']`.

`is_obb=True` on an empty `Detections` must not raise — image had no labels.

OBB export must correctly serialize N>1 detections per image.

`is_obb=True` without `'xyxyxyxy'` in data must fail loudly, not silently.

Test F1 score at different IoU thresholds

Test recall at different IoU thresholds

Test precision at different IoU thresholds

Boxes whose envelopes do not overlap are pruned by the gate to 0.0.

Test download_assets downloading a new file.

Boundary case: empty Detections carrying an OBB data field must return an empty area array (matches the mask / box_area branches).

Empty inputs trigger the early-return path; result is zero-filled.

Empty CompactMask collections should return a zero-shaped matrix.

Resize of an empty CompactMask returns empty with new image_shape.

Test F1 score with empty predictions but existing targets

Test recall with empty predictions but existing targets

Test precision with empty predictions but existing targets

No OBB predictions should produce an empty boolean keep mask.

Test completely empty predictions and targets

No OBB predictions should produce no merge groups.

Test F1 score with predictions but no targets

Test recall with predictions but no targets

Test precision with predictions but no targets

_mask_to_rle_counts output matches the public mask_to_rle encoder.

_mask_to_rle_counts -> _rle_counts_to_mask round-trip is lossless.

_mask_to_rle_counts produces COCO-compatible F-order RLE for known masks.

Parallel rotated bars share an envelope but not a body, so they score 0. Exercises the path where a pair passes the axis-aligned gate yet has

Logger uses the explicitly provided level.

_extract_class_names returns class names sorted by class index.

Without ORIENTED_BOX_COORDINATES, area mirrors box_area (AABB).

Non-OBB heavily-overlapping AABBs collapse to one under either method.

Predictions of absent class count as FPs under MICRO/MACRO; WEIGHTED excludes them by design (GT support=0 → weight=0, consistent with sklearn

Predictions of absent class count as FPs under MICRO/MACRO; WEIGHTED excludes them by design (GT support=0 → weight=0, consistent with sklearn

Predictions on an image with no targets must count as false positives.

Predictions on an image with no targets must count as false positives.

WEIGHTED F1 is 0.0 when all images are background (no GT anywhere).

WEIGHTED precision is 0.0 when all images are background (no GT anywhere).

Test is_md5_hash_matching when file exists and hash matches.

Test is_md5_hash_matching when file exists but hash doesn't match.

Test is_md5_hash_matching when file doesn't exist.

float32 masks with pixel count > 2**24 must give exact IoU via float64.

COCO files with 0-indexed category ids (written by supervision <=0.28.x) must still load and produce correct internal 0-indexed class_ids.

A non-data-URI mask must not skip the item's confidence and desync arrays.

Empty and non-empty results should produce string-kind class_name arrays.

Test the from_inference method with valid input.

Test the from_inference method with invalid input.

SDK objects with .dict() and empty predictions produce string-kind class_name.

Test the from_mediapipe method with valid input.

Test the from_yolo_nas method with valid input.