Partially overlapping masks: IoU should match the analytic value.

Test F1 score with perfect matching predictions and targets

Test recall with perfect matching predictions and targets

Test precision with perfect matching predictions and targets

Smoke test: MeanAveragePrecision accepts non-square OBB inputs without error. NOTE: MeanAveragePrecision uses the COCO evaluator path

Perfect non-square OBB predictions score 1.0 for metrics that use OBB IoU.

Regression test for #1746: _polygons_to_masks must return shape (N, H, W). The original PR rewrite processed only a single polygon and always ret

Test F1 score with different precision and recall scenarios

Prime-sized crops with non-integer scale ratios match cv2 exactly.

Verify that process_video works correctly with custom performance parameters. Scenario: Processing video with custom prefetch and buffer par

Verify that process_video correctly propagates exceptions from the callback. Scenario: Processing a video where the callback raises an excep

Verify that process_video handles exceptions correctly even with small buffers. Scenario: Processing a video with minimal buffering where an

Verify that process_video respects the max_frames parameter. Scenario: Processing only a limited number of frames using `max_frames`. Ex

Verify that process_video does not call _mux_audio when preserve_audio=False. Scenario: Default process_video call without setting preserve_

Verify that process_video calls _mux_audio when preserve_audio=True. Scenario: Processing a video with preserve_audio=True and ffmpeg availa

Verify successful video processing with a pass-through callback. Scenario: Successfully processing a video with a simple pass-through callba

ValueError when OBB data shape is wrong for area computation.

Non-positive epsilon_step must raise ValueError.

Percentage outside [0, 1) must raise ValueError.

An overlap metric other than IOU or IOS raises ValueError.

repack() is a no-op when bboxes are already tight.

repack() on an empty CompactMask returns another empty CompactMask.

repack() normalises an all-False mask to a 1x1 crop.

repack() shrinks the crop to the minimal True-pixel rectangle.

Test that reset() clears all stored data

Test that reset() clears all stored data

Test that reset() clears all stored data

_resize_crop falls back to cv2 for dense masks (above _L3_DENSITY_THRESHOLD). Checkerboard yields ~1 run per pixel, far above the 0.25 thresh

resize() dispatches to _rle_resize for sparse masks.

_rle_counts_to_mask reconstructs the correct boolean mask from run lengths.

Split then rejoin must reconstruct original mask for random inputs.

Round-trip must be lossless with inclusive xyxy from mask_to_xyxy.

_rle_resize matches cv2.resize(INTER_NEAREST) within 1-pixel tolerance.

Test that a second call to the same slicer instance does not re-emit the out-of-bounds warning even when detections are still out of bounds.

Test that a detection whose coordinates exactly equal the slice dimensions does not trigger the warning (boundary is exclusive: > not >=).

Test that no warning is emitted when callback returns detections within the slice bounds.

Test that the out-of-bounds warning is only emitted once even across multiple slices.

Test that exactly one warning fires even with thread_workers > 1, validating that the threading.Lock makes the check-and-set atomic.

Test that a warning is emitted when callback returns detections with negative coordinates, indicating wrong reference frame.

Test that a warning is emitted when callback returns detections with coordinates outside the slice bounds.

Every annotation's image_id must reference an image id present in the same file, even when a non-default starting_image_id is used.

An empty dataset writes a valid (but empty) COCO file and returns the starting ids unchanged so chaining still composes around it.

starting_image_id and starting_annotation_id below 1 must raise ValueError.

Dataset with images but zero detections per image: image ids are assigned sequentially but annotation list stays empty.

Resize scales shape and offsets proportionally.

Scenario 1: Success Case with imperfect match

Scenario 2: GT Present, No Prediction (Missed Detection)

Scenario 3: No GT, Prediction Present (False Positive)

Scenario 4: No GT, No Prediction (Category has no data)

Scenario 5: Only 1 of 3 Classes Present (Perfect Match)

Comparing a set with itself yields a symmetric matrix and a 1.0 diagonal.

Self-IoU diagonal must be 1.0 for masks with at least one True pixel.

len() and .shape must agree with the dense array.

Test precision calculation for single class with mixed results

Test recall calculation with some missed detections

Test F1 score calculation with mixed precision and recall

Test that a single perfect detection yields 1.0 recall.

Test that single perfect detection gets 1.0 mAP (not 0.0 due to ID=0 bug)

Single True pixel in a 3x3 mask upscaled preserves position.

A single OBB prediction should be returned as one singleton group.

Test skeleton enum has correct structure.

Test SKELETONS_BY_EDGE_COUNT dictionary mapping.

Test SKELETONS_BY_VERTEX_COUNT dictionary population.

Test SKELETONS_BY_VERTEX_COUNT uses last-in-wins for duplicate counts.

smooth=True must not crash for any unique-position count from 1 to 4. Each position is repeated twice to simulate brief holds between moves.

With <4 unique positions smooth=True output must match smooth=False. Verifies the dedup-then-fallback path: when unique_xy has ≤3 points,

smooth=False with a stationary tracker must not crash (regression guard). Ensures the refactor did not accidentally alter the smooth=False co

Moving tracker must produce a spline trace distinct from the raw polyline. Compares smooth=True output against smooth=False for the same move

A single annotate() call with smooth=True must not crash. When len(xy) == 1 the drawing guard skips cv2.polylines entirely; the dedup

Boxes average over window; confidence averages present values or None.

Two tracks with different confidence presence must not raise on merge.

update_with_detections warns and returns input when tracker_id is None.

Full window (length=3) averages all 3 frames, not just the last two.