Reject absolute file_name values that escape the images directory.

Reject relative traversal file_name values that escape the images directory.

Reject file_name resolving to a subdirectory inside images/ (is_dir guard).

Reject file_name resolving to the images directory itself (equality guard).

Mask presence depends on annotation format and OBB/segmentation flag.

Logger respects the LOG_LEVEL environment variable.

Logger respects the LOG_LEVEL=WARNING environment variable.

Test MACRO averaging with specific example

Test MACRO averaging with specific example

Test that `mAR @ K` limits detections per image, not per class. BUG SCENARIO (what was wrong): The previous implementation would limit d

Create a simple rectangular mask fixture

_rle_counts_to_mask(_mask_to_rle_counts(m)) == m for non-empty masks.

When both `mask` and `ORIENTED_BOX_COORDINATES` are present, area is computed from the mask.

_mask_to_rle_counts produces correct COCO F-order run lengths.

Known boxes for empty, single-pixel, and full masks.

Vectorized output is bit-identical to the per-mask scalar loop.

compact_mask_iou_batch must be numerically identical to dense IoU.

Using tight bounding boxes (mask_to_xyxy) must still be accurate.

Random masks give the same matrix as the explicit (N, M, H, W) reference.

mask_iou_batch on bool input must not emit any RuntimeWarning.

Large covariance with max_axis still produces a bounded ellipse.

Smoke test: MeanAveragePrecision accepts metric_target=ORIENTED_BOUNDING_BOXES. NOTE: MeanAveragePrecision uses the COCO evaluator path (box_iou_

Test that MEDIA_ASSETS has all VideoAssets and ImageAssets as keys.

Test that MEDIA_ASSETS values are tuples of (url, md5_hash).

Produces correct merged OBB corners.

Merges detection group correctly.

Test MICRO averaging with specific example

Test that mismatched prediction/target list lengths raise error

Test that mismatched prediction/target list lengths raise error

Test that mismatched prediction/target list lengths raise error

Equal pixel counts but different (H, W) must raise, not silently mismatch.

Mixed mask splits correctly per column.

Test mixed scenario with some classes having no detections

One CompactMask + one dense array must still work correctly.

Merging a CompactMask with a dense ndarray falls back to dense.

N=4 masks at different positions all scale correctly after resize.

Multi-segment masks merge correctly for both force_masks values.

Uneven-length segments load without ValueError (issue #1209 regression).

Test F1 score calculation for multiple classes

Test recall calculation for multiple classes

Test precision calculation for multiple classes

Test that multiple perfect detections get 1.0 mAP

Test recall when multiple predictions compete for one target

_mux_audio leaves the output file unchanged when muxing cannot complete.

_mux_audio calls os.replace with video_path as destination on success.

N x N pairwise IoU: diagonal must be 1.0 for non-zero-area masks.

Merge groups must match between CompactMask and dense inputs.

Identical masks: all predictions should merge into one group.

Non-overlapping masks: every mask should be its own group.

Identical masks: only the highest-confidence one should survive.

NMS keep-set is identical for CompactMask and the equivalent dense array.

Borderline IoU pair (≈ threshold) must agree — catches the resize bug. With resize-to-640, sub-pixel rounding on a pair whose true IoU is ver

Non-overlapping masks: all should be kept.

With overlapping tiles, NMS must suppress duplicates using RLE IoU.

compact_masks=True must not crash when callback returns no masks.

Test F1 score with predictions that don't overlap with targets

Test recall with predictions that don't overlap with targets

Test precision with predictions that don't overlap with targets

Non-overlapping masks should always produce IoU = 0.

Logger does not propagate to the root logger.

Non-square resize: 4x6 to 2x3 with independent axis scaling.

Test that OBB callbacks are serialized even when thread_workers > 1.

Empty OBB detections return empty result.

Axis-aligned OBB NMM produces same envelope as AABB NMM.

OBB NMM produces correct geometry and confidence.

Serializing an object must not write the 1-index offset back into the inputs.

Mutation guard holds when xyxy is a NumPy row-view (the actual bug scenario).

Offset into a larger destination image must preserve pixels.

IoU is invariant under per-axis (anisotropic) scaling. An affine map multiplies every area — intersection and union alike — by the same deter

Anchors with negative coordinates are excluded, not wrapped or clamped.

Small box inside large: IOU keeps both; IOS suppresses small.

_paint_masks_by_area is a no-op when detections carry no mask.

Thread-pool path produces the same result as the sequential path.

resize() with N=10 masks exercises ThreadPoolExecutor; output is correct.