Diagonal staircase OBBs: merged AABB equals axis-aligned union.

When the same tracker stays at an identical anchor point for several frames the trace buffer accumulates duplicate points. `scipy.spl

Logger has a StreamHandler pointing to stderr for warnings.

Logger has a StreamHandler pointing to stdout.

Detection straddling two adjacent zones is counted in exactly one zone. The old implementation clipped each detection box to the zone's bound

cm.sum(axis=(1,2)) must equal cm.area (the fast path).

`overlap_metric=IOS` divides by the smaller area, so a small box fully contained in a larger one scores exactly 1.0, while IoU is the area rat

Same class: lower-score OBB suppressed. Different class: both kept.

Single-row and single-col crops scale correctly with cv2 parity.

Test with 3 classes on single image - explicit N x K bug reproduction. THE BUG: mAR @ K was limiting detections per class per image, not

At threshold extremes, positive-overlap non-identical OBBs suppress at 0.0 and are both kept at 1.0.

Tight bounding boxes (mask_to_xyxy) must still produce identical IoU.

Logger has exactly two handlers: one for stdout, one for stderr.

CompactMask path: collect_union=True returns array matching dense.

Dense path: collect_union=True returns array covering all painted pixels.

Test unique vertices calculation from skeleton edges.

X-pattern OBBs: both survive under either method because OBB IoU < 0.5.

Area equals the rotated body's area regardless of rotation, not the AABB.

Verify that Vector.cross_product correctly calculates the scalar value. Scenario: Computing the cross product between a vector and a point.

Verify that Vector.magnitude correctly calculates Euclidean distance. Scenario: Calculating the magnitude (length) of a vector. Expected

Test that VideoAssets.list() returns all video filenames.

Test that VideoAssets enum members have correct attributes.

Verify that VideoInfo correctly retrieves metadata from a video file. Scenario: Retrieving metadata from a video file using `VideoInfo`.

Verify that VideoInfo preserves non-integer FPS values as floats. Scenario: Retrieving metadata from a video while OpenCV reports 23.976 fps

Edges with an endpoint marked not visible are not drawn.

Not-visible keypoints produce no ellipses.

Vertices marked not visible are not drawn.

When visible is None all vertices are drawn.

Edges with both endpoints visible are drawn.

Vertices marked visible are drawn.

Valid JSON of the wrong shape should yield empty results, not raise.

Test WEIGHTED averaging with specific example

Test download_assets with ImageAssets enum.

NMS filters overlapping keypoint skeletons.

NMS raises when required fields are missing.

Masks shifted fully outside should remain valid and decode to all-False.

with_offset must clip partial out-of-frame translations like move_masks.

Test download_assets with VideoAssets enum.

The result stays a valid polygon and respects the point budget. The exception is the 3-point floor: keeping a valid polygon (at least 3

xyxy values beyond the image boundary should be clipped silently.

Converts OBB corners to axis-aligned bounding boxes.

YOLO load with force_masks=True should produce a non-empty mask.

YOLO load/save round-trip should keep segmentation mask geometry stable.

An invalid bounding box should not crash from_dense.

Extreme downscale that collapses a 1px bbox returns valid 1x1 crop.

A percentage of 0 removes no points.

Scenario: 1 image with 3 classes - explicit bug reproduction. Demonstrates the N x K vs K issue: with 3 classes, the bug would allow 3 d

Processes the current frame, updating time durations for each tracker. Args: detections: The detections for the current frame, in

Tint image with solid color overlay at specified opacity. Args: image: The image to tint. color: Overlay tint color. Default

Convert bounding box to format `(min x, min y, max x, max y)`, i.e., `(top left, bottom right)`.

Get current position in bounding box format `(top left x, top left y, width, height)`.

Convert the result to a pandas DataFrame. Returns: The result as a DataFrame.

Convert the result to a pandas DataFrame. Returns: The result as a DataFrame.

Convert the result to a pandas DataFrame. Returns: The result as a DataFrame.

Convert the result to a pandas DataFrame. Returns: The result as a DataFrame.

Convert the result to a pandas DataFrame. Returns: The result as a DataFrame.

Scenario: 2 images with 2 classes with varying confidence levels. Tests that `mAR @ K` limits per image (not per class) by creating a case w