Reset random state before each test for reproducibility.

Execute the eval command.

Execute the track command.

Create sample GT and tracker MOT files for testing.

Restore a previously saved filter state. Args: state: Dictionary from `get_state`.

Fixture providing SportsMOT flat format test data.

Fixture providing SportsMOT MOT17 format test data.

Format as a table string. Args: columns: Metric columns to include. If `None`, includes all available metrics.

90° rotation of [0,0,2,4] gives axis-aligned enclosing [-4,0,0,2].

A track updated this frame survives even with 1 successful update.

Every class in _DISTRACTOR_CLASSES must be excluded from GT and suppress an overlapping tracker detection.

Frame where every GT row is a distractor class yields zero scored GT.

All three metric groups are present when all metrics requested.

All trackers return stable ascending unmatched indices on same pattern.

All detections are unmatched when no tracks exist; result is sorted.

--asset values are split on commas and whitespace-stripped.

--asset accepts comma-separated values.

Pairwise CIoU is at most 1; unlike IoU/DIoU/GIoU it can be < -1.

Tuner auto-adds the metric family required by the objective.

Axis-aligned R must propagate into P via A @ P @ A.T.

Batch path must update P when R is axis-aligned.

Vectorised batch call must equal applying the helper per element.

Larger buffer ratio yields equal-or-higher BIoU for near-miss boxes.

BIoU(buffer_ratio=0) should equal standard IoU for all box pairs.

CMC must preserve bbox width and height in xyxy space.

apply_cmc with an identity transform must leave state unchanged.

apply_cmc(None) must leave state and covariance unchanged.

A pure translation H must shift center by (tx, ty).

A bigger bounding box must produce strictly larger R diagonal values.

A bigger bounding box must produce strictly larger Q diagonal values.

BBox width/height stay positive even after many predictions.

time_since_update == maximum_frames_without_update is past the limit.

time_since_update == maximum_frames_without_update - 1 survives.

ByteTrack must call the configured IoU in its low-confidence association branch.

ByteTrack resets the consecutive-update counter after a missed frame.

confidence=None must route every detection to Stage 1 even when explicit-low-conf would not spawn. Asserts the actual semantic of treating `None`

Construction counts as the first successful consecutive update.

--cache-dir sets the cache directory.

Validate variant-vs-reference parity across shared geometric scenarios. `upper_bound` is used only for cases where the original tests expecte

CBIoUTracker must increment IDs monotonically within one session.

CIoU <= DIoU: aspect-ratio term is nonnegative after scaling by alpha.

When enable_cmc=False, passing frame to update is harmless.

reset() also resets the internal CMC state.

compute_clear_metrics produces the expected CLEAR-MOT counts across scenarios.

Output file exists and each line is valid 10-column MOT format.

Cross-axis R must leave P untouched (A=None branch).

Batch path must also freeze P when R has cross-axis terms.

Dataset is captured as positional argument.

Tracker(iou=None) must default to a standard IoU instance.

run_tune() passes all argparse.Namespace fields to tune() correctly.

Two clearly different frames must return a finite (2,3) matrix.

DIoU <= IoU: center-distance penalty is nonnegative.

A distractor-class GT (conf==1) must not be scored as GT, and a tracker detection overlapping it must be removed rather than counted FP.

downscale>1 must still return a finite (2,3) float32 transform.

Empty sequence list returns all-zero HOTA metrics.

Empty sequence list returns all-zero Identity metrics.

Frames with no detections are fed as sv.Detections.empty() without error.

Missing GT frame must produce no error and zero GT dets for that frame.

CMC.apply_batch with an empty list must not raise.

Empty input must return empty output.

EPS constant equals numpy float epsilon, matching the TrackEval reference.

Test evaluate_mot_sequences with DanceTrack flat format (auto-detected).

Test evaluate_mot_sequences with DanceTrack MOT17 format (auto-detected).

Test evaluate_mot_sequences with SportsMOT flat format (auto-detected).

Test evaluate_mot_sequences with SportsMOT MOT17 format (auto-detected).