Frame 1 spawns a track with a real ID (BoT-SORT special-case behavior).

On the first frame there is no previous frame, so identity is returned.

box_iou is numerically stable across sub-pixel and large-coordinate boxes.

Empty similarity matrix must return empty array unchanged.

Fusing with 0.5 scores must halve every IoU value.

Fusing with all-1 scores must leave IoU unchanged.

_fuse_score via normalize_for_fusion must preserve ranking for signed variants.

Each detection column must be scaled by its own score.

GIoU <= IoU always holds.

Two identical frames should produce near-zero translation.

Identity R with non-zero t must shift all corners by t.

GT (conf=0, non-distractor class) is neither scored GT nor a distractor. A tracker detection overlapping it must be kept, not suppressed. Be

An immature track that has been lost must be removed.

Test creating tracker with params dict like CLI would do.

box_iou raises ValueError when box_format is not 'xyxy' or 'xywh'.

box_ioa raises ValueError when box_format is not 'xyxy' or 'xywh'.

Invalid search_space value dicts raise ValueError at class definition.

json() includes HOTA fields when only HOTA computed.

--list sets the flag to True.

--list is False when omitted.

--list wins over dataset positional.

--list calls _print_available and returns 0.

Batch CMC must match per-track apply_cmc for a single track.

A mature track lost for a few frames survives inside the buffer.

Even a mature track must die once it exceeds the lost buffer.

--metrics defaults to ['CLEAR'] when not supplied.

HOTA depends only on association structure, not on the integer id values. Relabeling ground-truth and tracker ids with a consistent, non-mono

No dataset and no --list prints error to stderr and returns 1.

Heterogeneous state estimator types must raise TypeError immediately.

Mixed maturity and lost-time states — only valid tracks survive.

Consistent tracking across frames yields perfect IDTP/IDFN/IDFP.

Multi-frame sequences must accumulate GT/tracker dets across all frames.

Identical sequences aggregate to the same HOTA; TP counts are summed.

Multi-sequence aggregation sums raw counts and recomputes IDF1.

Batch CMC applies the same transform to every tracklet.

A track initialized at box A, then a detection at box B just outside should be associated by CBIoU (buffer expands boxes) but not by

Negative input coordinates must transform without sign-related artifacts.

Missing confidence should behave like usable detections, not suppress tracking.

Every tracker treats confidence=None the same as all-ones across multi-box batches. Multi-detection batches exercise confidence bucketing in trac

Non-contiguous unmatched indices are returned in ascending order. Scenario: 4 tracks, 5 detections; track 1 matches detection 2 (IoU=0.9).

Non-finite H values must propagate into track state without raising.

Passing None as frame_bgr must return identity without raising.

CMC.apply_batch with H=None must not change any tracklet state.

When --split and --asset are omitted, None is forwarded.

_normalize_type returns the expected base type for assorted annotations.

Objective string is normalized to uppercase regardless of input case.

ORU unfreeze fires correctly after exactly one missed frame. Copilot raised a concern that ``_observed`` stays ``True`` throughout the first

Optional arguments have correct defaults when omitted.

-o value is forwarded to download_dataset.

-o sets the output directory.

-o is an alias for --output.

estimate() must always return a (2, 3) float32 matrix.

Vectorized implementation is bit-identical to pre-vectorization sequential version. Guards the hot path against future silent drift. Compares

_print_available runs without raising and does not leak output.