Canonical Geometry

Geometry mode is explicit in both the structure type and TrackerSpec.

Mode	Canonical detections	Canonical tracks	Packed rows
AABB	Boxes(float32[N,4]) in xyxy	Tracks with Boxes	AABB6 input / AABB8 output
OBB	OrientedBoxes(float32[N,5]) in cxcywha	Tracks with OrientedBoxes	OBB7 input / OBB9 output

Scores and geometry are float32; class IDs, track IDs, and detection indices are int64. Tensors are CPU-contiguous. Constructors reject invalid dtype, device, rank, contiguity, finiteness, extent, score range, or row alignment instead of correcting inputs silently.

Those dtypes describe canonical structures. A plain NumPy input uses the packed row boundary and returns a C-contiguous float64 AABB8 or OBB9 matrix; a Detections input returns canonical Tracks.

A tracker instance has one fixed geometry mode. Its update validates that Detections.geometry matches. Evaluation also validates the dataset and build geometry before replay.

Tracks.detection_indices maps a track row to the current detection row. -1 marks a propagated track without a current detection.

OBB continuity

OBB angles use radians and remain finite but unwrapped. Trackers resolve equivalent width/height/angle representations relative to the current state, then retain damped angular velocity to avoid flip artifacts.

All registered Python trackers support OBB detections. Native factory validation rejects any unsupported tracker/geometry combination before state is created.