boxmot.engine.calibration.kalman_replay

Replay detection-only Kalman histories and score held-out box forecasts.

DEFAULT_HORIZONS module-attribute

DEFAULT_HORIZONS = (1, 3, 5)

box_metrics

box_metrics(
    predicted: ndarray, truth: ndarray, *, is_obb: bool
) -> tuple[float, float]

Score physical boxes in common units, using exact oriented overlap.

batch_box_metrics

batch_box_metrics(
    predicted: ndarray, truth: ndarray, *, is_obb: bool
) -> ndarray

Score corresponding boxes in linear space, retaining exact rotated IoU.

replay_track

replay_track(
    track: CalibrationTrack,
    *,
    tracker_name: str,
    geometry: str,
    config: Mapping[str, Any],
    camera_warps: ndarray | None = None,
    horizons: Sequence[int] = DEFAULT_HORIZONS,
) -> dict[int, dict[str, Any]]

Replay detection-only states and bounded open-loop forecasts from each observation.

Forecasts use the same observed history and then withhold all intermediate observations. Fixed-step, nonadaptive forecasts are batched from bounded snapshots of posterior states. Adaptive noise and measured timestamps keep forecasts at their observation origin so future innovations cannot enter prediction. With CMC, each delivered frame applies its image-derived warp after prediction, including intermediate frames with detections withheld. These are detection-gap replays, not forecasts of unavailable future images.

summarize_forecasts

summarize_forecasts(
    records: Sequence[tuple[CalibrationTrack, dict]],
    *,
    primary_horizon: int | None = None,
) -> dict[str, Any]

Weight trajectories, sequences, and supported horizons equally in order.