benchmarks.hota.tracker.build_learned_tracker
def build_learned_tracker(height: int, width: int, cost_threshold: float = 0.5, embed_dim: int = EMBED_DIM, checkpoint: str | Path = DEFAULT_LEARNED_CKPT) -> unitrack.TrackerBuild a MOTR-style learned appearance tracker from a trained checkpoint.
Wires the same cosine-distance association as :func:`build_cosine_tracker`,
but the embedding state is filtered by learned modules: a
:class:`~.learned_modules.Propagator` (predict, via ``LearnedProcess``) and a
:class:`~.learned_modules.Fuser` (update on match, via
``LearnedObservation``), loaded from ``checkpoint``. Raises
:class:`FileNotFoundError` if the checkpoint is missing.
``height`` / ``width`` are accepted for a uniform factory signature but unused
(appearance matching needs no mask state).
Source: unitrack/benchmarks/hota/tracker.py:345