costs.gallery

Single-embedding matching (one stored vector per tracklet) is brittle when an object's appearance changes: a fresh detection from a new viewpoint can look unlike the last stored embedding even though it is the same identity. DeepSORT / MeMOT-style trackers keep a small **gallery** of recent embeddings per tracklet and match a detection against the *best* (or mean) of them, so a single good historical view is enough to re-associate. :class:`GalleryCost` reads a tracklet's gallery buffer and fill count (maintained by :class:`~unitrack.states.GalleryAppend`) and reduces the per-slot cosine similarities to an ``(N, M)`` cost matrix.