
SAM 3 seed-and-propagate video tracker. A detector seeds objects (boxes or masks); SAM 3 memory propagates each object's mask across frames with a stable id. Mask-native resolved oneof: [TrackedSegmentation] (default), tracks, objects, boxes, or ids.
A detector seeds objects and SAM 3 propagates their masks, which are rendered over the live frame.
A recorded clip is seeded and tracked, then the per-object behaviour is summarised for reporting.
Image and a seed list, resolved once as either [BoundingBox] (box detector) or [Segmentation] (mask detector). The seed is a trigger, not a per-frame requirement — frames with an empty seed list run pure propagation and every already-registered object keeps being tracked.
seed_iou_gate is absorbed (the existing id and memory are kept); a seed below the gate against every live track registers a NEW id and pins that seed's class id. Lower the gate to re-seed more eagerly, raise it to suppress duplicate ids on noisy detections.
confidence_threshold is the per-frame presence-score floor; an object scoring below it is treated as absent on that frame (occluded / out of frame) and not emitted, but its memory is retained so it re-appears later with the SAME id.
[TrackedSegmentation] for id + per-pixel segmentation (the default, mask-native arm), [Track] for id + mask-bbox + center trajectory, [TrackedObject] for id + confidence, [BoundingBox] for id-less mask boxes, or [UInt64] for the bare id set. Every arm derives from the same per-frame propagated-instance list, so counts and ids stay consistent across arms.
[Track] arm fills each track's `trajectory` with a bounded history (last 64) of mask-bbox center points, oldest first, accumulated per track id across frames; an id that disappears drops its history. The box and segmentation carry the class id pinned from the seed that started the track, so the class stays meaningful downstream.
Automated release