Reidentify Objects (Torchreid)

How it fits

Typical backends

Caveats

• I/O contractThe output shape is chosen once from the connected sink and fixed for the run, mirroring the tracker arms: [Track] (box + trajectory unchanged, id rewritten to the global id), [TrackedObject] for id-plus-confidence, [BoundingBox] for the boxes only, or [UInt64] for the global-id set. Every arm derives from the same per-track records, so the arms agree id-for-id.
• Hard constraintA single OSNet checkpoint is required in the model slot and must match the backbone named by reid_model_name; the architecture is built first and the file is loaded into it, so a mismatch fails at startup. The backbone and offline loader are a vendored subset of KaiyangZhou/deep-person-reid; no weights are ever downloaded at runtime.
• State lifetimeThe identity bank persists across frames inside the running container and grows with the number of distinct objects; forget_after evicts a track after that many frames without an observation, bounding memory.
• Parameter interactionThe match threshold decays from sim_base toward sim_min with time since last seen, so a longer absence is matched more leniently; raise sim_min to forbid loose re-entries at the cost of fragmenting identities across long gaps.
• Parameter interactionmax_gap_frames, base_radius_px and radius_growth_px gate candidates by a motion-predicted search window that widens with the gap; an object that reappears far from its predicted position outside this window is treated as new even if its appearance matches.
• CostEach track's box is cropped and run through the embedding network per frame, so cost scales with object count. device `cuda` is strongly preferred; CPU inference is too slow for real-time streams.