Object File Tracker¶
Stateful tracker — instantiates Object-Centered Feature Binding
Keeps a persistent bundle for each object and carries its identity across time, updating or splitting it as new frames arrive.
An Object File Tracker holds an open "file" for each object it believes exists and keeps that file the same file across time as fresh observations arrive. Its defining commitment — the thing that separates it from a per-frame binder — is temporal identity: the question it answers is not "which features belong together right now?" but "is this new feature the continuation of an object I already have?" Each file accumulates a history, predicts where its object should appear next, and matches incoming features against that prediction, so a briefly occluded object can be re-bound to its existing identity rather than reborn as a stranger. The whole mechanism is organized around persistence across gaps: it commits a new observation into an existing file only when confidence clears a bar, and it will hold or split a file rather than let one object's identity leak into another's.
Example¶
A soccer broadcast system is tracking every player on the pitch to drive automated camera cuts and stats overlays. Each player is an object file: a bundle carrying that player's recent positions, jersey color, and running speed. On a clean frame this is easy. The test comes when two players cross — a defender and an attacker briefly overlapping near the touchline, their bounding boxes merging into one blob for half a second.
A memoryless per-frame binder would emerge from the overlap and re-bind the two boxes arbitrarily, sometimes swapping the players' identities. The Object File Tracker instead leans on each file's motion history: before the crossing, it predicts where each player will be a moment later given their speed and heading, and after the blob separates it matches each reappearing box to the file whose prediction it best continues, only committing the re-bind when the match confidence is high enough. The defender stays the defender through the occlusion; the attacker stays the attacker. Where the two truly cannot be told apart, the tracker holds both files as provisionally uncertain rather than guessing — and if a single tracked blob turns out to have been two players all along, it splits the file into two. The output is stable player identities that survive contact, occlusion, and clutter.
How it works¶
The tracker's substance is prediction-and-match over time. Each object file carries a state and a learned motion/association prior that projects that state forward to the next observation. Incoming features are scored against every file's prediction, and the best match is committed into that file only if it clears a confidence bar; a feature matching nothing well enough seeds a candidate new file, and a file that goes unmatched for too long is retired. Two behaviors distinguish it from a static binder: it reuses identity — the same file persists across frames, so history compounds — and it can split a file when the evidence inside it stops cohering (one tracked bundle whose parts diverge becomes two). Crucially, prediction is a prior, not proof: a low-confidence match is held open, not forced.
Tuning parameters¶
- Motion/association prior strength — how confidently the tracker projects a file forward. A strong prior sails through occlusion but can drag a file onto a plausible-but-wrong continuation when the object behaves unexpectedly.
- Commit threshold — how good a match must be to be absorbed into an existing file. High thresholds resist identity swaps but spawn more fragment files; low thresholds keep identities continuous at the risk of silent leakage.
- Coast duration — how long a file survives with no matching observation before retirement. Longer coasting recovers from long occlusions but risks re-binding a returning stranger to a dead file.
- Split sensitivity — how much internal divergence triggers splitting one file into two. Eager splitting catches merged objects early but fragments a single noisy object.
When it helps, and when it misleads¶
It is the right mechanism whenever identity must persist through gaps — occlusion, dropout, brief noise — and where losing or swapping identity is costly. Its temporal memory is exactly what lets it out-perform per-frame binding through contact and clutter. The very idea of a persistent perceptual bundle carried across time and re-matched to reappearing features has a long pedigree.[1]
Its failure mode is identity drift: because a file wants to continue, a confident-but-wrong match can be absorbed and then defended by the accumulating history, so the tracker glides smoothly onto the wrong object and only a hard contradiction dislodges it. The classic misuse is trusting a long-coasting file through a heavy occlusion and re-binding whoever emerges first to the oldest waiting identity. The guarding discipline is to keep the commit threshold honest, cap coast duration, and let low-confidence continuations sit as held rather than being written into the file — and to hand genuinely tangled cases to a review step rather than resolving them by momentum.
How it implements the components¶
object_hypothesis_set— each open file is a standing object hypothesis, maintained as a persistent unit across time rather than re-proposed each frame.learned_association_prior— a motion/continuity model projects each file forward so reappearing features can be matched to the identity they continue.binding_confidence_threshold— a new observation is committed into an existing file only when the match confidence clears the bar; below it, the continuation is held.
It does not fuse across sensing modalities via a feature_channel_inventory or a combined binding_evidence_vector — that cross-channel role is Multimodal Fusion Tracker, its nearest twin. Both are "trackers," but this one carries one object's identity across TIME within a channel, while the fusion tracker binds features across MODALITIES at an instant.
Related¶
- Instantiates: Object-Centered Feature Binding — supplies temporal identity persistence to the archetype's bundles.
- Sibling mechanisms: Multimodal Fusion Tracker · Assignment Optimization Matcher · Merge/Split Review Queue · Temporal Coincidence Detector
Editorial Notes¶
Form Classification¶
Form family: Monitoring, Sensing & Alerting
Rationale: Object File Tracker operates as ongoing observation, sensing, or alerting that detects and surfaces state without itself executing the response because it keeps a persistent bundle for each object and carries its identity across time, updating or splitting it as new frames arrive.
Independent corroboration: The frozen evidence defines Object File Tracker as 'Keeps a persistent bundle for each object and carries its identity across time, updating or splitting it as new frames arrive', so its operative form is Monitoring, Sensing & Alerting.
Nearest alternative: Control, Automation & Runtime — Object File Tracker includes features of a live operational control that automatically routes, enforces, adapts, or responds during execution, but its defining operation is ongoing observation, sensing, or alerting that detects and surfaces state without itself executing the response.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Cognitive Science
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Cognitive science introduced object files as persistent representations that bind features and maintain identity across motion and re-encounter.
Related originating lineages:
- Computer Science & Software Engineering — Computer vision and multi-object tracking independently implement persistent identity records.
- Neuroscience — Visual binding and tracking research contributes candidate neural mechanisms for object continuity.
- Psychology — Experimental perception research supplied the behavioral evidence and tracking paradigms underlying object continuity.
Review resolution: Both independent reviews agree on primary origin cognitive_science; reconciliation resolves reported_ambiguity, alternate_origin_disagreement, origin_mode_disagreement. Formative alternate lineages retained: psychology, computer_science, neuroscience. The broader reach of later applications is kept separate as domain_reach=multi_domain; origin_mode=cross_disciplinary_synthesis describes the historical relationship among lineages. Confidence is conservatively reconciled to medium, and encyclopedia_synthesis=true preserves the reviewers' boundary judgment.
Attribution caveat: The title joins the psychological object-file construct with an engineered stateful tracker.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; medium confidence.
References¶
[1] The object file — a persistent mental representation that binds an object's features and carries its identity across time and motion, updated as it is re-encountered rather than recreated (Kahneman, Treisman & Gibbs, 1992). The mechanism borrows the name and the commitment to identity-through-time. withdrawn registry ↩