ReWorld-Track: A Recursive Event World Model for Language-Guided Multi-Camera Tracking

1 INTRODUCTION

• Incomplete camera coverage creates blind gaps, and similar candidates at a return can make a wrong match corrupt the history used for later handoffs.Figure 1 illustrates multiple plausible routes during an 8.8 s mask and a target and competitor in the same return frame.Source: §1 · p. 1, Figure 1 · p. 2
• The unresolved challenge is how an association should change the state for the next forecast: a single match can erase plausible routes, while continued absence depends on camera availability.The update must carry candidate locations, motion histories, and absence evidence through successive handoffs.Source: §1 · pp. 1–2
• ReWorld-Track learns a task-specific predictive state over future target observations, forecasting camera, arrival time, and entry region and recursively revising that state after association.The model is framed as a structured-event predictor rather than a general-purpose or pixel-generative world model.Source: §1 · p. 2
• 65.19 HOTA on CityFlowV2 and 45.36 on MTMMC; on MTMMC, ReWorld-Track gains 1.24 HOTA points over adapted CRTracker and 0.50 over a comparably sized generic learned updater.Its identity-retention advantage grows across repeated handoffs.Source: §1 · p. 2
• Candidate- and null-conditioned states contribute through posterior weights and a learned recurrent projection, with later handoffs supervising what the state retains.A shared belief couples forecasts to identity decisions, and elapsed time and camera availability update it during waits.Source: §1 · pp. 2–3
• Matched controls connect improved return forecasts with stronger identity retention across scenes and repeated handoffs.Natural-gap and camera-masking tests examine recovery without fresh observations, while a matched-coverage diagnostic measures the role of waiting.Source: §1 · p. 3

Identity memory and uncertain association

• Graph association, track queries, object-permanence models, and probabilistic or multiple-hypothesis tracking retain identity evidence or alternative histories; ReWorld-Track learns the predictive state passed between cameras from weighted candidate and null states.The cited examples include Whareformer [Chalk et al., 2026] and Polycepta [Nagy et al., 2026].Source: §2 · p. 3

Predicting returns across cameras

• Trajectory Tensors [Styles et al., 2022] forecast a future camera, arrival time, and location, while ReWorld-Track feeds candidate- and null-conditioned states into the next forecast to retain alternatives across transitions.SMO-MCTF extends location-tensor forecasting to multiple objects.Source: §2 · p. 3

Language as identity evidence

• Language-guided retrieval and tracking use descriptions to identify targets, but descriptions may fit multiple candidates or refer to hidden attributes; ReWorld-Track combines query support with appearance and the event prior.Examples include CRTracker [Chen et al., 2025], ViewSAM [Ge et al., 2026], and LaMMOn.Source: §2 · p. 3

Predictive latent dynamics and world models

• World models learn predictive latent states for future observations and downstream decisions; ReWorld-Track applies this perspective to structured target-observation events rather than future pixels.Its state predicts the next camera, arrival time, and entry region and is recursively revised by uncertain association outcomes.Source: §2 · p. 3

3.1 FROM A QUERY TO SUCCESSIVE HANDOFFS

• A query and source observation initialize the identity to follow; prediction advances its belief through a gap, association evaluates current candidates, and feedback prepares the state for the next observation.The tracker receives camera-local candidates with timestamps, boxes, appearance, and observed track prefixes, and may continue waiting instead of linking a candidate.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 PREDICTING THE NEXT OBSERVATION

• The predictive belief represents identity attributes, camera or blind-transition location, motion, and network presence, and recursive filtering propagates it over elapsed time on a fixed camera graph.A 256-dimensional recurrent state summarizes identity and observation history; graph-message layers constrain routes while the recurrent state adapts their probabilities to the target history.Source: §3.2 · p. 4
• The event forecast predicts a next camera, arrival time, and entry position; camera availability determines whether continued non-observation weakens a route or supplies no observation evidence.Candidate priors integrate event probability over camera, time, and entry region; appearance and query evidence then help score candidates against a temporary-null alternative.Source: §3.2 · pp. 4–5, §3.3 · p. 5

3.3 ASSOCIATING CANDIDATES OR CONTINUING TO WAIT

• Association normalizes event priors with candidate likelihoods to produce posterior probabilities for candidates and a temporary-null alternative, while acceptance requires a posterior threshold and margin for two consecutive updates.Null represents no match among current candidates; per-camera matching resolves competing identities, and camera-local association permits one view to wait while another observes the identity.Source: §3.3 · p. 5

3.4 CARRYING UNCERTAINTY INTO THE NEXT FORECAST

• Feedback weights candidate-conditioned and null states by their posteriors, then summarizes route probabilities, motion uncertainty, and identity evidence for the next forecast.ReWorld retains weighted hypothesis structure alongside a learned recurrent summary; the comparison with fixed-moment feedback and a generic set encoder uses the same predictor, scorer, training data, and rollout depth.Source: §3.4 · pp. 5–6

3.5 LEARNING FROM SUCCESSIVE RETURNS

• Training uses successive observation updates and handoffs so that later forecast errors can train an earlier posterior projection.Windows contain 32 updates and at most four handoffs; a right-censored waiting likelihood trains the probability that the target remains unseen through an observed interval.Source: §3.5 · p. 6

4.1 EXPERIMENTAL SETUP

• Evaluation uses separate RGB training and scene-disjoint tests for CityFlowV2 vehicles and MTMMC pedestrians, with queries limited to initialization-time observations.Detector misses remain eligible, and an incorrect first commitment counts as failure.Source: §4.1 · pp. 6–7

4.2 IDENTITY CONTINUITY IN VEHICLE AND PEDESTRIAN DOMAINS

• Vehicle AssA rises from 65.53% to 73.58%, and pedestrian AssA rises from 43.92% to 46.39%, with detection accuracy nearly unchanged in both domains.Across five paired seeds, mean pedestrian HOTA improves by 1.24 points over adapted CRTracker.Source: Table 1 · p. 6, §4.2 · p. 7

4.3 HOW FEEDBACK PRESERVES IDENTITY

• On MTMMC, removing the prior lowers HOTA from 45.36 to 40.02, while ReWorld-Track beats a similarly sized generic updater by 0.50 HOTA points.Next-camera accuracy rises from 86.03% to 87.41%, and median arrival error falls from 0.78 to 0.71 s.Source: §4.3 · p. 7, Table 2 · p. 7

4.4 MISSING OBSERVATIONS AND REPEATED HANDOFFS

• Across repeated handoffs, identity retention reaches 95.79% with feedback versus 82.24% with restarted prediction; at 96% attained reacquisition in the test-set sweep, false matches are 3.18% with null versus 30.27% without.Source: §4.4 · p. 9

4.3 HOW FEEDBACK PRESERVES IDENTITY

• Without temporary null, false matches rise from 4.90% to 12.74% as mean delay falls from 1.03 to 0.61 s, while IDF1 and HOTA decline.Temporary null permits evidence to accumulate before commitment; the table reports MTMMC controls across 2,749 handoffs and 10,683 absence decisions.Source: Table 2 · p. 7, §4.3 · p. 8

4.5 LANGUAGE EVIDENCE AND GENERALIZATION

• Natural descriptions add 10.9 HA points on visually ambiguous MTMMC candidates, compared with 2.1 on easy cases; feedback also remains useful on unfamiliar routes.With partly unobservable attributes, the full scorer remains above query omission while the hard attribute gate falls below it; on unseen topology, HA is 74.6%, 3.8 points above one-way prediction.Source: §4.5 · p. 9

5 CONCLUSION

• Matched controls connect the posterior update to stronger identity continuity across independent scenes and repeated handoffs.On the supplied camera graph, the belief links recovery after absence to history used for later returns.Source: §5 · p. 9
• Misleading descriptions and unfamiliar route statistics reduce recovery, leaving accurate identity continuation under these shifts an open challenge.Source: §5 · p. 9

AI USE STATEMENT

• Generative AI assisted literature research, experimental support, and language polishing; the authors reviewed and verified assisted content and take responsibility for the manuscript and reported results.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The study uses benchmark vehicle and pedestrian imagery under the corresponding access terms, with benchmark identities used for evaluation rather than real-world names.The ethics statement also cautions that language and visual evidence can be incomplete or biased and that uncertain predictions require care in consequential applications.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The source package includes eight figure PDFs, numerical data, and deterministic table builders for reported comparisons.The reproducibility statement points to sections and appendices documenting the model, protocols, and evaluations.Source: REPRODUCIBILITY STATEMENT · p. 10

A PEDESTRIAN IDENTITY CONTINUATION

• Figure 5 follows three MTMMC s17 identities through query initialization, intermediate observations, matching, and continuation along annotated transitions.GT20 and GT97 follow c08–c03–c12, while GT3 takes the reverse route; returns change the view of query attributes and matching evidence.Source: A PEDESTRIAN IDENTITY CONTINUATION · p. 15, Figure 5 · p. 15

B A VEHICLE EXAMPLE OF THE RECURSIVE UPDATE

• For GT396, observations in c023 at 49.9 s and c025 at 54.9 s establish history; c026 candidates at 58.9 s supply evidence whose posterior updates the state for the next handoff.Candidate and null probabilities weight conditioned states, preserving alternatives so later observations can refine an uncertain association.Source: B A VEHICLE EXAMPLE OF THE RECURSIVE… · p. 16, Figure 6 · p. 16

C LANGUAGE-CONDITIONED VEHICLE CASES

• For g1, c026-local P(null) is 0.72 and 0.90 at 54.8 and 55.4 s, then 0.07 and 0.04 at 58.9 and 61.9 s as support shifts to the c026 candidate.GT396 remains visible in available c025 during the earlier c026 null decisions.Source: C LANGUAGE-CONDITIONED VEHICLE CASES · p. 16

D ENLARGED LANGUAGE-CONDITIONED VEHICLE CASES

• Figure 7 follows GT396, GT334, and GT336 from source views through waiting, matching, and continuation, with identity labels and candidate alternatives shown in each frame.Historical observations are dotted trails, and a c026 null stage can occur while the target remains visible elsewhere in the network.Source: Figure 7 · p. 17, D ENLARGED LANGUAGE-CONDITIONED… · p. 18

D.1 GT396 WITH DISPLAY IDENTITY g1

• GT396 is rejected against current c026 alternatives at 55.4 s while still annotated in c025, then matched at 58.9 s and continued at 61.9 s.Figure 4(d) supplies first-appearance and acceptance timing, while Table 23 measures aggregate calibration.Source: §D.1 · p. 18

D.2 GT334 WITH DISPLAY IDENTITY g2

• The silver SUV GT334 retains display identity g2 across initialization at 48.3 s, c026 alternatives at 51.4 s, accepted match at 54.9 s, and continuation at 58.5 s.Source: Pred ID: g2 · p. 19

D.3 GT336 WITH DISPLAY IDENTITY g3

• GT336’s image-local candidate letters change, while its display identity g3 persists through the accepted match at 110.8 s and continuation at 114.8 s.The source query initializes g3 at 103.0 s, with local alternatives visible at 105.4 s.Source: Pred ID: g3 · p. 20

E PAIRED PEDESTRIAN RECOVERY WITH MISSING CAMERAS

• With four of sixteen MTMMC cameras masked, g1’s association-null probability rises to 0.93 at frame 1842, then falls to 0.06 at frame 1876 as the same identity continues.Cameras c08 and c12 remain available for initialization and later matching while c03 is masked in all three cases.Source: E PAIRED PEDESTRIAN RECOVERY WITH… · p. 20

F OBSERVATION-MISSING SENSITIVITY AND EVENT PREDICTION

• At an 8.8 s mask, HA is 87.76% for the full model and 67.15% for reactive memory; four-camera masking gives full-model HA of 76.46%.Arrival-time absolute error has median 0.71 s and 90th percentile 2.07 s; recursive-model next-camera top-1 accuracy is higher in all three gap bins.Source: F OBSERVATION-MISSING SENSITIVITY AND… · p. 20

G VEHICLE FEEDBACK AND WAITING

• Through three handoffs, recursion retains 95.79% of identities versus 82.24% without recursion, a 13.55-percentage-point difference.This comparison evaluates complete identity chains.Source: G VEHICLE FEEDBACK AND WAITING · p. 20
• At 96.0% correct reacquisition, explicit null reduces forced-match error by 27.09 points.Removing null increases false associations before the correct target appears; removing recursion increases errors after reappearance and missed reacquisitions.Source: G VEHICLE FEEDBACK AND WAITING · p. 20
• The paired pedestrian recovery figure follows g1’s P(null) among current candidates and temporary null, conditional on an active target, through a masked-camera gap and reappearance.Four MTMMC s17 cameras are masked while prior global state and physical routes remain.Source: Figure 8 · p. 21

H IMPLEMENTATION AND LEARNING

• A shared recurrent target state supports forecasting, association, and posterior feedback, with the posterior supplying the state for the next observation prediction.Both accepted matches and waits use weighted state updates; an acceptance threshold controls only the emitted identity decision.Source: H IMPLEMENTATION AND LEARNING · p. 22, §H.3 · p. 23

H.1 CAUSAL ENCODERS AND PERSISTENT STATE

• Frozen CLIP visual and text encoders feed normalized 256-dimensional projections, while each identity stores a recurrent state, camera-region distribution, velocity estimate, and network-presence probability.A 256-unit GRU and two graph-message layers propagate the state using region probabilities, velocity moments, installation features, and elapsed time; masking leaves physical routes unchanged.Source: §H.1 · p. 22

H.2 OBSERVATION-EVENT DISTRIBUTIONS AND CANDIDATE MASS

• The event decoder predicts next-camera probabilities, arrival-time and entry-position mixtures, and continued network presence; availability and detection opportunity determine observable event mass.The remaining mass includes unavailable views, missed detections, and cases outside candidate support; a calibrated target-versus-background log-likelihood ratio contributes to association.Source: §H.2 · p. 22

H.3 CONCURRENT OBSERVATIONS AND COMMITMENT

• Same-time camera posteriors begin from a shared prior and form one mixture update, while candidate and null conditioned states contribute posterior-weighted region and velocity moments to the next recurrent state.The same weights inform match emission; waits still update the posterior, and null states retain identity until sequence end.Source: §H.3 · pp. 22–23

H.4 LOSSES, ROLLOUTS AND MODEL SELECTION

• The learning objective combines event, association, and censored-wait likelihoods, then adds presence and identity supervision.Camera and association use cross-entropy, time and entry use negative log-likelihoods, censored intervals use right-censored time likelihood, presence uses binary cross-entropy, and identity uses supervised contrastive loss.Source: §H.4 · p. 23
• Training windows contain 32 observation updates and at most four handoffs, with gradients passing through handoffs and teacher forcing decreasing to zero over the first 20 epochs.Five seeds are used, checkpoints are selected by validation HOTA, and operating thresholds are frozen before testing.Source: §H.4 · p. 23

I DATASET-SPECIFIC COMPONENT ABLATIONS AND COUNTS

• Removing entry prediction or presence estimation slightly improves one dataset's IDF1 but raises false-match rate, motivating both handoff and trajectory metrics.CityFlowV2 uses the full configuration as its ablation baseline, while MTMMC uses the reference configuration.Source: I DATASET-SPECIFIC COMPONENT ABLATIONS… · pp. 23–24

J COMPLETE PAIRED LANGUAGE CONDITIONS

• Natural descriptions yield 559 correct CityFlowV2 handoffs versus 520 without query and 429 MTMMC handoffs versus 399, while misleading descriptions reduce handoff accuracy below no-query.Partly unobservable attributes favor the full model over hard gating; comparisons use the same tracking events within each dataset.Source: J COMPLETE PAIRED LANGUAGE CONDITIONS · p. 25, Table 7 · p. 25

K WITHIN-DOMAIN CAMERA AND SCENE CONDITIONS

• Under unseen-scene evaluation, the full model reaches 68.65% MTMMC handoff accuracy, while CRTracker retains higher HOTA in matched-scene and unseen-camera settings.The method ranking changes with camera or scene exposure and the identity-continuity metric.Source: K WITHIN-DOMAIN CAMERA AND SCENE… · pp. 25–26, Table 8 · p. 26

L.1 POPULATIONS AND SPLIT CONSTRUCTION

• Evaluation uses scene-disjoint splits, keeping all cameras in a scene together and separating matched-scene, unseen-camera, and unseen-scene conditions.The principal and component-removal studies share fixed event populations; query studies use 637 vehicle and 509 pedestrian events.Source: §L.1 · p. 26

L.2 REFERENCE AND FULL CONFIGURATIONS

• The full configuration adds hypothesis-conditioned candidate and null states with calibrated posterior probabilities, plus recursive training rollouts of up to four handoffs.Later prediction errors supervise preceding posterior updates; the detector, tracker, encoders, candidate construction, camera graph, and test protocol remain unchanged.Source: §L.2 · p. 27

L.3 ELIGIBILITY, ABSENCE DECISIONS AND TRAJECTORY METRICS

• Eligible handoffs require at least 1 s of observed history, a 0.5 s gap, and reacquisition within 5 s; a wrong first commitment counts as failure.Absence decisions use a model-independent 2 Hz grid, and trajectory identity metrics are computed over synchronized camera sequences without cross-camera box matches.Source: §L.3 · p. 27

L.4 BASELINE IDENTITIES AND RESOURCES

• Baselines include causal appearance matching, reactive feature memory, fixed route-and-time priors, and adapted trackers with method-specific representations.Internal controls share causal prefixes and validation-selected thresholds.Source: §L.4 · pp. 27–28

L.5 COMPARISON UNITS AND AGGREGATION

• Handoff accuracy is aggregated over eligible handoffs, false-match rate over absence decisions, and identity metrics over resulting trajectories.Paired sample analysis resamples the same scenes and identity chains for both methods.Source: §L.3 · p. 27, §L.5 · p. 28

M.1 MATCHED COMPONENT INTERVENTIONS

• Matched interventions remove event, arrival, entry, or presence forecasts, block posterior feedback, disable null handling, or remove query evidence before retraining.The recursion intervention retains causal appearance memory and prediction heads; the null intervention forces candidate selection when candidates exist.Source: §M.1 · p. 28

M.2 QUERIES AND OBSERVED ATTRIBUTES

• Five query conditions vary text reliability while holding source history and tracking events fixed, with attributes annotated only from historical frames.The full model scores query-candidate similarity, whereas the hard gate rejects confidently observed contradictions and treats unknown attributes as neutral.Source: §M.2 · pp. 28–29

• 4.61 HA points over appearance-plus-time association for vehicle-to-pedestrian transfer and 5.22 points for the reverse transfer, with false-match rates also decreasing in both directions.The source-trained tracker is evaluated with transferred checkpoints and thresholds; target supervision yields further gains on the same target-domain events.Source: N TRANSFER RESOURCES AND ADDITIONAL… · p. 29, Table 9 · p. 29
• 42 additional correct handoffs in each object domain accompany 85 fewer vehicle and 117 fewer pedestrian false accepts versus one-way event prediction.The comparison also includes appearance-plus-time memory and uses fixed event groups within each domain.Source: Table 10 · p. 29, N TRANSFER RESOURCES AND ADDITIONAL… · p. 30, Table 11 · p. 30

O AUXILIARY SPATIAL OUTPUTS AND METRIC CONVENTIONS

• Relative to the GRU predictor, the full model reduces ADE/FDE by 0.16/0.35 m, raises occupancy IoU by 5.63 points, and lowers Chamfer by 0.03 m.The auxiliary evaluation uses calibrated ground-plane coordinates, six forecast samples over 3 s, and 1000 valid windows.Source: O AUXILIARY SPATIAL OUTPUTS AND METRIC… · p. 30, Table 12 · p. 30

P.1 REFERENCE-CONFIGURATION PAIRED OUTCOMES AND UNCERTAINTY

• In the reference MTMMC comparison, ReWorld-Track and CRTracker correctly accept 2257 and 2243 handoffs out of 2749, with 57 versus 43 discordant events.The paired outcomes determine uncertainty in the difference between correct-handoff rates.Source: §P.1 · p. 31, Table 13 · p. 31
• Retaining null lowers false-match rate by 2.07 points versus recursive threshold but increases mean delay by 0.35 s; on shared accepted events, delays are 1.18 versus 0.91 s.The mechanism study uses 2800 handoffs; missed handoffs count against handoff accuracy, while correct accepts determine mean delay.Source: Table 14 · p. 31, Table 15 · p. 31, §P.1 · p. 32

P.2 FULL-MODEL PRINCIPAL COMPARISON

• 45.36 HOTA for the full MTMMC configuration exceeds CRTracker by 1.24 points on the same test manifest; IDF1 is 58.12 and IDSW is 298.Relative to the reference configuration, the full model adds 40 correct handoffs and removes 47 false accepts.Source: Table 16 · p. 32, §P.2 · p. 32

P.3 PREDICTOR-MATCHED FEEDBACK CONTROLS

• With the predictor held fixed, learned posterior feedback gains 0.94 HOTA points over fixed-moment soft association and lowers false-match rate by 0.63.Adding null to feedback plus threshold lowers handoff accuracy from 84.58 to 83.56 while false-match rate falls from 6.66 to 4.90.Source: Table 17 · p. 32, §P.3 · p. 32

P.4 INDEPENDENT TRAINING RUNS

• 45.36 mean HOTA for ReWorld-Track versus 44.12 for CRTracker across five paired seeds, with a 1.24-point mean difference and two-sided Student-t interval [1.10, 1.38].The interval has four degrees of freedom; the paired differences have sample SD 0.12 on the fixed test set.Source: Table 18 · p. 33, §P.4 · p. 33

P.5 THRESHOLD SENSITIVITY AND WAITING

• Raising the commitment threshold from 0.50 to 0.95 cuts false accepts from 1004 to 266, while correct handoffs fall from 2340 to 2115 and mean delay rises from 0.72 to 1.52 s.Operating thresholds are selected on validation data and remain fixed during test evaluation.Source: §P.5 · p. 33, Table 19 · p. 33

P.6 FIXED-MEMBERSHIP MISSING-CAMERA DURATION STUDY

• At 8.8 s masking, the full model retains 527 correct handoffs (87.83%) versus memory’s 403 (67.17%), with all 600 events kept in the denominator.Figure 3 follows a separate event group at durations of 1, 3, 5, 7, and 8.8 s.Source: §P.6 · p. 33, Table 20 · p. 33

Q.1 COMPONENT INTERVENTIONS AND UNCERTAINTY

• Removing the event prior lowers HA, IDF1, and HOTA by 6.36, 5.51, and 5.34 points, while removing recursion lowers them by 4.82, 3.76, and 4.08 points.Removing null increases HA by 1.49 points and reduces delay by 0.42 s, but FM rises by 7.84 points and IDF1/HOTA fall by 0.58/0.38 points, indicating greater handoff coverage alongside more false identities.Source: §Q.1 · p. 34

Q.2 NATURAL BLIND-GAP DURATION

• Full-model minus-memory HA gains grow from 2.5 to 5.8, 9.7, and 15.5 points across natural MTMMC blind-gap duration bins.Each bin contains a different event group, with methods matched within that bin; all groups favor the evolving event belief over reactive memory.Source: §Q.2 · p. 34, Table 21 · p. 34

Q.3 LANGUAGE UNDER APPEARANCE AMBIGUITY

• Natural queries improve HA by 2.1 points on easy cases and 10.9 on ambiguous cases, while misleading descriptions reduce no-query scores by 5.2 and 6.2 points.Paraphrases remain within 0.5 and 1.3 points of natural queries. For the full MTMMC candidate/null distribution, joint belief ECE is 2.8%, versus 5.0% after temperature scaling and 8.7% for raw similarity; NLL and Brier score are 0.391 and 0.127, versus 0.447 and 0.156 after temperature scaling.Source: §Q.3 · p. 35, Table 22 · p. 35, §Q.4 · p. 35

Q.5 TRACKING-WORKLOAD LATENCY AND SCALE

• Under Full 19-camera, end-to-end mean and P95 latency are 69.4 and 92.6 ms, so at least 95% of updates finish within the 100 ms interval.Full 19-camera retains 1186 gated association pairs versus 412 for Heavy, about 2.9 times as many, accompanying a core mean increase from 25.4 to 38.7 ms.Source: Table 24 · p. 35, §Q.5 · p. 36

Q.6 CAMERA AND TOPOLOGY SHIFT

• Full-model HA decreases from 85.1 to 80.6 and 74.6 as training exposure shifts from seen topology to unseen cameras and unseen topology, while feedback retains gains over both baselines.Gains over reactive memory are 2.8, 3.8, and 6.7 points; gains over one-way prediction are 1.7, 2.3, and 3.8 points.Source: §Q.6 · p. 36, Table 25 · p. 36

R.1 SEPARATING ACTIVE-TARGET ASSOCIATION FROM NETWORK EXIT

• The model distinguishes temporary null association from target exit, and weights non-observation by camera availability and the target’s possible paths.An active target may remain in the network without an observation; when all streams are masked, the no-event likelihood is one, while paths through available cameras with high detection probability lose support if no target is observed.Source: §R.1 · pp. 36–37, §R.2 · p. 37

U POSTERIOR UPDATES, PREDICTION AND CONTINUITY

• ReWorld’s posterior update gains 0.50 HOTA points over a generic learned updater and averages 0.48 points higher across three held-out scenes.It gains a further 0.44 HOTA points over fixed posterior moments. Unlike the generic updater, ReWorld weights candidate- or null-conditioned states by calibrated posterior probabilities and retains region distribution, velocity moments, and network presence for the next prediction.Source: U POSTERIOR UPDATES, PREDICTION AND… · p. 40, §U.1 · p. 40, §U.2 · p. 41

U.3 PREDICTION AFTER THE UPDATE

• ReWorld improves next-camera Top-1 by 1.38 points over the generic updater, while median arrival error falls from 0.78 to 0.71 s.The 90th-percentile arrival error falls from 2.22 to 2.07 s; the forecasts provide the prior for evaluating candidates at the next return.Source: §U.3 · p. 41, Table 29 · p. 41

U.4 IDENTITY THROUGH SUCCESSIVE RETURNS

• 0.27, 1.03, and 2.65 points are the ReWorld-minus-generic identity-retention differences after one, two, and three handoffs, respectively.Table 30 reports identity retention as IR@k (%).Source: §U.4 · p. 41, Table 30 · p. 41
• Fixed-moment feedback improves identity retention over resetting but retains fewer identities than either learned update.The passage does not provide numerical values for these comparisons.Source: §U.4 · p. 41
• Errors in the carried history can influence several associations after the first return.The passage connects the growing predictive-state difference to identity decisions across later handoffs.Source: §U.4 · p. 41