Find your next learning step
The best local match can be the wrong trajectory. When identity evidence is nearly tied, deferring the decision or retaining alternatives can be more truthful than forcing a permanent label.
Open the chapter contents · Remote Sensing root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Define confidence and deferral
Route 2 · Work the ambiguity ledger
Route 3 · Test low confidence and repair
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE CONFIDENCE AND DEFERRAL
1. Identity confidence belongs to a link
A detection confidence estimates whether an observation contains the target class. Identity confidence asks a different question: how strongly does the evidence support linking this observation to a particular track? High detection confidence can coexist with ambiguous identity.
Keep separate scores for detection, motion compatibility, appearance similarity and final association. A single number makes it difficult to discover why an identity was accepted.
CHAPTER 2 OF 12 · DEFINE CONFIDENCE AND DEFERRAL
2. Deferred assignment preserves options
When two links are nearly tied, an online tracker can mark the identity unresolved and wait for later evidence. Deferred assignment avoids locking one uncertain choice into every later frame. It trades immediate completeness for fewer irreversible switches.
Retrospective relinking revisits earlier fragments after future observations become available. It must be labelled as offline inference, because an online system did not possess that future evidence at the original time.
| Decision state | Evidence position | Appropriate output |
|---|---|---|
| decisive | one candidate clearly passes gates and score margin | assign with calibrated confidence |
| ambiguous | several candidates remain plausible | defer or retain alternatives |
| impossible | no candidate passes physical gates | start a new track or leave unmatched |
CHAPTER 3 OF 12 · WORK THE AMBIGUITY LEDGER
3. Work an ambiguity ledger
Track A can link to detection X with cost 0.31 or Y with cost 0.33, where lower is better. Both pass motion and appearance gates. A greedy rule chooses X, but the margin is only 0.02. The next observation gives continuation costs of 0.18 from Y and 0.62 from X.
A deferred two-step decision selects Y with combined cost 0.51 rather than X with 0.93. The later frame did not change the earlier pixels; it changed which trajectory explanation is coherent.
| Hypothesis | First-link cost | Continuation cost | Two-step total |
|---|---|---|---|
| A→X | 0.31 | 0.62 | 0.93 |
| A→Y | 0.33 | 0.18 | 0.51 |
CHAPTER 4 OF 12 · WORK THE AMBIGUITY LEDGER
4. Calibrate rather than decorate confidence
If links reported at 0.8 confidence are correct only half the time, the score is not calibrated. Bin predicted confidence and compare it with observed correctness on held-out sequences. Stratify by gap length, object density and sensor condition.
Ranking and calibration are different. A score can rank candidates well while overstating certainty. Operational thresholds should reflect both discrimination and the cost of false merges.
CHAPTER 5 OF 12 · WORK THE AMBIGUITY LEDGER
5. Keep hypotheses alive without duplicating truth
Multi-hypothesis tracking can preserve several association branches until evidence separates them. Probabilistic association can distribute belief across candidates. These methods represent uncertainty; they do not mean the physical object had multiple identities.
Prune branches with documented rules and retain the discarded mass or reason. Otherwise confidence may rise simply because alternatives were deleted.
CHAPTER 6 OF 12 · TEST LOW CONFIDENCE AND REPAIR
6. Changed context: lower the detector score
Suppose Y’s detection confidence falls after haze, but its motion and trajectory continuation remain coherent. Rejecting it solely for low detector confidence breaks the track. Accepting every low-confidence box, however, raises false links.
This transfer check motivates staged association: strong detections establish stable links, while weaker detections are considered under tighter trajectory gates. The dormant-track guide supplies the long-gap reachability logic.
CHAPTER 7 OF 12 · TEST LOW CONFIDENCE AND REPAIR
7. Separate online repair from retrospective repair
Online repair can reconnect a dormant track using only information available so far. Retrospective relinking can split or merge tracklets after a complete sequence is processed. Both can be useful, but their latency and evidence budgets differ.
Archive the original online output before offline repair. Otherwise a polished final trajectory can conceal how often the live system was uncertain or wrong.
CHAPTER 8 OF 12 · TEST LOW CONFIDENCE AND REPAIR
8. Evaluate the decision policy
Report ID switches, fragmentation, false merges, missed relinks and unresolved duration. Add coverage: what fraction of eligible observations received an identity? A conservative policy may look precise because it leaves hard cases unassigned.
Score online and retrospective outputs separately. Stratify by confidence band and verify whether deferral actually improves later decisions rather than merely postponing errors.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and misconceptions
Confidence is not probability unless calibrated against relevant data. Appearance embeddings can drift across illumination, view and sensor changes. Motion models can be overconfident when registration or manoeuvres are uncertain.
The worked costs are illustrative and additive. Real trackers use learned similarities, assignment constraints and lifecycle rules. Retrospective relinking can also create false certainty if future fragments are themselves misdetected.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Build an evidence check
Keep per-candidate gates, component scores, margins, alternative hypotheses and decision times. Construct sequences with verified identity across occlusion, low confidence and crossings. Replay them under immediate, deferred and retrospective policies.
Use the object-persistence guide for lineage rules and Temporal Frame Ties, Event Steps and Seasonal Aliasing Across Sensors when clock alignment affects trajectory continuity.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Is detection confidence identity confidence?
No. One scores the observation; the other scores a particular association.
Is deferral always safer?
No. It costs latency and can still fail if later evidence is weak. Its value must be measured.
May an offline system rewrite IDs?
Yes, if retrospective processing is part of the declared method and the original online state is preserved for evaluation.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Continue through verified routes
Begin with the Remote Sensing root, Gold Standard of Education and How X Works. Continue through dormant-track reconnection, object persistence and temporal ties above, plus Cluster Size Thresholds, Boundary Uncertainty and Split–Merge Events.
Primary research routes include Bae and Yoon’s confidence-based data association, Liu and colleagues’ confidence-score appearance learning and hierarchical association, and Zhang and colleagues’ video-SAR re-identification with multi-stage association. The two-step ambiguity ledger is an original teaching example.
Top · Contents · Previous · How X Works