Find your next learning step
A tracking graph must decide not only which fragments connect, but how many trajectories should exist. Birth and death terms make that hidden count decision visible and testable.
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 path-count costs
Route 2 · Work the count ledger
Route 3 · Test priors and circulation
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE PATH-COUNT COSTS
1. Let the graph choose how many paths exist
A tracking graph needs more than transition edges. It must decide when a trajectory begins, when it ends, which detections are used and how many object paths exist. Birth and death terms turn those choices into part of the optimisation.
Without them, a solver may link every fragment into a few implausibly long paths or start a new object at every weak detection. Object count is therefore an outcome of the declared costs and constraints, not a free truth returned by the algorithm.
CHAPTER 2 OF 12 · DEFINE PATH-COUNT COSTS
2. Separate four cost families
Use detection-selection, transition, birth and death costs. A fragmentation penalty may be expressed through repeated birth/death charges or an explicit higher-order term. Keep the sign convention visible because some formulations use negative rewards for selecting confident detections.
| Cost family | Encourages when low | Failure when mis-set |
|---|---|---|
| birth | starting a path | too cheap fragments; too costly misses entries |
| death | ending a path | too cheap fragments; too costly forces links |
| transition | continuing identity | poor calibration creates switches |
| detection selection | using a node | weak clutter or missed true objects |
CHAPTER 3 OF 12 · WORK THE COUNT LEDGER
3. Work a count-selection ledger
Suppose two tracklets A and B are separated by a short gap. Selecting each tracklet gives a reward of 8, represented as cost −8. Starting a path costs 5, ending it costs 5, and linking A→B costs 3.
One combined path costs 5 − 8 + 3 − 8 + 5 = −3. Two separate paths cost 2 × (5 − 8 + 5) = 4. The lower-cost solution selects one object path. If the transition cost rises to 12, the combined path costs 6; two separate paths cost 4 and the inferred count becomes two.
| Scenario | One linked path | Two paths | Selected count |
|---|---|---|---|
| transition cost 3 | −3 | 4 | 1 |
| transition cost 12 | 6 | 4 | 2 |
| birth and death each 1 | −11 | −12 | 2 |
CHAPTER 4 OF 12 · WORK THE COUNT LEDGER
4. Read birth and death as priors
Birth cost can encode how plausible it is for an object to appear at a location and time. Death cost can encode exits, occlusions or the end of sensor coverage. These are priors or penalties, not observed identity evidence.
Scene boundaries, ports, roads and acquisition gaps may justify spatially varying terms. Document them. A map-aware prior can improve tracking, but it can also erase unexpected events if made too strong.
CHAPTER 5 OF 12 · WORK THE COUNT LEDGER
5. Distinguish fragmentation from multiplicity
Two paths may mean two objects or one fragmented object. The graph decides using available links and costs. Penalising every birth and death reduces fragmentation, but excessive penalties can merge distinct objects.
Report both estimated count and path continuity. A low identity-switch score with the wrong number of objects is not a complete success. Count error, fragmentation and false paths reveal different failures.
CHAPTER 6 OF 12 · TEST PRIORS AND CIRCULATION
6. Changed context: open a scene boundary
In the first ledger, place B at a known entry boundary and reduce its birth cost from 5 to 1. Two paths now cost (5 − 8 + 5) + (1 − 8 + 5) = 0, still above the linked cost −3. If transition cost is 6, the linked path costs 0 and the two-path solution ties.
The transfer check shows how a geographic prior can change an ambiguous count decision without changing appearance or motion. Ties should be surfaced; arbitrary solver tie-breaking can look like certainty.
CHAPTER 7 OF 12 · TEST PRIORS AND CIRCULATION
7. Use circulation when count is unknown
Minimum-cost flow formulations may require a search over the number of objects. Minimum-cost circulation can integrate path count into the optimisation by allowing negative-cost cycles corresponding to beneficial trajectories.
Wang, Wang and Yu describe this motivation explicitly. The formulation can be globally solvable while still being wrong if detection rewards, priors or graph gates are misspecified.
CHAPTER 8 OF 12 · TEST PRIORS AND CIRCULATION
8. Calibrate across sensor regimes
A birth prior learned from dense video may fail in satellite revisits where objects appear after long acquisition gaps. Death at the image edge differs from disappearance under cloud or a sensor handoff. Stratify by view geometry, resolution and revisit interval.
Use Tracklet Graphs, Global Flow and Nonlocal Identity Decisions for the transition graph. The present question is how path creation and termination alter count.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
Birth and death costs are not biological event probabilities unless explicitly calibrated. A negative total path cost does not guarantee a real object. Count selection can be highly sensitive to a shared offset applied to all detections.
The ledger assumes independent linear costs. Group motion, split–merge events and mutual exclusion may require higher-order constraints. Tracklet errors remain inside nodes unless the model permits them to be broken.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Build an evidence check
Store every graph node and edge, detection reward, birth/death term, transition cost, capacity and selected cycle or path. Sweep birth and death costs over a plausible range and plot count, fragmentation and false-path rates.
Test boundary, occlusion, clutter and long-gap strata separately. Compare a fixed-count flow, count-search flow and circulation formulation on identical detections. Inspect near-optimal alternatives when count changes under a small cost perturbation.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Questions readers should ask
Is the inferred path count the object count?
It is an estimate under the graph model. Fragmentation and missed detections can make the two differ.
Why not make births very expensive?
That suppresses fragments but can miss real entries or force unrelated tracklets together.
What should happen at an exact tie?
Retain or report ambiguity where possible. A deterministic tie-break is not additional evidence.
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 the tracklet-graph guide above, Identity Confidence, Deferred Assignment and Retrospective Relinking and Cluster Size Thresholds, Boundary Uncertainty and Split–Merge Events.
Primary research routes include Zhang, Li and Nevatia’s global association using network flows, Pirsiavash, Ramanan and Fowlkes’ globally optimal tracking formulation, and Wang, Wang and Yu’s minimum-cost circulation framework. The count-selection ledger is an original teaching example.
Top · Contents · Previous · How X Works