FAQ

FAQ

Is IncidentGraph a single neural network?

No. It is a framework and graph contract that can combine local recognizers, association logic, confidence calibration, contradiction handling and evidence provenance.

Is the reported graph score a real MEVA end-to-end score?

No. The graph score comes from the controlled 60-incident synthetic diagnostic.

What is the primary real-data result?

For the reported three-seed comparison, the unweighted actor-centric fully fine-tuned model has the strongest mean result: accuracy 0.2134 ± 0.0399, macro-F1 0.0718 ± 0.0127, weighted-F1 0.1414 ± 0.0290, and macro-mAP 0.1517 ± 0.0203.

Why is the macro-F1 low?

The evaluated subset is strongly imbalanced and several classes remain poorly recognized. The project reports these limitations rather than hiding them behind accuracy or a best-seed result.

Does actor cropping mean the system has solved detection?

No. The reported actor crop uses annotation-derived tubes and is an oracle-localization condition.

Why keep contradictory hypotheses?

Because deleting losing claims makes the final reconstruction harder to audit and can hide real uncertainty in the evidence.

Why represent missing evidence explicitly?

Because a camera gap means the evidence is unknown, not that an event did not occur.

Can I add a stronger activity or re-identification model?

Yes. The architecture is designed so that local models can be replaced while preserving the neutral observation/event interfaces and graph contract.

Can raw MEVA videos be uploaded to the repository?

Only if the dataset license explicitly permits that redistribution. The public project design assumes raw video and restricted data remain outside normal Git tracking.