Research Boundaries

Research Boundaries

This page records what the current project evidence supports and what it does not.

Supported by current real-data experiments

The real MEVA study supports statements about:

• local activity-evidence recognition on the evaluated bounded subset;
• actor-crop versus full-frame behavior under the controlled ablation;
• frozen versus fully fine-tuned backbone behavior;
• class-imbalance strategy comparisons across three seeds;
• calibration diagnostics;
• class-wise errors and source-video dependence.

Not yet supported as a real end-to-end claim

The current evidence does not establish real-world end-to-end superiority for:

• cross-camera identity fusion;
• temporal incident reconstruction across a real camera network;
• graph-level contradiction resolution on real evidence;
• real sensor-gap reasoning;
• graph-level uncertainty under real distribution shift;
• a learned end-to-end IncidentGraph model against directly comparable SOTA baselines.

Oracle localization

The actor-centric experiment uses annotation-derived actor tubes. Report it as oracle localization.

Synthetic benchmark

The 60-incident diagnostic verifies software/evaluation semantics under generated ground truth. Its graph score must remain labeled synthetic/controlled.

Strongest next validation

A high-value real end-to-end study would compare, under the same real camera evidence:

time-only
→ appearance-only
→ naive fusion
→ fusion without provenance
→ fusion without uncertainty
→ complete IncidentGraph

with real identity association, event linking, graph-level metrics and camera/sensor stress tests.