Skip to content
Anubha Parashar edited this page Aug 10, 2026 · 2 revisions

IncidentGraph Wiki

Welcome to the IncidentGraph project Wiki. This documentation covers the evidence model, actor-centric activity recognition, cross-camera association, provenance-preserving graph reconstruction, contradiction handling, missing evidence, calibration, real-MEVA experiments, the controlled graph diagnostic, reproducibility, testing and responsible use.

Important

The current real-data study validates the local activity-evidence component. The complete graph pipeline is additionally tested with a controlled synthetic diagnostic. Do not present the synthetic graph score as real end-to-end multi-camera performance.

Start here

  1. Getting Started
  2. Architecture
  3. Evidence Model
  4. Activity Recognition
  5. Cross-Camera Association
  6. Graph Reconstruction
  7. Provenance and Uncertainty

Main capabilities

  • heterogeneous observation records for RGB, thermal, depth and other evidence sources;
  • actor-centric local activity recognition;
  • reliability-conditioned cross-camera association;
  • typed entity and event graph construction;
  • non-empty provenance on accepted event edges;
  • explicit contradiction groups and alternative hypotheses;
  • explicit missing-evidence objects;
  • classifier calibration and uncertainty reporting;
  • controlled system benchmarking under shared graph/evaluation interfaces;
  • real-data and synthetic-diagnostic claim boundaries.

Wiki map

System design

Recognition and experiments

Software and operations

Governance and planning

Caution

Do not publish private camera footage, personal identifiers, access credentials, restricted dataset material, or outputs that could expose individuals.

IncidentGraph · evidence lineage preserved · conflicts retained · missing evidence remains unknown

Clone this wiki locally