How ductai199x/forensic-knowledge-graph-eccv26 works
This repository currently documents a research framework, not an implemented codebase: the tree contains a README and assets only, and the README says code, pretrained models, and FKG-50K are forthcoming. The documented workflow takes an image through forensic trace extraction and region analysis, constructs a grounded Forensic Knowledge Graph (FKG), then uses Iterative Context Refinement (ICR) with a VLM to produce a justified report. The graph reflects those documented stages and capabilities; paths are null because no implementation paths are present.
Components 15
Forensic analysis
- Input image
- Trace extraction
- Region proposals
- Forensic reasoner
- Authentication findings
- Forgery localization
- Camera provenance
Knowledge graph
- Forensic Knowledge Graph
- Causal dependencies
- Scene links
Grounded explanation
- Iterative Context Refinement
- Explanation VLM
- Justified forensic report
Dataset
- FKG-50K
Other components
- Image submitter
Connections 14
- Image submitter → Input image: submits
- Input image → Trace extraction: analyzes
- Trace extraction → Region proposals: informs
- Region proposals → Forensic reasoner: informs
- Forensic reasoner → Authentication findings: identifies
- Forensic reasoner → Forgery localization: localizes
- Forensic reasoner → Camera provenance: attributes
- Forensic reasoner → Forensic Knowledge Graph: constructs
- Forensic Knowledge Graph → Causal dependencies: encodes
- Forensic Knowledge Graph → Scene links: links
- Forensic Knowledge Graph → Iterative Context Refinement: conditions
- Iterative Context Refinement → Explanation VLM: guides
- Explanation VLM → Justified forensic report: generates
- FKG-50K → Forensic Knowledge Graph: provides ground truth