A human-centric software platform where papyrologists keep full control while AI helps rank plausible joins,
preview candidate placements, and document the full reconstruction process in a reproducible way.
Built for actual reconstruction work: recto and verso inspection, ranked candidate joins, manual validation, and documented decision-making in one research-facing workspace.
Recto + VersoSynchronized analysis
Expert-ledHuman remains in control
Project overview
What is JoinPap?
JoinPap is developed to assist papyrologists in reconstructing fragmented documents through a human-centered workflow.
It provides expert-facing tools for arranging fragments, notes, and project state, and incorporates machine-generated matching proposals as decision support.
The interface is designed for real research practice: synchronized views, recto/verso inspection, interactive browsing of suggestions, and fast validation of competing hypotheses.
A joint effort between the Istituto Papirologico "Girolamo Vitelli" in Florence, and the Istituto di Scienza e Tecnologie dell'Informazione "A. Faedo" of the National Research Council (ISTI-CNR), Pisa.
Access points
Resources
Access the core software, the companion AI pipeline, and project information from a single landing page.
JoinPap Software
The main application for interactive reconstruction, with dedicated tools for fragment management, split recto/verso visualization, and annotation of reconstruction decisions.
Deep-learning pipeline that scores candidate joins by combining geometric boundary compatibility and visual continuity (fibers/ink), then exports results for JoinPap.
Some visual examples of the JoinPap interface in action.
Model workflow
AI Tool Interface & Inference
The model evaluates a potential positioning between two fragments by scoring one main touch-point patch plus nearby boundary patches, then averaging them into a match score for that positioning.
Scores from recto and verso are combined on a shared grid to rank candidate placements and fragment pairs; on our validation dataset, the classifier reaches strong separation between plausible and implausible joins (89.4% accuracy).
Following, some visual examples of the AI tool interface in action:
Research context
Funding & Research Project
This research is funded within the PRIN PNRR 2022 project
“Reconstructing Fragmentary Papyri through Human-Machine Interaction”.
Learn more on the project website:
www.joinpap.unifi.it.
