Healing Stones

Fragmented cultural heritage artifacts — such as Mayan stelae displaced from their original sites — exist today as scattered, incomplete pieces that are nearly impossible to physically reassemble. This project builds an AI-driven virtual reconstruction pipeline for Stela #43 from the Naranjo archaeological site, operating on 3D-scanned .PLY fragment data. The system combines a Point Transformer V2 backbone for geometric feature learning, a Siamese network for fracture surface matching, a Graph Transformer for globally consistent assembly reasoning, and a Trimmed ICP alignment module with pose graph optimization. A synthetic fracture dataset with realistic surface degradation simulation addresses the limited availability of real labeled data. Every learned component has a geometric fallback, and uncertainty quantification ensures low-confidence matches are flagged for human expert review rather than committed automatically. Deliverables include: a modular PyTorch reconstruction pipeline, a reusable synthetic fracture data generation tool, trained model checkpoints, a robust coarse-to-fine alignment module, a global pose optimization module, a full quantitative evaluation report comparing results against the existing geometric baseline, and complete technical documentation. The system is designed to augment archaeologist judgment — not replace it — by pre-screening thousands of potential fragment matches and providing interpretable, auditable match explanations.