GARF-CH: Generalizable Flow-Matched 3D Reassembly of Fragmented Cultural Heritage Point Clouds

Abstract. Digital restoration of cultural heritage (CH) assets is essential for preserving historical structures fragmented by time, environmental degradation, or catastrophic events. Although learning-based approaches have demonstrated impressive results in 3D fragment reassembly for everyday objects and small artifacts, their application to large-scale CH structures is largely unexplored. To overcome this gap, this study examines the feasibility and reliability of neural point cloud reassembly methods in complex CH scenarios. Building upon the state-of-the-art GARF framework, we introduce GARF-CH, a parameter-efficient adaptation of GARF tailored to architectural-scale heritage data. We have built a novel fractured dataset derived from the ArCH dataset by generating realistic fractures on large-scale CH point clouds. This enables a systematic evaluation of these methods in the CH domain. Extensive quantitative and qualitative experiments on 15 real-world CH structures show that GARF-CH reduces rotation error by 1.76^◦, translation error by 0.3×10⁻², and Chamfer Distance by 0.71×10⁻³ compared to the original GARF model, while maintaining the same part accuracy (83.33%). Furthermore, we assess robustness under missing-fragment conditions and show that the proposed approach maintains coherent reconstructions even with incomplete data. These results highlight the potential and limitations of learning-based reassembly for large-scale CH restoration and establish a foundation for future research on scalable, robust, and autonomous digital reconstruction pipelines for complex CH environments.