MatchMakerNet: Enabling Fragment Matching for Cultural Heritage Analysis

Ariana M. Villegas-Suarez, Cristian Lopez, Ivan Sipiran; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 1632-1641

Automating the reassembly of fragmented objects is a complex task with applications in cultural heritage preservation, paleontology, and medicine. However, the matching subtask of the reassembly process has received limited attention, despite its crucial role in reducing the alignment search space. To address this gap, we propose MatchMakerNet, a network architecture designed to automate the pairing of object fragments for reassembly. By taking two point clouds as input and leveraging graph convolution alongside a simplified version of DGCNN, MatchMakerNet achieves remarkable results. After training on the Artifact (synthetic) dataset, we achieve an accuracy of 87.31% in all-to-all comparisons between the fragments. In addition, it demonstrates robust generalization capabilities, achieving 86.93% accuracy on the Everyday (synthetic) dataset and 83.03% on the Puzzles 3D (real-world) dataset. These findings highlight the effectiveness and versatility of MatchMakerNet in solving the matching subtask.