Automatic Reconstruction of Glass Relics Using Manifold Learning

Abstract

The risk imposed by manual reassembly on valuable broken relics necessitates automating this process by leveraging computer vision for 3D data acquisition, and data science to extract features of interest from high dimensional data. This thesis proposes a solution for the automatic reassembly of broken glass relics. The solution first relies on digitizing the broken shards. After that, contours of shards are extracted and segmented. Next, the proposed system maps the segments into the space of manifolds to quantify the similarity in their local geometry and uncover pairwise matches among them. Finally, a global optimization step finds the overall solution of the reassembly problem and the shards are aligned to visualize a digitally reassembled relic. This digital solution is then used in an application that runs on a head-mounted AR device to guide users through the process of sequentially reconstructing the real relic. The focus of this thesis is on the reconstruction part of the problem. The proposed system is discussed and verified over a dataset of broken glassware. Experiments on ten manually broken glass relics validate the success of the proposed approach by estimating the correct position of each shard in the reassembled relic. Moreover, the performance of the proposed system was tested for two extreme scenarios: missing shards and intruder shards. The system showed robust performance in finding matches among shards, however, the alignment of shards around missing pieces was effected.