Reconstructing the past: Applying Deep Learning to Reconstruct Pottery from Thousands Shards

Abstract

A great deal of time, patience, and effort are required to excavate pottery. For example, archaeologists dig hundreds to thousands of pottery shards from an excavation site. However, restoring pottery is a time-consuming and challenging process, requiring considerable amounts of expertise, experience, and time. Therefore, computer-assisted restoration methods are indispensable to assist the pottery restoration process. However, existing restoration approaches mostly resort to heuristic-based approaches, which are computationally expensive to match and align different shards together. It is often infeasible to handle and process a large number of shards to reconstruct pottery in 3D. In this paper, we propose a deep learning-based pottery restoration algorithm to classify a pottery shard to a specific pottery type and further predict the exact shard location in the pottery type. We use a novel 3D Convolutional Neural Networks and Skip-dense layers to achieve these objectives. Our model first processes a 3D point cloud data of each shard and predicts the shape of the pottery, which a shard possibly belongs to. We first apply Dynamic Graph CNN to effectively perform learning on 3D point clouds of shards and use Skip-dense layers for a classifier. In particular, we generate features from the 3D scanned point cloud of each shard using spatial transform and edge convolution, then classify shards into one of the pottery shape types using Skip-dense. We achieve 98.4% of classification accuracy over 5 different pottery types and 0.032 RMSE for shard location prediction.