Roman silver coins no longer need pairwise eyeballing: unsupervised clustering groups them by die, cutting validation comparisons by at least 91.8%
Synopsis
Die studies of ancient coinage require comparing every coin with every other, and the number of comparisons grows with the square of the sample, which has long kept large-scale studies out of reach. An unsupervised pipeline first selects keypoints on coin relief with a Gaussian process, then clusters coins by Bayesian distance microclustering on keypoint dissimilarities, inferring both the number of dies and their assignments. On 1135 Roman silver denarii struck in 64–66 C.E., the obverse sample reached a normalized mutual information of 0.959, and visual validation required at least 91.8% fewer comparisons than an all-pairs review.
Interpretation
Die study of ancient coins moves from manual coin-by-coin comparison to an unsupervised pipeline that proposes candidate die groups for expert verification. Manual study requires at least N(N−1)/2 visual comparisons for N images, more than 1.1 million for 1500 obverses, which makes large coinages infeasible. On 1434 obverse images of Neronian denarii, clustering matched 297 die labels established by manual verification with a normalized mutual information of 0.959.
The pipeline represents each coin as a set of automatically chosen relief keypoints and measures whether two coins came from the same die by how well those keypoints match. General-purpose feature detectors stack keypoints across scales and are not designed for die analysis; here Gaussian-process uncertainty sampling places keypoints on letters, locks of hair and eye contours. Across 2866 images of 1135 coins, the obverse sample reached sensitivity 0.840 and false discovery rate 0.067; swapping in off-the-shelf ORB keypoints within the same pipeline gave clustering indistinguishable from random assignment on large datasets containing many worn specimens.
Perspective
The demonstrated setting is 2866 obverse and reverse images of 1135 Neronian denarii struck in 64–66 C.E., with the three lowest preservation grades making up 26% of the sample; final die attribution still rests on expert verification and correction, and the pipeline is positioned as turning an all-pairs comparison into verification of a small number of candidate clusters.
Coins from the same die but in very different preservation states can have larger dissimilarity than well-preserved coins from different dies, or than coins from different dies by the same engraver, so the distributions overlap; the authors also note that a universal threshold for graph-based clustering is harder to choose for datasets with more, smaller classes.
