Metric Learning for Ancient Coin Identification¶
, Jason B. Forsyth, , and
IEEE International Conference on Machine Learning and Applications (ICMLA 2026) · Oct 5, 2026 To appear
JMU faculty highlighted in ; student co-authors in .
Abstract¶
Ancient coin identification is an instance-level recognition problem with a difficult data regime: large public collections provide many coin identities, but usually only one image per physical coin, while confirmed same-coin pairs from independent sources are scarce. Because such paired datasets are costly and likely to remain small, we ask how they can best be combined with abundant single-image public data. We test whether scarce paired data can be used effectively as an adaptation signal for tuning augmentations, which then provide synthetic repeated views of the public single-image identities used to train an embedding model. Our ArcFace-based pipeline tunes augmentation hyperparameters on repeated-image validation data and trains on public coin identities. For controlled evaluation, we introduce a curated dataset of 3,708 images covering 309 coin sides under varied lighting and reproduction conditions, including print–scan catalog-style artifacts. In this controlled setting, direct paired training achieves 98.01% Recall@1, while augmentation-guided public-data training achieves 94.57% without using paired identities as supervised training classes. Ablations identify augmentation tuning and preservation of intermediate spatial features as the principal contributors to public-data performance.
Source code: https://github.com/COIN-Research-Group/coin-embedding
BibTeX¶
@inproceedings{sprague2026metric,
author = {Nathan R. Sprague and Jason B. Forsyth and Trevor Schonbrun and Dhanshree D. Atre},
title = {{Metric Learning for Ancient Coin Identification}},
booktitle = {IEEE International Conference on Machine Learning and Applications (ICMLA 2026)},
year = {2026},
note = {To appear}
}