A Large-Scale Evaluation of Correspondence-Based Coin Classification on Roman Republican Coinage

Abstract

While being of great use for the numismatic community, the problem of image-based ancient coin classification has shown to be a challenging problem in the past [KZ08, Ara10, ZK12, ZKK14]. This is caused by the high number of classes, the low number of available samples per class, as well as the general conditions of ancient coins. Another critical problem is the availability of training data. Learning-based methods rely on large amounts of training data to capture the variability within classes, and it is shown in [ZKK14] that the learning-based method proposed in [Ara10] is heavily affected by a low number of training samples. Hence, the method proposed in [ZKK14] is based on a coin-to-coin similarity metric from matched local features and thus suffers less from the training data problem. Accordingly, the method clearly outperforms all previously proposed methods for ancient coin classification [KZ08,Ara10,ZK12] when only one reference coin is available per class. Although the results in [ZKK14] show the superiority of the proposed method, they are based on a 60-class dataset and it is unclear how the classification performance is influenced by much larger number of classes. In this work, we aim to fill this gap by empirically evaluating the method's performance by means of a dataset consisting of 418 coin classes.