EGRAPHSEN is a case study on image classification in the humanities, specifically on painter attribution on Attic vase paintings. This study aimed to explore the new perspective that artificial intelligence (AI) can offer when studying traditional methods and heterogeneous domains. When we translate the task (painter attribution), we have to consider the idiosyncrasies of the data domain (Attic vase paintings). This is challenging for both, classical archaeologists and computer scientists. In this paper, we address how to approach the challenges in the creation of the dataset. We carefully selected and prepared the data, reflected on potential biases and trained a convolutional neural network (CNN) accordingly. Specifically, we developed sampling criteria to combat the biases and a hierarchical labelling system to segment the images into details. Our model architecture was designed to process sets of images instead of only one individual image, which enables us to experiment with different combinations of image segments. This forms the basis for an analysis framework, which allows us to go beyond mere painter attribution and to explore the ambiguity of image similarity itself.
Kipke et al. (Fri,) studied this question.