Generating Cuneiform Signs with Cycle-Consistent Adversarial Networks

The efficient use of machine learning methods in the context of word spotting on cuneiform datasets is usually limited by the number of annotated training samples available. In view of the large amount of required data, their manual generation is a tedious task for 2D and 3D datasets. In this work we present an automatic approach for generating training samples using Generative Adversarial Networks for Domain Adaptation. Our approach adapts between the visual domains of hand-drawn cuneiform autographs and 2D-projections of 3D-scanned cuneiform tablets without any class information. This way, promising results are achieved, as shown in our experiments.