Period Classification of 3D Cuneiform Tablets with Geometric Neural Networks

Clay tablets are the oldest handwritten documents using the cuneiform script named after the wedge shaped imprints of a rectangular stylus left in their surface. Therefore the most suitable documentation technique is 3D acquisition using e.g. structured light. The tablets are heterogeneously shaped and may contain damage in varying degrees. Convolutional neural networks enabled large advances in the analysis of historical script written with ink on paper. However, cuneiform tablets remain inaccessible to common raster-image based recognition methods for handwriting, because the characters are actually represented by 3D shapes. With the release of the Heidelberg Cuneiform Benchmark Dataset (HeiCuBeDa) a plethora of machine learning opportunities for the underlaying representation, 3D surface meshes, was made accessible. Additionally HeiCuBeDa contains transliterations and meta-data such as language-type for approx. 1/3 of the 1.977 tablets. In this work, we combine and adapt the convolution operation of SplineNet with the pooling from PointNet++ to predict the time-period of a tablet directly on basis of its mesh representation. The classification tasks were performed on 336 tablets for training and 158 tablets for testing. Our proposed approach reaches a classification accuracy of 84 % on 4 time-period classes.