Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A novel network for classification of cuneiform tablet metadata

Frederik Hagelskjær

TL;DR

A convolution-inspired architecture that gradually down-scales the point cloud while integrating local neighbor information is developed, and the final down-scaled point cloud is then processed by computing neighbors in the feature space to include global information.

Abstract

In this paper, we present a network structure for classifying metadata of cuneiform tablets. The problem is of practical importance, as the size of the existing corpus far exceeds the number of experts available to analyze it. But the task is made difficult by the combination of limited annotated datasets and the high-resolution point-cloud representation of each tablet. To address this, we develop a convolution-inspired architecture that gradually down-scales the point cloud while integrating local neighbor information. The final down-scaled point cloud is then processed by computing neighbors in the feature space to include global information. Our method is compared with the state-of-the-art transformer-based network Point-BERT, and consistently obtains the best performance. Source code and datasets will be released at publication.

A novel network for classification of cuneiform tablet metadata

TL;DR

A convolution-inspired architecture that gradually down-scales the point cloud while integrating local neighbor information is developed, and the final down-scaled point cloud is then processed by computing neighbors in the feature space to include global information.

Abstract

In this paper, we present a network structure for classifying metadata of cuneiform tablets. The problem is of practical importance, as the size of the existing corpus far exceeds the number of experts available to analyze it. But the task is made difficult by the combination of limited annotated datasets and the high-resolution point-cloud representation of each tablet. To address this, we develop a convolution-inspired architecture that gradually down-scales the point cloud while integrating local neighbor information. The final down-scaled point cloud is then processed by computing neighbors in the feature space to include global information. Our method is compared with the state-of-the-art transformer-based network Point-BERT, and consistently obtains the best performance. Source code and datasets will be released at publication.
Paper Structure (13 sections, 5 equations, 2 figures, 6 tables)

This paper contains 13 sections, 5 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Network Structure
  5. Neighbor Features
  6. Network Training
  7. Fine tuned Point-BERT
  8. Experiments
  9. Period Classification
  10. Seal and Left Side Sign
  11. Tablet Front
  12. Ablation
  13. Conclusion

Figures (2)

  • Figure 1: The network structure of our presented method. First a point cloud is sampled from the CAD model. The point cloud is then processed and down-sampled. The orange layers compute neighbors in spatial space. In the blue layer neighbors are found the feature space. The tablet is "HS 2274" mara2019breaking.
  • Figure 2: Three tablets shown from the left side ("HS 1238", "HS 2271", "HS 2274"). The larger curvature on the left side is seen, except for tablet "HS 2274" which is wrongly oriented. On the two leftmost tablets the left side sign is present.