Search papers, labs, and topics across Lattice.
A novel neural network architecture is introduced for classifying cuneiform tablet metadata from 3D point cloud representations, addressing the challenge of limited annotated data and the need for automated analysis. The architecture uses a convolution-inspired approach for downscaling the point cloud while integrating local neighbor information, followed by feature-space neighbor computation to incorporate global context. Experiments demonstrate that this method outperforms the transformer-based Point-BERT on this task.
A convolution-inspired network beats transformers at classifying cuneiform tablet metadata from 3D point clouds, despite limited labeled data.
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.
