IberianVoxel: Automatic Completion of Iberian Ceramics for Cultural Heritage Studies

Pablo Navarro, Celia Cintas, Manuel J. Lucena, José Manuel Fuertes, Antonio Jesús Rueda, Rafael Jesús Segura, Carlos J. Ogáyar-Anguita, Rolando González-José, Claudio DelrieuxIJCAI 2023 · DOI

TL;DR

This paper introduces IberianVoxel, a novel 3D Autoencoder Generative Adversarial Network (3D AE-GAN) that automatically reconstructs missing parts of fragmented Iberian ceramics. By learning from a large dataset of 3D voxelized pottery, IberianVoxel can generate realistic completions, aiding archaeologists in their analysis and preservation efforts.

What Problem Does This Paper Address?

Archaeological excavations frequently unearth fragmented pottery, which are crucial for understanding past cultures, trade routes, and chronological periods. However, reconstructing these fragments is a labor-intensive, manual process that can further damage delicate artifacts. Existing computational methods often rely on 2D representations or struggle with realistic fragment handling. This paper addresses the need for an automated, 3D approach to reconstruct fragmented ceramics, preserving their integrity and providing a more efficient analysis tool for archaeologists.

Methods

The core of IberianVoxel is a 3D Autoencoder Generative Adversarial Network (3D AE-GAN). The generator (G) takes a voxelized fragment as input and, through an encoder-decoder architecture, generates the missing portion of the pottery. The discriminator (D) then learns to distinguish between real and generated complete pottery.

Key methodological contributions include:

  • 3D AE-GAN Architecture: A specialized GAN designed for 3D voxel data, capable of learning and generating complex ceramic shapes.
  • Artificial Fragment Generation: A procedure adapting the Discrete Voronoi Chain algorithm to create realistic 3D fragments from complete voxelized models. This allows for the generation of a diverse dataset of fragmented pottery for training and evaluation.
  • Data Pipeline: A process to convert 2D pottery profiles into 3D voxelized models by employing a "spin" technique and then fragmenting these models.

Datasets

The paper introduces a new benchmark dataset:

  • IberianVoxel Dataset: A collection of 1001 voxelized 3D Iberian pottery models. These models were fragmented 16 times each, resulting in a total of 16016 fragmented models with corresponding complete pottery references. The fragments are categorized by size (15-20%, 20-30%, and 30-100% of the original ceramic volume).

Key Insights and Ideas

  • Bridging the Gap between 2D and 3D: The paper successfully moves beyond 2D profile analysis to a full 3D reconstruction, offering a more comprehensive understanding of ceramic forms.
  • Generative Power for Archaeology: The use of GANs demonstrates their potential not just for artistic generation but also for practical applications in cultural heritage, specifically in artifact reconstruction.
  • Fragment Size Robustness: The model shows consistent performance across various fragment sizes, indicating its adaptability to different levels of damage.
  • Archaeologist Validation: The inclusion of domain expert evaluation provides crucial qualitative validation, ensuring the generated reconstructions are not only geometrically plausible but also culturally relevant.

Key Results

  • Quantitative Performance: The averaged Mean Squared Error (MSE) was 0.06, and the Dice Similarity Coefficient (DSC) was 0.70 across all classes and fragment sizes, indicating good reconstruction quality.
  • Class-Specific Performance: Certain pottery classes (e.g., Class 10 and 11, identified as open shapes like plates) showed improved reconstruction, potentially due to a larger number of examples in the dataset.
  • Shape Classification: A binary shape classifier (open/closed) performed similarly on both original and reconstructed pottery, suggesting the generated completions retain essential shape characteristics.
  • Expert Evaluation: Archaeologists rated the quality of reconstructed pottery with a mean score of 2.09 (out of 3) and assessed their Iberian style with a mean score of 3.93 (out of 5), indicating satisfactory results.

Limitations and Future Work

  • Asymmetry: Some generated samples exhibit a lack of perfect symmetry, which the authors attribute to the limited examples in certain classes. Future work could involve post-processing steps or penalization factors in the loss function to address this.
  • Irregularities: Occasional irregularities in reconstruction or fragment alignment were observed. Morphological operations are suggested as a potential post-processing solution.
  • Visualization Tools: Experts suggested improvements in visualization tools, such as scale factors and better edge inspection, to aid in evaluation.
  • Generalizability: While the framework is presented as generalizable, further testing on different types of brittle artifacts (e.g., projectile points, ancient buildings) would be beneficial.

Summary updated

BibTeX
@inproceedings{navarro2023iberianvoxel,
  author = {Pablo Navarro and Celia Cintas and Manuel J. Lucena and José Manuel Fuertes and Antonio Jesús Rueda and Rafael Jesús Segura and Carlos J. Ogáyar-Anguita and Rolando González-José and Claudio Delrieux},
  title = {{IberianVoxel: Automatic Completion of Iberian Ceramics for Cultural Heritage Studies}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/647},
}
Figure