Computational Visual Ceramicology: Matching Image Outlines to Catalog Sketches

Barak Itkin, Lior Wolf, Nachum Dershowitz

Abstract

Identifying archaeological pottery fragments (potsherds) is a labor-intensive task traditionally requiring expert comparison with vast paper catalogs. This paper presents OutlineNet, a deep learning framework that identifies sherds from single smartphone photos by matching 2D fracture outlines to standardized catalog sketches. By leveraging a synthetic data pipeline and a novel "CareLoss" function to address class imbalance, the system successfully narrows down a sherd's origin from hundreds of possibilities, facilitating rapid field identification without the need for expensive 3D scanning.

What Problem Does This Paper Address?

Pottery identification is a cornerstone of archaeological dating, yet most unearthed fragments are small, undecorated, and difficult to classify. The standard practice involves manually comparing these "potsherds" to massive, often non-digitized catalogs of 2D sketches representing complete vessel types.

A potsherd photo with a checkerboard for scale and a semi-automatically traced fracture outline.

Several challenges hinder automation in this domain:

  • Field Constraints: Active excavations require fast, low-cost tools; high-fidelity 3D reconstruction is currently too resource-intensive for field use.
  • Data Scarcity and the "Sim-to-Real" Gap: There is a lack of labeled real-world sherd imagery. Models must bridge the gap between noisy, broken fragments in photos and the clean, idealized sketches found in catalogs.
  • Class Imbalance: Common pottery types dominate datasets, leading standard models to neglect rare but historically significant classes to optimize overall accuracy.

Methods

The authors introduce OutlineNet, a neural network architecture optimized for processing 2D outlines represented as sequences of $(x, y)$ coordinates.

The proposed pipeline: (a, d, e) 3D models for synthetic data, (b) real-world input, (c) catalog sketch, and (f) extracted outlines for matching.

Outline Representation and Architecture

Instead of raw pixel data, the system uses semi-automatically traced inner and outer fracture lines. To maintain the distinction between these surfaces, a "group-hot" encoding is employed, coupling spatial data with its categorical meaning (inner vs. outer surface). The network uses a dual-path architecture to process both spatial coordinates and angular information (the "bend" of the outline) before merging them for final classification.

CareLoss and Synthetic Data

A key contribution is CareLoss, a loss function designed to combat "class neglect." Unlike standard functions that only penalize based on ground truth, CareLoss adds a weight based on predicted labels. This forces the model to be more discerning for classes it frequently misidentifies, ensuring rare categories are not ignored.

To overcome the lack of training data, the authors developed a synthetic pipeline. Rather than full 3D "shattering" simulations, they mathematically generate millions of synthetic fracture outlines directly from 2D catalog sketches, effectively simulating breaks at a fraction of the computational cost.

Datasets and Results

The model was evaluated on several Roman-era pottery datasets:

  • TSI (Terra Sigillata Italica): 65 classes, 435 sketches.
  • TSH (Terra Sigillata Hispanica): 98 classes.
  • TSSG (South Gaulish Terra Sigillata): 94 classes.

On the TSI dataset, OutlineNet achieved a 73.7% top-10 accuracy on real-world samples. While top-1 accuracy is lower (22%), the high top-10 and top-2 (32.7%) performance demonstrates the tool's effectiveness in significantly narrowing the search space for experts. The pipeline also showed robustness when tested on unseen pottery families (TSH and TSSG) without retraining, reaching up to 81% top-10 accuracy.

Limitations and Future Work

The system currently relies on human-in-the-loop tracing to identify "inner" vs. "outer" surfaces, as automated edge detection remains difficult in noisy excavation environments. Additionally, the model is sensitive to scale, requiring a physical ruler (like a checkerboard) in the photo. Future work will focus on improving automation and addressing potential mislabeling in historical catalogs, which can limit the model's performance.

Summary updated

Abstract

Field archeologists are called upon to identify potsherds, for which they rely on their professional experience and on reference works. We have developed a recognition method starting from images captured on site, which relies on the shape of the sherd's fracture outline. The method sets up a new target for deep-learning, integrating information from points along inner and outer surfaces to learn about shapes. Training the classifiers required tackling multiple challenges that arose on account of our working with real-world archeological data: paucity of labeled data; extreme imbalance between instances of different categories; and the need to avoid neglecting rare classes and to take note of minute distinguishing features of some classes. The scarcity of training data was overcome by using synthetically-produced virtual potsherds and by employing multiple data-augmentation techniques. A novel form of training loss allowed us to overcome classification problems caused by under-populated classes and inhomogeneous distribution of discriminative features.
BibTeX
@inproceedings{itkin2021computational,
  author = {Barak Itkin and Lior Wolf and Nachum Dershowitz},
  title = {{Computational Visual Ceramicology: Matching Image Outlines to Catalog Sketches}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17740},
}
Figure