Automatic Die Studies for Ancient Numismatics

This repo contains the implementation of the paper Automatic Die Studies for Ancient Numismatics

[arXiv]

Install

Clone with image matching submodules (including XFeat weights)

git clone --recurse-submodules <URL>

Then install dependencies

pip install -r requirements.txt
pip install -r accelerated_features/requirements.txt
cd RoMa
pip install .

Add Datasets

To add a collection of coins, create a datasets/<COLLECTION-NAME> subfolder containing an image for each coin. Possible formats for the images are PNG or JPG. A subfolder should contain only coins that are likely to be struck by the same die. In particular reverse and obverse of the same coin should be in different subfolders. <COLLECTION-NAME> can be a symbolic link to a folder.

As some collections have images containing both obverse and reverse sides of coins, the Python utility function utils.split_obverse_reverse may be used to split in distinct directories.

Facultative : If clustering evaluation is needed (clustering.evaluate_clustering), add a datasets/<COLLECTION-NAME>_Dies.xlsx excel file, with columns File and Die. Utility utils.create_labels_excel may be used.

Run an Automatic die study

Automatic die studies are ran from run.py, using the configuration described in config.yaml. Main options are:

• Matching Algorithm (XFeat | RoMa)
• Values of XFeat's $top_k$ and RoMa's confidence threshold
• Filtering (True | False)
• Dataset used (Paphos | Tanis1986 | Any added dataset)

Example Scripts

Example scripts for computing similarities and clustering are to be found in examples/.

• example_clustering.py : Run and evaluate all 3 clustering methods on precomputed XFeat similarities, for the Paphos collection.
• example_similarities.py : Compute similarities on a collection (both read from config.yaml). Save the results as similarities/matches.npy.

Fully Automatic CADS Implementation

We provide two Taylor's CADS implementations (namely $\textit{CADS-AG}^*$ and $\textit{CADS-HDBSCAN}$) that are fully automatic. The corresponding Python functions are cads_hdbscan and cads_ag_star in the file cads.py.

Reference

Clément Cornet, Héloïse Aumaître, Romaric Besançon, Julien olivier, Thomas Faucher, Hervé Le Borgne, Automatic Die Studies for Ancient Numismatics, arXiv:2407.20876 , 2024

@misc{cornet2024automaticdiestudiesancient,
      title={Automatic Die Studies for Ancient Numismatics}, 
      author={Clément Cornet and Héloïse Aumaître and Romaric Besançon and Julien Olivier and Thomas Faucher and Hervé Le Borgne},
      year={2024},
      eprint={2407.20876},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.20876}, 
}