Corpus Nummorum - Coin Image Dataset

Corpus Nummorum - Coin Image Dataset

This dataset is a collection of ancient coin images from three different sources: the Corpus Nummorum (CN) project, the Münzkabinett Berlin and the Bibliothèque nationale de France, Département des Monnaies, médailles et antiques. It covers Greek and Roman coins from ancient Thrace, Moesia Inferior, Troad and Mysia. This is a selection of the coins published on the CN portal (due to copyrights). 

The dataset contains 115,160 images with about 29,000 unique coins. The images are split in three main folders with different assignment of the coins. Each main folder is sorted with the help fo subfolders which hold the coin images. The "dataset_coins" folder contains the coin photos divided into obverse and reverse and arranged by coin types. In the "dataset_types" folder the obverse and reverse image of the coins are concatenated and transformed to a quadratic format with black bars on the top and bottom. The images here are sorted by their coin type. The last folder "dataset_mints" contains the also concatenated images sorted by their mint. An "sources" csv file holds the sources for every image. Due to copyrights the image size is limited to 299*299 pixels. However, this should be sufficient for most ML approaches. 

The main purpose for this dataset in the CN project is the training of Machine Learning based Image Recognition models. We use three different Convolutional Neural Network based architectures: VGG16, VGG19 and ResNet50. Our best model (VGG16) archieves on this dataset a 79% Top-1 and a 97% Top-5 accuracy for the coin type recognition. The mint recognition achieves an 79% Top-1 and 94% Top-5 accuracy. We have a Colab notebook with two models (trained on the whole CN dataset) online.

During the summer semester 2023, we held the "Data Challenge" event at our Department of Computer Science at the Goethe-University. We gave our students this dataset with the task to achieve better results than us. Here are their experiments:

Team 1: Voting and stacking of models

Team 2: Multimodal model

Team 3: Transformer models

Team 4: Dockerized TIMM Computer Vision Backend & FastAPI

• Approach | Type Dataset | Mint Dataset
• Ours                  79%                79%
• Team 1                -                    86%
• Team 2              86%                  -
• Team 3              88%                58%
• Team 4                -                      -

Now we would like to invite you to try out your own ideas and models on our coin data.

If you have any questions or suggestions, please, feel free to contact us.