3v3r51nc3/eurocoin-vision-dataset

Eurocoin Vision Dataset

A compact computer vision dataset for euro coin detection and classification in realistic RGB images.

The dataset is annotated for multi-class object detection and is suitable for training and evaluating YOLO-style detectors on euro coins captured under non-ideal conditions.

Class Mapping

The dataset uses integer category IDs mapped to coin classes in the order defined by classes.txt.

Category ID	Class Name
0	10_cent
1	1_cent
2	1_euro
3	20_cent
4	2_cent
5	2_euro
6	50_cent
7	5_cent

Notes

• Bounding boxes are stored in objects.bbox
• Class IDs for each bounding box are stored in objects.category
• objects.category[i] is the class of objects.bbox[i]

Example

{
  "file_name": "gp1__18.png",
  "objects": {
    "bbox": [
      [1053.59, 1785.54, 862.64, 495.16],
      [2650.41, 1293.16, 814.52, 450.27]
    ],
    "category": [2, 2]
  }
}

HF to Label Studio Converter

Use hf_to_label_studio.py to convert the dataset's Hugging Face-style metadata.jsonl annotations into Label Studio import files.

Requirements

• Python 3.9+
• Pillow (pip install pillow)

Run

python hf_to_label_studio.py

Input

• Dataset root: this repository directory
• Annotation source: first metadata.jsonl found under data/
• Images: files under data/ with extensions .png, .jpg, .jpeg, .webp, .bmp

Output

Generated under label_studio_import/:

• tasks.json: Label Studio tasks with rectangle annotations converted to percentage coordinates
• label_config.xml: Label Studio labeling interface with all 8 coin classes

Notes

• The script maps category IDs using this class order: 10_cent, 1_cent, 1_euro, 20_cent, 2_cent, 2_euro, 50_cent, 5_cent.
• Image paths in tasks are emitted in Label Studio local-files format: /data/local-files/?d=<dataset-folder>/<relative-path>.

Main use cases

This dataset is intended for:

• object detection of euro coins in cluttered scenes,
• multi-class coin classification,
• coin counting based on detections,
• total value estimation as a downstream task computed from detected classes,
• benchmarking classical and deep vision pipelines.

Annotation format

The dataset contains annotations for the following 8 coin classes:

• 10_cent
• 1_cent
• 1_euro
• 20_cent
• 2_cent
• 2_euro
• 50_cent
• 5_cent

Category ids:

• 0 → 10_cent
• 1 → 1_cent
• 2 → 1_euro
• 3 → 20_cent
• 4 → 2_cent
• 5 → 2_euro
• 6 → 50_cent
• 7 → 5_cent

Dataset metadata:

• year: 2026
• version: 1.0
• contributor: Label Studio

Important note

This dataset is not annotated with direct scene-level labels such as:

• total number of visible coins,
• total monetary value of the image.

However, both quantities can be derived from object detection annotations by:

1. detecting all coins,
2. counting the number of detections,
3. summing coin values according to predicted classes.

So, counting and value estimation are supported as derived tasks, but the dataset is fundamentally structured as a multi-class detection dataset.

Data characteristics

This dataset does not assume ideal acquisition conditions. It reflects practical difficulties commonly encountered in real vision systems.

Common challenges

• Complex backgrounds with textures similar to coin contours
• Camera tilt, making circular coins appear elliptical
• Reflections, causing broken or duplicated edges
• Coins placed close together, making separation harder
• No reference object, so absolute scale cannot be recovered directly
• Low saturation images, where color cues become unreliable

These properties make the dataset useful for studying the limits of:

• YOLO-style object detectors,
• Hough-based circle detection,
• scale estimation without calibration,
• color/material heuristics,
• robust counting in cluttered scenes.

Dataset scale

• 106 images in total

Recommended tasks

Recommended primary task:

• multi-class object detection

Recommended secondary tasks:

• coin counting
• value estimation from detections
• coin family / material grouping (bronze, gold, bimetal), if classes are further regrouped in post-processing

Class coverage

The dataset currently includes only the following denominations:

• 1_cent
• 2_cent
• 5_cent
• 10_cent
• 20_cent
• 50_cent
• 1_euro
• 2_euro

So this is a subset of all euro coin denominations, not necessarily a complete representation of every minting variation or side design.

Limitations

• Small dataset size
• Non-calibrated acquisition
• Potential sensitivity to illumination and viewpoint
• Value estimation depends on detection quality
• No direct scene-level regression labels

Summary

Eurocoin Vision Dataset is a small but challenging benchmark for realistic euro coin detection and classification. Its annotation structure is best suited for object detection, while counting and total value computation should be treated as downstream tasks derived from detections.