COIN-Research-Group/sawhill-dataset

Sawhill Numismatic Collection Dataset

Dataset Description

This dataset contains video recordings and extracted images of coins from the MacKenzie Art Gallery's Sawhill Numismatic Collection. The dataset is designed for research in automated coin identification, cultural heritage digitization, and computer vision applications in numismatics.

Dataset Summary

• Source: MacKenzie Art Gallery Sawhill Numismatic Collection
• Content: Handheld video recordings of coins, extracted frame sequences, and reference embeddings
• Purpose: Automated coin identification using deep learning
• License: CC-BY-4.0

Supported Tasks

• Image Classification: Identify coins from video frames or images
• Feature Extraction: Pre-computed 512-dimensional ArcFace embeddings for coin matching
• Object Detection: Coin segmentation and cropping from video frames

Dataset Structure

sawhill-numismatic-collection/
├── videos/                      # Raw video recordings (.MP4)
│   └── MVI_XXXX.MP4            # Handheld video of single coin
├── data/
│   ├── reference_embeddings.npy # (N, 512) float32 - ArcFace embeddings
│   ├── reference_labels.npy     # (N,) str - Accession numbers or filenames
│   ├── reference_paths.npy      # (N,) str - Source image paths
│   └── draft_labels/            # Auto-generated labels (CSV + JSON)
│       ├── MVI_XXXX.csv         # Per-frame predictions
│       └── MVI_XXXX_summary.json # Video-level summary
└── output/                      # Extracted frames from videos
    └── MVI_XXXX/
        ├── frame_0000.jpg       # Individual extracted frames
        └── ...

Data Fields

Reference Embeddings

• reference_embeddings.npy: Pre-computed 512-dimensional ArcFace embeddings

  • Shape: (N, 512) where N = number of reference images
  • Type: float32
  • Normalization: L2-normalized

• reference_labels.npy: Labels for each embedding

  • Format: Accession numbers (e.g., "2024.1.11") or filenames (e.g., "MAC 03 Coins-15")
  • Type: Unicode strings

• reference_paths.npy: Source paths for reference images

  • Format: Absolute paths to original images
  • Type: Unicode strings

Draft Labels (Auto-generated)

Each video has two output files:

CSV Format (MVI_XXXX.csv):

frame_idx,timestamp_sec,predicted_label,confidence,candidate_2,score_2,candidate_3,score_3,flagged
0,0.00,2024.1.11,0.2571,MAC 03 Coins-20,0.2355,2024.2.137,0.2252,True

JSON Summary (MVI_XXXX_summary.json):

{
  "video": "MVI_0158.MP4",
  "total_frames_sampled": 31,
  "auto_accepted": 0,
  "flagged_for_review": 31,
  "predicted_label": "2024.1.11",
  "prediction_confidence": "low"
}

Data Collection

Video Recording

• Equipment: Handheld camera (iPhone/digital camera)
• Setting: Natural lighting, plain background
• Method: ~30-second video rotating coin to show both obverse and reverse
• Frame Rate: 29.97 fps
• Resolution: 1920x1080

Reference Images

Reference images sourced from:

1. Named coins: MacKenzie Art Gallery catalog with accession numbers (YYYY.M.N format)
2. Loose images: Additional catalog images labeled by filename

Processing Pipeline

1. Frame Extraction: Sample frames at 1 fps intervals (every 30th frame)
2. Coin Segmentation:
   • Border color statistics in LAB color space
   • Otsu thresholding
   • Morphological operations
   • Connected components analysis
   • Centroid-based cropping to 224×224
3. Embedding: ArcFace model (512-dim) trained on coin images
4. Matching: L2-normalized cosine similarity via dot product

Usage

Download Dataset

# Install dependencies
pip install huggingface_hub

# Download from your computer
git clone https://github.com/COIN-Research-Group/extract-sawhill-dataset
cd extract-sawhill-dataset
python hf_pull.py --repo-id COIN-Research-Group/sawhill-dataset

# Or download only specific components
python hf_pull.py --repo-id COIN-Research-Group/sawhill-dataset --data-only
python hf_pull.py --repo-id COIN-Research-Group/sawhill-dataset --videos-only

Upload New Data

# After adding new videos or processing new data
python hf_push.py --repo-id COIN-Research-Group/sawhill-dataset

# Upload only videos
python hf_push.py --repo-id COIN-Research-Group/sawhill-dataset --videos-only

# Upload only processed data
python hf_push.py --repo-id COIN-Research-Group/sawhill-dataset --data-only

Load Reference Embeddings

import numpy as np
from huggingface_hub import hf_hub_download

# Download files
embeddings_path = hf_hub_download(
    repo_id="COIN-Research-Group/sawhill-dataset",
    filename="data/reference_embeddings.npy",
    repo_type="dataset"
)
labels_path = hf_hub_download(
    repo_id="COIN-Research-Group/sawhill-dataset",
    filename="data/reference_labels.npy",
    repo_type="dataset"
)

# Load data
embeddings = np.load(embeddings_path)  # Shape: (N, 512)
labels = np.load(labels_path)          # Shape: (N,)

print(f"Loaded {len(embeddings)} reference embeddings")
print(f"Unique coins: {len(np.unique(labels))}")

Process New Videos

# See the main repository for full pipeline
# https://github.com/COIN-Research-Group/extract-sawhill-dataset

# Quick start:
from autolabel import autolabel_videos

autolabel_videos.autolabel_video(
    video_path="path/to/new_video.mp4",
    reference_embeddings=embeddings,
    reference_labels=labels,
)

Model Information

ArcFace Embedding Model

• Architecture: ResNet50 backbone + ArcFace head
• Embedding Dimension: 512
• Training: Trained on coin image pairs
• Normalization: L2-normalized embeddings
• Weights: arcface_main.pth (not included in dataset, see main repository)

Performance

• Test Accuracy: Correct identification on MVI_0158.MP4 (target: 2024.1.11)
• Confidence Scores: 0.18-0.27 range (reflects domain gap between handheld video and studio photography)
• Database Size: 787 reference embeddings (311 named coins × 2 faces + 165 loose images)

Limitations

1. Domain Gap: Reference images are high-quality studio photos; videos are handheld with variable lighting
2. Confidence Threshold: Current threshold (0.82) may be too high for cross-domain matching
3. Segmentation Quality: Performance varies with background complexity and lighting conditions
4. Coverage: Reference database may not include all coins in the full collection

Citation

If you use this dataset in your research, please cite:

@dataset{sawhill_numismatic_2026,
  title={Sawhill Numismatic Collection Dataset},
  author={MacKenzie Art Gallery},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/COIN-Research-Group/sawhill-dataset}}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.

You are free to:

• Share: Copy and redistribute the material
• Adapt: Remix, transform, and build upon the material

Under the following terms:

• Attribution: You must give appropriate credit to the MacKenzie Art Gallery

Contact

For questions or contributions, please open an issue on the main repository.

Acknowledgments

• MacKenzie Art Gallery for providing access to the Sawhill Numismatic Collection
• Contributors to the coin-embedding project for the ArcFace model architecture