LEI

Datasets

Detection (LEI-Detection)

Scanned index cards with bounding-box annotations for three object classes (stamp, etymon, content).

• Train:
  • Stamp: 6,115 instances
  • Etymon: 12,731 instances
  • Content: 6,349 instances
• Validation:
  • Stamp: 644 instances
  • Etymon: 1,230 instances
  • Content: 663 instances
• Total:
  • Stamp: 6,759 instances
  • Etymon: 13,961 instances
  • Content: 7,012 instances

Download:
https://next.hessenbox.de/index.php/s/75afxwmToTge2gX

---

Recognition (LEI-Stamps)

Cropped textual stamp images used for stamp recognition.

• Total images: 170,400
• Unique stamp classes: 3,817
• Train split:
  • 161,652 images
  • 3,817 stamp classes
• Validation split:
  • 8,748 images
  • 2,916 stamp classes (stamps with ≥4 instances)
  • 3 images per class in validation

Download:
https://next.hessenbox.de/index.php/s/QFHKTqwzpwg6Tce

---

Benchmark (LEI-Benchmark)

Full index-card images (no bounding boxes) for end-to-end benchmarking of stamp detection and recognition, including cards without stamps.

• Total index cards: 1,369
• Cards with stamp: 1,348
• Cards without stamp: 21
• Unique stamps: 638

Download:
https://next.hessenbox.de/index.php/s/oXjMwWSXY4w23Ra