small-models-for-glam/archival-index-card-detector

Archival Index Card Detector

A YOLO26n object detector for archival index cards. Upload any archival scan — a single card, a multi-card sheet, or a full archival page — and the model will draw a box around every card it finds.

What it's for

The model handles the detection / localisation step in a typical archival processing pipeline:

full archival page scan → this model: where are the cards? → crop each card → OCR / VLM extraction → structured records

It's been trained to work across very different "card" presentations:

• —Full archival pages with one or more cards visible (NLS Advocates Library style)
• —Multi-card sheets with 2–9 cards stacked or arranged in a grid (US Navy Nurse Corps biographical records)
• —Pre-cropped single cards from library catalogs (Boston Public Library) and manuscript catalogs (Duke Rubenstein)
• —Handwritten and typed content
• —Background / cover pages (returns zero detections — see the binder-cover example)

Performance

Single class (card). Validated on a held-out 285-image set, mAP@50 ≥ 0.99 across all four training collections; mAP@50:95 ≥ 0.977. The 3×3 multi-card grid example in the demo is detected as 9/9 cards with mean confidence 0.98 — the case that motivated the project, where the original NLS-only detector returned zero.

Use the model directly

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(
    repo_id="small-models-for-glam/index-card-detector-v5",
    filename="best.onnx",  # or "best.pt"
)
model = YOLO(weights)
results = model.predict("your_scan.jpg", conf=0.25, imgsz=1024)[0]
for box in results.boxes.xyxy.cpu().tolist():
    print(box)  # [x1, y1, x2, y2] in pixel coords

• —Model: `small-models-for-glam/index-card-detector-v5`
• —Training data: `small-models-for-glam/index-card-detection-v5`
• —Base model: `NationalLibraryOfScotland/archival-index-card-detector`

Limitations & feedback

Trained on English-language US/UK archival material. The model may over-predict on newspaper clippings or photographs that appear in mixed archival scans (we have very few non-card negatives in training). Found a failure case? Open an issue on the model repo with the image — it becomes training data for v6.

Configuration (Space owner)

Set environment variables to point at a different model:

• —MODEL_REPO — defaults to small-models-for-glam/index-card-detector-v5
• —MODEL_FILE — defaults to best.onnx