Commonforms Eval

commonforms-eval

CommonForms: A Large, Diverse Dataset for Form Field Detection — Barrow et al. (2025) (arXiv:2509.16506, 2025)

What this evaluates

Evaluates an object detection model's ability to locate and classify form field widgets (text inputs, checkboxes/radio buttons, and signatures) on scanned or digital form pages. It probes sensitivity to input resolution and robustness across different languages and document domains.

Datasets

• CommonForms — total 480000; splits: test (-1); repo https://github.com/jbarrow/commonforms

Metrics

• mAP50-95 (primary) — range: percent
  • Mean Average Precision computed across IoU thresholds from 0.50 to 0.95 in 0.05 increments. Per-class AP is also reported for Text, Choice (checkboxes/radio buttons), and Signature fields.

Input / output format

Input: High-resolution rasterized form page images (e.g., 1216px width/height, or 640/960/1536px for ablation studies).

Output: Bounding box coordinates and class labels for detected form fields (Text, Choice, Signature).

Scoring recipe

def compute_mAP(predictions, ground_truth, iou_thresholds=np.arange(0.50, 0.96, 0.05)):
    aps = []
    for iou in iou_thresholds:
        # Match predictions to GT based on IoU >= iou
        # Compute precision-recall curve and average AP
        aps.append(calculate_ap(predictions, ground_truth, iou))
    return np.mean(aps) * 100  # Reported as percentage

Common pitfalls

• Form fields are fine-grained; using low-resolution inputs (e.g., 640px) drastically reduces performance (~20 point drop) compared to high-resolution (1216px+).
• Choice Buttons (checkboxes/radio buttons) and Signatures are particularly resolution-sensitive and often missed by commercial tools or low-res models.
• Aggressive data filtering improves detection efficiency (+4 AP) but reduces the total training pool, requiring careful trade-off analysis.

Evidence (verbatim from paper)

All results are reported as $mAP_{50-95}$. Results are reported on the CommonForms test set, so are directly comparable with all other results reported in the paper.

Citation

@misc{barrow2025commonforms,
  title={CommonForms: A Large, Diverse Dataset for Form Field Detection},
  author={Barrow et al. (2025)},
  year={2025},
  note={arXiv:2509.16506}
}

• arXiv: 2509.16506