mini-CommonForms

Mind the Gaps: A Curated Benchmark for Form Field Detection

Iheb Brini · Omar Moured · Hamza Gbada · Elisa Barney

Document Analysis Systems (DAS), LNCS, pp. 459–473

Curated Form Field Detection benchmark: given a document page, predict the location and type of every fillable field. A cleaned 47k-page curation of CommonForms, in COCO format.

_{text (blue) · choice (orange) · sig (green)}

At a glance

Split	Images	Annotations	text	choice	sig	Negatives (images with no bbox)
train	27,338	609,491	426,550	179,500	3,441	5,692 (20.8%)
val	7,400	345,816	247,217	96,679	1,920	1,189 (16.1%)
test	12,617	308,228	217,429	89,074	1,725	1,536 (12.2%)
total	47,355	1,263,535	891,196	365,253	7,086	8,417 (17.8%)

Note: Categories are 1 = text, 2 = choice, 3 = sig.

Citation

@InProceedings{brini2026mindthegaps,
  author    = {Brini, Iheb and Moured, Omar and Gbada, Hamza and Barney, Elisa},
  editor    = {Shafait, Faisal and Ulges, Adrian and Moetesum, Momina and Yin, Xu-Cheng},
  title     = {Mind the Gaps: A Curated Benchmark for Form Field Detection},
  booktitle = {Document Analysis Systems},
  series    = {Lecture Notes in Computer Science},
  publisher = {Springer Nature Switzerland},
  address   = {Cham},
  pages     = {459--473},
  isbn      = {978-3-032-36207-0},
  doi       = {10.1007/978-3-032-36207-0_27}
}