Nutrient benchmark

Form-field detector leaderboard

Object detection of Text, Choice (checkbox / radio) and Signature widgets on document pages, ranked by COCO mAP50-95 — one shared pycocotools scorer on 5,914 held-out pages of the public form-field-v1-benchmark.

How to read it. Cells show mAP50-95 (COCO, the ranking metric), with AP50 and per-class Text / Choice / Sig average precision alongside. pages/s is single-thread CPU throughput, reported for our own models only. The three tabs split the benchmark into empty, filled and handwritten forms. Open any row for its full per-class, per-variant breakdown.

Rank	System	Params	mAP50-95	AP50	Text	Choice	Sig	pages/s

open Apache-2.0, downloadable · commercial Nutrient commercial weights · baseline external reference detector.

Submit or reproduce a result

Score predictions on the frozen benchmark with the reference scorer, then submit the resulting JSON for review.

python score_detector.py --benchmark-repo nutrientdocs/form-field-v1-benchmark \
  --predictions preds.json --out result.json

The benchmark and scoring contract are public; the fixed 5,914-page set will not change silently.

Models benchmarked: nutrientdocs/form-field-v1-nano · nutrientdocs/form-field-v1-small · nutrientdocs/form-field-v1-medium · jbarrow/FFDetr · jbarrow/FFDNet-L · jbarrow/FFDNet-S

About the author
This project is maintained and funded by Nutrient - The deterministic document infrastructure enterprises run their highest-stakes workflows on: replayable output, clear exceptions, and full audit trails on the messy, regulated documents where AI alone breaks.