Four pages, run through both models, no cherry-picking of the boxes. Two of them the model wins, one is a tie, one it loses. The losing one is here on purpose: a page of examples where everything works tells you nothing about where the thing breaks.
Every panel is 640 px, which is what the model actually sees. Left is the baseline, right is this model, and the strip underneath is the forensic head — eight page-condition estimates that the baseline does not produce at all.
5 regions against 76. field and checkbox are not classes the baseline
has, so on a form it can only report that a form exists. This is the clearest
gap between the two, and it is a difference in ontology rather than accuracy.
Clean, it is a tie at 5 regions each — and the baseline gets the figure block that this model misses. Blurred and noised, the baseline returns nothing and this model still returns 5. Across 120 held-out degraded pages the baseline returned no regions at all on 14 of them.
Watch the condition strip between the rows: blur goes 0.00 → 0.98 and
native goes 1.00 → 0.17. That is the model telling you the page is a
degraded raster before you decide whether to trust the boxes.
3 regions against 9. Both models find text; only one separates the paragraphs, which is the difference between an extractor getting five blocks and getting one.
21 regions against 2. Dense multi-column microfilm is the worst case for this model, and the README says so. If your pages look like this, use the baseline.
Every page is a United States federal work under 17 USC 105. There is no copyright in any of them, so there is nothing to attribute and nothing for you to comply with when you copy this repository.
| image | source | where it came from |
|---|---|---|
| 01 | IRS Form 1040 (2025) | https://www.irs.gov/pub/irs-pdf/f1040.pdf |
| 02 | NASA NTRS 20140016745 | https://ntrs.nasa.gov/citations/20140016745 |
| 03 | NASA NTRS 19670020207 | https://ntrs.nasa.gov/citations/19670020207 |
| 04 | USPTO Official Gazette vol. 872 (1970) | https://archive.org/details/officialgazette872unit |
The NASA items were selected by the API's own copyright.determinationType
field reading GOV_PUBLIC_USE_PERMITTED, with no third-party material flagged.
The gazette was digitised by the Internet Archive itself, which its item
metadata records as contributor: Internet Archive.
The degradation in images 02 and 03 is synthetic, applied by this project to those same federal pages. Nothing here is a third-party scan whose reproduction rights are asserted rather than checked.
The figures are generated, not hand-assembled:
./training/.venv/bin/python training/scripts/make_examples.py \
--src <directory of the source PDFs> --out publish/github/examples/imagesbrowser.html is the same model in a page, and detect.py is the same model
in twenty lines of Python. Both are in this directory.