Unlimited-OCR — segment a composite figure into panels

Self-Evolving Skill: improves through use. The behaviour below was measured, not assumed — see ../../references/EMPIRICAL.md. Fix this file the moment a claim stops holding.

Why this exists

Unlimited-OCR localises charts and transcribes nothing inside them. Fed a 1080×1504 image containing nine matplotlib panels, it returned nine <|det|>chart [box]<|/det|> markers with perfect boundaries and zero characters of text — not even the panel titles, which were legible.

That is a dead end if you wanted a transcription. It is excellent if you wanted a segmenter: the boxes were accurate enough to crop nine clean, individually-interpretable panels out of one dense collage.

The pipeline that follows from it:

composite image ──[Unlimited-OCR: localise]──> N panel crops ──[a chart-reading model]──> N descriptions

A vision model handed one chart at a time, filling the frame, does markedly better than the same model handed a nine-panel grid and asked to describe all of it at once. This skill is the first stage; the describer is whatever you already use.

Use it

S=~/eon/cc-skills/plugins/unlimited-ocr/scripts/unlimited_ocr.py

# Every detected region, cropped
uv run --no-project $S segment --input figure.png --output ./panels

# Only charts and tables, ignoring text blocks
uv run --no-project $S segment --input figure.png --output ./panels --categories chart,table

# Wider margin — useful when axis labels sit outside the plotted area
uv run --no-project $S segment --input figure.png --output ./panels --pad-pixels 24

Output is <stem>_<index>_<category>.png per region, plus a segments.json manifest carrying the source size, both coordinate systems, and any text the model did associate with each region.

`--pad-pixels` defaults to 12, and should not be 0

The model's boxes hug the plotted area. A pixel-exact crop of a chart therefore cuts off the x-axis tick labels sitting just below it — verified on the nine-panel figure, where every crop at padding 0 lost its date axis. Those labels carry the units and the date range, which is most of what makes a chart interpretable by the next model in the chain.

Twelve pixels is a starting point, not a law. Raise it for figures with outboard captions.

Reading the manifest

{
  "source_size": [1080, 1504],
  "regions_detected": 9,
  "regions_written": 9,
  "segments": [
    {
      "index": 0,
      "category": "chart",
      "normalized_bbox": [0, 0, 999, 152],
      "pixel_bbox": [0, 0, 1080, 241],
      "path": "…_000_chart.png",
      "text": ""
    }
  ]
}

regions_detected and regions_written differ when --categories or --min-pixels filtered something out. They are reported separately on purpose: a silent gap between "what the model found" and "what you got" is how a segmentation job quietly under-delivers.

Limits

Boxes are normalised to 0–1000 on both axes and converted to pixels using the image's real size. A region whose box the model omitted is skipped rather than guessed at.
Segmentation quality is only as good as the layout detection. Verify a sample visually the first time you point this at a new document class — the nine-panel result was checked by eye before it was written down here.
This does not describe anything. It cuts. Pair it with a describer.

Post-Execution Reflection

After this skill completes, check before closing:

Did any crop clip its content? — raise --pad-pixels and record the document class that needed it. The default of 12 exists because 0 clipped axis labels off every panel of a real figure.
Were fewer regions emitted than the figure visibly contains? — the model omitted a box, not the CLI. Note the document class in references/PITFALLS.md; silent under-delivery is the failure mode this skill warns about.
Did you pair it with a describer? — this skill cuts and never describes. If the describer needed a specific crop size or format, write that requirement down here.

unlimited-ocr-segment-figure