Template Symbol Match

The bug this exists to prevent

Ask a vision model to count the sprinklers on a floor plan and it does well. Ask it *where they are* and it produces coordinates that look right, land near plausible features, and are wrong. The failure is specific and consistent: models identify well and localise badly.

This matters because the coordinates are the deliverable. A count you cannot place is not a takeoff, and the fabricated positions are individually plausible — nothing in the output looks like a guess, so nothing downstream rejects it.

The split that works: the model names the symbol, this finds it. Classic template matching cannot tell you what a symbol is, and it will find every pixel-similar repeat exactly. The two techniques fail in opposite directions, which is what makes the pairing worth more than either.

The three ways this goes wrong on its own

• It hallucinates on a blank sheet. Correlation always returns a best match; on empty paper
• that best match is noise. There is a threshold and a test asserting a blank sheet yields
• nothing, because "the highest score wins" quietly returns garbage.
• A near-blank template matches everything. An exemplar that is mostly white has almost no
• signal, so it correlates with every empty region on the page. That input is refused rather
• than run — it would otherwise return hundreds of hits and look like a productive result.
• Two nearby stamps collapse into one. Without non-maximum suppression, overlapping high
• scores around a single symbol read as one hit, and two symbols closer together than the
• template width become one. NMS is tuned so that case still yields two distinct hits.

What this does NOT do

No symbol recognition, no classification, no OCR, no PDF parsing, no rasterisation. You supply the pixels and one exemplar crop; it returns centres.

It is not rotation- or scale-invariant. A symbol stamped at 90° is a different template — which is honest for drawing takeoff, where symbol libraries are stamped at fixed orientations, and would be a real limitation elsewhere.

Verified

8 tests: every stamped instance found at exact centres, no hallucinated matches on a blank sheet, near-blank template refused, NMS keeping two close stamps distinct, downsampling arithmetic, and RGBA-to-grey luminance conversion.

Not covered: no test on a real scanned drawing, so behaviour under speckle, JPEG artefacts and faint reproduction is unmeasured. No performance benchmark — this is a straightforward correlation and large sheets will be slow.