One real target, scored end-to-end by the production pipeline. Seven rounds fired — one landed off paper, six scored. The fixture is regenerated by a script, never hand-edited.
Also builtsingle-shot live pipelineBatch mode splits work between Claude Vision (judgment — “how many holes, roughly where?”) and OpenCV (precision — with a single shared diff and a global matcher that prevents double-counting when anchors overlap). Scroll through the pipeline.
Same SIFT + FLANN + Lowe + RANSAC pipeline the live page uses — shared code, shared fixture format. The only difference here is what comes next: a single rectified canvas that every downstream anchor reuses, rather than a before/after pair feeding a temporal diff.
Claude receives the aligned photo and returns approximate positions for every perforation it can confidently identify. The hint (“about 7 rounds fired”) is deliberately framed as a hint, not a quota — Claude is free to return fewer shots when evidence is ambiguous, rather than inventing them from smudges or shadows. On this fixture, the hint was 7 and the scout returned 6. That gap is by design.
The aligned photo is differenced against a brightness-normalised template — the template scan lives near saturation, handheld photos are usually dimmer, so we shift their median brightness to match before the diff. Printed ring and numeral edges are then zeroed out via a feature mask (≈11.1% of pixels), so the diff only contains real perforations. Crucially this runs once— the same diff image is reused inside every anchor’s ROI, not recomputed per shot.
Each anchor is pre-assigned to its nearest ring group; that group’s ring_width drives its search radius and blob-area filters. The bounds come from the physical pellet/ring ratio (a 4.5mm pellet on an 8mm-ring target is always ~0.56 ring widths across) — so adding a new template at a different print DPI works without per-template recalibration. On this fixture every anchor gets a search radius of 339px (5.5 × ring_width) with a blob-area window of 229–22915px².
Every anchor’s ROI surfaces 2–4 candidate observations, and with a 5.5× ring-width radius, ROIs overlap heavily — multiple anchors often see the same physical hole. Observations across all ROIs are clustered (complete-linkage, so adjacent pellets can’t merge through a bridge observation), and a Hungarian solver finds the globally minimum-cost anchor↔cluster assignment. Edge cost is the distance from the anchor to its own observation in that cluster — so the returned centroid is always one the anchor actually saw, never a neighbour’s. The heatmap below is the solver’s cost matrix: darker means cheaper, the ringed cells are the assignment Hungarian picked.
| c0 | c1 | c2 | c3 | c4 | c5 | ||
|---|---|---|---|---|---|---|---|
| a0 | 361 | 452 | 272 | 92 | · | · | 4.8r4 |
| a1 | · | 156 | 34 | 365 | · | · | 10.0r10 |
| a2 | · | 90 | 239 | · | 349 | · | 8.8r8 |
| a3 | · | 224 | 332 | · | 74 | 342 | 2.5r2 |
| a4 | 265 | · | · | · | 309 | 60 | 4.8r4 |
| a5 | · | 478 | · | · | 273 | · | 5.7r5 |
Batch has no before-frame to fall back on — without alignment, the sniper has no template to diff against, and there’s no temporal signal to isolate holes. If SIFT can’t find enough keypoints (too blurry, target cropped, extreme angle), the endpoint returns 422 with a hint to retry rather than guessing.