johko/print-sheet-qc-pairs

Print-sheet QC pairs (synthetic, reference-based)

Companion to johko/print-sheet-qc-synthetic. This dataset trains reference-based quality control: every example is a triplet of

column	content
reference_image	the "PDF" ground-truth raster — clean sheet, items at ideal grid positions, continuous tone, no camera effects
camera_image	the camera view — same 1024×768 frame, per-item print jitter (±5 px), occasional skewed cards, injected defects, sensor noise / vignette / slight defocus
defect_mask	pixel-accurate binary mask (L-mode) of defect regions in the camera frame
objects	COCO boxes [x, y, w, h] + category (same 7 classes as the detection dataset)
defect_types	list of ALL defect classes applied to this sheet, exactly consistent with the mask (includes skewed_item)

Classes (7): order, smear, scratch, misprint, missing_element, misregistration, skewed_item. ~30% of sheets are defect-free (mask all-zero, defect_types empty).

Coordinate frame

Reference and camera images share the same frame (no global perspective is applied), so a raw difference image is already meaningful. The only geometry differences are the per-item jitter/rotation that real print registration produces. Global sheet-to-PDF homography estimation is deliberately downstream work, not baked into the data.

Mask construction

• smear, scratch, misregistration: drawn on a separate RGBA layer alpha-composited onto the camera image — the mask is that layer's alpha > 0 (pixel-exact).
• misprint, missing_element: the region rectangle (the recolor/cover is painted on the camera image).
• skewed_item: the rotated card's bounding rectangle.

Intended use (hybrid reference-based QC pipeline)

1. Geometric registration of camera image vs reference (with this data they are already aligned; for real sheets estimate a homography from sheet corners/fiducials first).
2. Coarse comparison (NCC / SSIM / color difference) between reference_image and camera_image to generate candidate deviation regions — deterministic, pixel-exact, self-explanatory.
3. Learned decision layer on candidate regions: train a patch-comparison (siamese-style) network on reference-patch + camera-patch pairs from this dataset, supervised by defect_mask, to separate real defects from benign differences (halftone, paper, lighting, small registration shift).

The objects column also keeps this dataset usable for plain detection training, same as the original dataset.

Previews

previews_pairs/ contains side-by-side strips: reference | camera | camera with mask overlay (red).

Generator

generate_print_qc_pairs.py — the deterministic, seed-reproducible generator used to build this dataset (image_id is the render seed). Regenerate or rescale with:

uv run generate_print_qc_pairs.py --num-train 600 --num-validation 100 \
    --repo-id johko/print-sheet-qc-pairs --push

Known limits

• Synthetic appearance only (flat-color vector-style rendering, no real halftone or paper texture); expect a domain gap to real camera images — same caveat as the detection dataset.
• Mask granularity is exact for smear/scratch/misregistration, rectangular for misprint / missing_element / skewed_item.
• The reference raster stands in for the PDF: in production you would rasterize the actual order PDF at the matched resolution instead.