variation op• Data kinds: video → table
• Call: import fullseye as fs; fs.ledger.varmodel_alignment(stack, max_misalign_px=0.5) (to call the implementation directly, import varmodel; varmodel.varmodel_alignment(stack, max_misalign_px=0.5); from the registry, opsspc.get("varmodel_alignment"))
• GPU: yes (mechanical GPU path from the allow-list, CuPy). device="auto" chooses the GPU from the measured size (2,097,152 pixels) up
Translation offset of every good sample against the typical position of the stack, and the margin to the tolerance (table).
Each image of stack (n, H, W) is phase-correlated (filters_freq.phase_correlation_fft) against a reference built from the stack itself (the most typical sample, then the registered mean); the integer peak is refined to 1/20 px by an upsampled DFT (Guizar-Sicairos et al. 2008). Offsets are relative to the median offset, so a common shift of every sample is not a misalignment. Returns offset (n, 2) = (row, col) in px (positive = the sample sits further down / right than the typical one), abs_offset (n,) = the larger of the two, margin_px = max_misalign_px - abs_offset (negative = this sample breaks the tolerance), ok (all margins >= 0), worst (index of the most shifted sample) and psr (peak-to- sidelobe ratio of the correlation), shift_sd_px (Cramér–Rao lower bound of the offset's
spread: noise variance over the structural gradient energy of the registered mean; the window makes the real spread a few times larger) and observable. A featureless or pure-noise stack has no structure, so its offsets cannot be measured (observable=False, offsets reported as 0, shift_sd_px capped at the image size) — and for the same reason a shift cannot fatten its band, so such a stack is not refused.
Why it matters: a per-pixel spread learned from a misaligned stack is inflated wherever the image has gradient (edges), so the tolerance band gets fatter and real defects there are missed.
Raises ValueError on an empty stack, a stack that is not (n, H, W) / a list of equal 2-D images, non-finite values, or max_misalign_px <= 0.
• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).
• Operator provenance and references — the sources of the research/methods this op family came from.
• The canonical algorithm (author, year) and its uses are named in the family usage guide above.
• poc_good_sample_variation_model — py -3.11 examples/poc_good_sample_variation_model.py
table as input)msa_anova_table · msa_gauge_rr · msa_bias_linearity · msa_attribute_agreement · gum_standard_uncertainty · gum_propagate · gum_expanded · gum_monte_carlo
variation)varmodel_band_k · varmodel_fit · varmodel_calibrate · varmodel_score
*Provenance: varmodel.py — SPC operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.