Replace the two-pass closed-form deskew (getPerspectiveTransform +
per-axis scale corrections) with an alternating-minimisation loop around
cv.findHomography (internal Levenberg–Marquardt). Each outer iteration
recomputes per-datum point correspondences from the current H and the
datum's shape constraint, then findHomography refines H. Confidence
drives per-correspondence replication; primary gets a 3× gauge boost.
- Add EllipseDatum type (center + two conjugate semi-axis endpoints +
known diameter) with 3-handle Konva rendering and coin presets.
- Generalise primary selection to any datum type. Priority rect >
ellipse > line; within type, confidence then image size. Warm-start
anchors: rect = 4 axis-aligned corners; ellipse = 4 conjugate-axis
samples on a world circle; line = 2 endpoints + 2 synthetic
perpendicular points (isotropic image-scale assumption).
- Direction-agnostic shape residuals: Procrustes-fit ideal (w × h) rect
to projected corners; midpoint-preserving line rescale; radial-snap
ellipse samples to a circle at projectPoint(H, center).
- Drop the "at least one rectangle" requirement. Any datum combination
works; diagnostics widgets auto-pick a scale reference across types.
- Diagnostics: replace X/Y axis-correction cards with RMS residual +
iteration count; per-datum table shows a residual breakdown column
(edge %, perp Δ°, iso/skew/dia).
- Detect period-2 oscillation in the outer loop and warn to console.
- Relative convergence threshold so the affine and perspective entries
of H are weighted comparably.
- Guard diagnostic diameter via geometric-mean-radius for non-circular
conics; guard collinear-axes ellipses; fix Mat leak in
solveHomography on the exception path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
b87f933b9e