Finding (empirical, kiha40 carbody vs a lofted STEP)
The heatmap rendered one entire body flank +4.5mm proud (red) and the opposite flank −4.5mm shy (blue) — reading as a dramatic roll/registration error. Two independent positional instruments (a dedicated flank-split metric + a synthetic-validated roll estimator, OCCTReconstruct PR #166) measure the same pair of bodies near-symmetric (−0.076mm). The heatmap's SIGN was the artifact:
The reference is an open, thin-walled mesh whose sections contain BOTH the outer skin and inner wall surfaces. For a solid sample point, the nearest reference triangle is sometimes the outer skin (normal outward → one sign) and sometimes the inner wall (normal toward the solid → opposite sign). Winding-number sign is undefined on an open mesh; pseudonormal sign flips per nearest-surface. Result: large, spatially-coherent (per-flank) sign patterns with no real positional asymmetry — the worst kind of artifact, because it looks like a systematic shape error.
Ask
- Document the caveat in the tool description: signed output is only trustworthy against watertight/single-surface references; against open thin-wall meshes prefer magnitude + section overlays.
- Robust mode: e.g. sign from the reference's OUTER envelope only (filter inner surfaces by ray/visibility or largest-shell), or an explicit
unsigned: true render, or per-sample nearest-surface-normal agreement check that greys out sign-ambiguous regions instead of colouring them.
Magnitudes remain useful; it's the sign channel that lies. (Reported from the OCCTReconstruct deep review; the false "roll" diagnosis cost one dispatch cycle.)
Finding (empirical, kiha40 carbody vs a lofted STEP)
The heatmap rendered one entire body flank +4.5mm proud (red) and the opposite flank −4.5mm shy (blue) — reading as a dramatic roll/registration error. Two independent positional instruments (a dedicated flank-split metric + a synthetic-validated roll estimator, OCCTReconstruct PR #166) measure the same pair of bodies near-symmetric (−0.076mm). The heatmap's SIGN was the artifact:
The reference is an open, thin-walled mesh whose sections contain BOTH the outer skin and inner wall surfaces. For a solid sample point, the nearest reference triangle is sometimes the outer skin (normal outward → one sign) and sometimes the inner wall (normal toward the solid → opposite sign). Winding-number sign is undefined on an open mesh; pseudonormal sign flips per nearest-surface. Result: large, spatially-coherent (per-flank) sign patterns with no real positional asymmetry — the worst kind of artifact, because it looks like a systematic shape error.
Ask
unsigned: truerender, or per-sample nearest-surface-normal agreement check that greys out sign-ambiguous regions instead of colouring them.Magnitudes remain useful; it's the sign channel that lies. (Reported from the OCCTReconstruct deep review; the false "roll" diagnosis cost one dispatch cycle.)