From bc11f74636b216224b59a3a97f05b050a5ab7988 Mon Sep 17 00:00:00 2001 From: Johannes Hofmann Date: Mon, 16 Nov 2026 23:20:12 +0100 Subject: core: drop the absolute-RMS rejection threshold The rms > 30 px gate was a heuristic backstop, and a flawed one: pixels are the wrong unit (30 px means something different on a 600 px crop than on a 4000 px panorama), and over-determined fits - 4+ markers, unmodeled distortion, honest marker slop - routinely sit at tens of px RMS while being perfectly good fits. Rejecting those would be worse than the garbage-commits it was guarding against, which the principled checks (marker order, same-side feasibility, validity filter) already cover exactly. RMS is still computed and printed to the fit session log as a purely informational line (final rms=NN.Npx). --- crates/gipfel-core/src/lsq.rs | 24 +++++++++++------------- 1 file changed, 11 insertions(+), 13 deletions(-) (limited to 'crates/gipfel-core/src') diff --git a/crates/gipfel-core/src/lsq.rs b/crates/gipfel-core/src/lsq.rs index 182cd84..078d421 100644 --- a/crates/gipfel-core/src/lsq.rs +++ b/crates/gipfel-core/src/lsq.rs @@ -636,19 +636,17 @@ pub fn comp_params( )); } - // The fit must actually MATCH the markers. cost = 0.5 * sum(resid^2) - // over 2n residuals, so rms = sqrt(cost/n). A huge residual means the - // marked geometry cannot be represented in this projection mode at - // all - e.g. a panorama (linear pixel spacing) fitted rectinearly has - // provably no exact solution for wide spans; the optimizer then - // commits absurd parameters (pitch -80 deg, scale at the sanity - // floor) that pass the other guards but never match the markers. - let rms = (final_cost / n as f64).max(0.0).sqrt(); - if final_cost.is_finite() && rms > 30.0 { - return bail(format!( - "Fit cannot match the markers (RMS {rms:.0} px) - this projection cannot represent the marked geometry. Is this a panorama? Switch to cylindrical mode." - )); - } + // Informational only: RMS pixel error of the marked peaks in the + // final solution (cost = 0.5 * sum(resid^2) over 2n residuals, so + // rms = sqrt(cost/n)). NOT a rejection criterion - over-determined + // fits (4+ markers, unmodeled distortion) legitimately sit at tens + // of px, and the pixel unit does not translate across image sizes. + // Unsatisfiable geometries are caught by the principled checks + // (marker order, same-side feasibility, validity filter). + flog(format!( + "final rms={:.2}px", + (final_cost / n as f64).max(0.0).sqrt() + )); *parms = new_parms; Ok(()) -- cgit v1.2.3