summaryrefslogtreecommitdiff
path: root/crates/gipfel-core/src/lsq.rs
AgeCommit message (Collapse)Author
4 dayscore: a fit must actually fit - reject huge residual RMSJohannes Hofmann
User session log (browser console) showed the failure mode precisely: a panorama (Saeuling az 11.2 at px 150, Zugspitze az 91.7 at px 3398, linear pixel spacing) switched to rectilinear mode. Every seed landed at cost ~12900 (RMS ~80 px) and the fitter committed the garbage (pitch -81 deg, scale 91) because it merely passed the sanity filters. Rectilinear geometry cannot represent linear pixel spacing over wide spans: for an azimuth pair d apart, the x-ratio has a hard minimum (~5 for d = 40.9 deg, and it grows with d), so such marker sets have provably NO exact solution - the old 'always solvable' assumption held only for tan-consistent spacings. comp_params now computes rms = sqrt(cost/n) from the final cost and rejects rms > 30 px with: Fit cannot match the markers (RMS 80 px) - this projection cannot represent the marked geometry. Is this a panorama? Switch to cylindrical mode. Regression test uses the user's exact captured marker data. Also fixed a synthetic test placement (800/300 px for a 40.9 deg pair) that was itself rectilinearly impossible - the guard caught it, which is a nice independent confirmation.
4 dayscore+web: paste-ready fit session logs in the browser consoleJohannes Hofmann
comp_params now records every decision in a drainable per-thread log: input markers (name, azimuth, elevation, pixel positions), incoming parameters (incl. distortion), the derived d_img/scale/center seeds, EVERY seed attempt (center, scale, pitch, cost, valid/INVALID), the chosen solution or the plain fallback, and every rejection reason. The web front end drains it after each refit and prints it to the browser console (DevTools -> Console) under 'gipfel fit session'. Failing placements can now be reported by simply copying the console block - it contains everything needed to reproduce the fit verdict offline. Example: comp_params proj_view=60deg n=2 start[center=0.00 ...] markers: [0] Zugspitze az=58.40 ... px=(500,-200) ... d_img=-40.90deg scale0=1401 center0=37.95deg scale_seeds=[1401 1341] seed c= 38.0 s= 1401 p= 0.00 cost= 0.000 valid chosen cost=0.000 center=37.396 pitch=-3.294 roll=5.681 scale=1341.1
4 dayscore: reject wrong left-right marker order with a plain messageJohannes Hofmann
A rectilinear image lists peaks in azimuth order (real roll stays far below 45 deg), so marking the second peak on the wrong side of the first cannot be fitted - but until now the fitter either returned a barely-valid garbage solution (center parked at a marker azimuth +/-90 deg, scale at the sanity floor) or a confusing domain error. comp_params now checks the marker order up front (rectilinear only - cylindrical panoramas legitimately wrap) and answers e.g. Wrong marker order: Saeuling (az 17.5 deg) must be left of Zugspitze (az 58.4 deg) in the image Tests: swapped Zugspitze/Saeuling order rejected with that message; correct order fits at any sane spacing, either side of center.
4 dayscore: validity-filter the multi-start, don't guard after the factJohannes Hofmann
The real cause of the placement-dependent 'fit failed': the exactly- determined two-marker system has SEVERAL EXACT solutions - cost ties. Degenerate ones (scale ~30 px, center at a marker azimuth +/- 90 deg, where the tangent saturates) match both markers perfectly, won the lowest-cost race arbitrarily, and were then rejected by the post-hoc guards. A good solution existed but lost the tie -> random failures that flipped when a marker moved a few pixels. Fix: validity (scale >= 50, every known marker inside the projection domain) now filters candidates INSIDE the seed loop, before the cost comparison. First near-exact valid result (cost < 1 px^2) ends the search early, which also makes the typical fit ~75x faster. If no seed yields a valid fit, a plain run follows so the guards still produce the most informative error. Measured on a simulated Thaneller flow (Zugspitze fixed, Saeuling dragged over a 41x21 grid around its true spot): failures dropped from 163/861 to 0/861. On a deliberately absurd 61x17 grid (markers up to 3000 px off): domain rejections 103 -> 0, remaining failures are honest 'cannot determine scale' cases (marker in the other peak's x-column).
4 dayscore: spread six center seeds across the full circle in the fitJohannes Hofmann
Experiment requested while chasing the remaining placement-dependent 'fit failed' reports: the derived center seed (midpoint of the two marker azimuths in the image-space branch) can itself be far off when marker placements are rough, and the optimizer then never sees the basin containing the true solution. The multi-start now tries six center seeds 60 deg apart around the full circle, crossed with the scale and pitch seeds (24 tiny LM runs - test suite time unchanged). Whatever the marker placement, a seed now starts within 30 deg of any conceivable true center.
4 dayscore: relax fit guard to the projection's true domain, not the 60 deg heuristicJohannes Hofmann
The view-angle guard introduced earlier used get_view_angle() (60 deg for rectilinear) - but that is the C original's VISIBILITY heuristic ('peaks this far off center are usually outside the photo'), not the projection's domain. A wide-angle photo with half-FOV > 60 deg (ultrawide phones, stitched shots) legitimately shows peaks 60-85 deg off center; the TRUE fit got rejected, and whether the fitted center kept both markers under 60 deg depended on exactly where the markers were placed - the remaining placement-dependent 'fit failed' reports. The guard now uses the actual domain: +/-90 deg for rectilinear (the tangent flips sign beyond it), 180 deg for cylindrical. It still catches camera-looking-backwards solutions and grossly misplaced markers, now with the limit in the error message. Tests: rejection moved to 200 deg-apart markers (truly impossible); new test asserts a +/-70 deg wide-angle photo fits cleanly in rectilinear mode.
4 dayscore: multi-start the two-marker fit to kill basin rouletteJohannes Hofmann
The exactly-determined two-marker system (4 residuals, 4 parameters) has multiple solution basins: the tangent repeats every 180 deg in the view center (a camera looking the opposite way projects identically), and pitch/roll can absorb marker y errors. With a single start, the basin the optimizer landed in depended on where the user last dragged a marker - so the new view-angle guard fired 'randomly', and moving a marker vertically could flip success and failure. comp_params now runs the level-1 fit from a small seed set and keeps the lowest-cost result: * scale seeds: the linear estimate plus a tangent-aware variant (rectilinear geometry puts dx = s * 2 * tan(d/2), which the linear estimate underestimates by up to 60% at wide spans) * pitch seeds: zero plus a y-derived estimate (tan p ~ tan elev + y/s per marker, averaged) instead of always starting at pitch 0 lm/lsq now return the final cost for the comparison. Golden fits unaffected; all 38 core + 13 golden tests pass.
4 dayscore: reject fits that park markers outside the projection's viewJohannes Hofmann
Investigating the Thaneller report (mark Zugspitze + Saeuling in rectilinear mode, 'jump totally weird' when Saeuling sits east of north): a numeric sweep of the Saeuling marker across north with the real geography shows comp_params is perfectly CONTINUOUS there - it recovers truth exactly at the true position. What is NOT benign: a fit whose solution places a marked peak beyond the projection's own view angle (rectilinear: +/-60 deg). The tangent-compressed solution that matches far-apart markers is exact for the two markers but scatters every other peak, and is_visible() hides the marked ones - exactly 'totally weird'. comp_params now validates every known marker against proj.get_view_angle() after fitting and returns a descriptive error (suggesting cylindrical mode) instead of committing the parameters. Cylindrical's limit is 2pi, so panorama fits are untouched. The web status line shows 'fit failed: <reason>' in that case. Tests: rectilinear fit of markers 150 deg apart is rejected; the same markers fit fine in cylindrical mode; golden fits unaffected.
4 daysweb: refit view parameters when switching projectionJohannes Hofmann
View parameters are projection-specific: a cylindrical scale is linear in azimuth, a rectilinear one tangent-based. Switching projections kept the old parameters, so the display reinterpreted them; the first marker drag then refit in the NEW projection, converging to parameters that are valid there but garbage for the other one - and switching back showed that garbage until yet another refit happened to save it (exactly the reported panorama -> rect -> drag -> back -> crazy -> tiny-drag-fixes-it chain). * web: the projection dropdown now refits to the markers right after switching (no-op without markers), so parameters always match the active projection. * core: the wide-branch initial guess (markers >180 deg apart in the image, chosen by the sign of the pixel difference) is now gated to projections that can show more than a half-plane. Rectilinear (120 deg max view) always uses the shortest arc. * regression test: panorama fit -> rectilinear interlude -> refit recovers the panorama parameters exactly.
8 daysgrey out distortion controls in panorama mode; make core skip the fitJohannes Hofmann
Confirmed: the cylindrical projection formulas contain no k0/k1/x0 and its distortion derivative trait defaults are NaN, so a level-2 (distortion) fit cannot move anything there. Worse, comp_params zeroed the distortion parameters before the (doomed) calibration - so enabling 'Fit distortion' in panorama mode silently destroyed distortion a rectilinear fit had established. * core: comp_params probes whether the projection actually provides distortion derivatives (NaN trait default = no) and skips both the calibration reset and the level-2 lsq if not. Regression test asserts k0/k1/x0 survive a fit_distortion=true fit in cylindrical mode. * web: the fit-distortion checkbox AND the whole distortion-profiles group now sit in a <fieldset disabled> in cylindrical mode (opacity 0.45), so the dead controls are visibly inert. * dist: rebuilt.
8 dayscore: fix fit for wide panoramas (markers >180 deg apart in image)Johannes Hofmann
The initial guess in comp_params assumed the two markers are less than 180 deg apart: a_center started on the first marker and the scale came from the shortest-arc azimuth difference. In panoramas wider than 180 deg the markers can straddle the seam - e.g. at -170 deg and +170 deg of the view center they are 340 deg apart in the image but only 20 deg on the compass. The start was then off by up to 170 deg in azimuth and a factor of 17 in scale, and the fit never recovered. Fix: derive the image-space azimuth difference from the SIGN OF THE PIXEL DIFFERENCE (the user's marker placement tells us which branch was meant; the magnitude is unique modulo 2pi for panoramas below 360 deg), and start the view center between the markers in that branch. The golden fit tests now compare a_center modulo 2pi: the C harness settles wherever its >2pi input azimuths put it (380 deg), the Rust port normalizes (20 deg) - same physical solution. Tests: fits with markers 340 deg apart, 350 deg apart in reversed marker order, and 200 deg apart (just over the seam). All three fail without the fix (verified), pass with it.
8 dayscore: fix fit when north lies between the two markersJohannes Hofmann
With markers at e.g. 350 deg and 10 deg the raw azimuth difference is 340 deg instead of 20 deg. Two places choked on that: * ProjArgs::from_params fed the raw difference into the projection formulas. The cylindrical projection uses it LINEARLY (x = scale * (m_view - c_view)), so the 350 deg peak was projected 350 degrees of scale to the right of center instead of 10 degrees to the left - no parameter set can satisfy both markers and the panorama-mode fit fails. The rectilinear projection is periodic and unaffected. Fix: represent m_view as c_view + wrapped difference, so every consumer (fit residuals, Jacobian, project_point) sees the shortest arc. get_coordinates already normalized for exactly this reason; the fit path bypassed it. * comp_params derived the initial scale from |d_alph| = 340 deg, an initial scale off by a factor of 17 - this hit rectilinear mode too. Use the wrapped difference. Regression tests: two-marker fits across north for both projections (the cylindrical one fails without the fix), plus a direct check that a 350 deg peak projects to -10 deg * scale.
2026-10-26Replace the C++ implementation with the Rust portJohannes Hofmann
The Rust port (formerly gipfel-rs) is now the official successor to the FLTK/GSL C++ code base. This squashes the port into a single commit and removes the C++ sources (they remain in the git history below). Layout: crates/gipfel-core headless core: hills, panorama geometry, projections (maxima-generated derivatives), pure-Rust LM fit, gipfel: JPEG comment + EXIF I/O, profiles, stitching crates/gipfel-cli command-line tools (export, info, stitch, profiles) crates/gipfel-app Cosmic DE (libcosmic) GUI, binary name `gipfel` codegen/ maxima scripts + regen.sh for the projection code tests/ golden files + C++ reference harnesses (historical) The original README is kept for now; it will be updated to describe the Rust build. gipfel.dat, AUTHORS, ChangeLog and NEWS are unchanged.