Single-plane profile_extrude produces wrong geometry + fitter regresses to 4 spheres on sample_plane_cyl.ply #24

New issue

Closed

opened 2026-06-20 19:22:33 +00:00 by Mark · 1 comment

Mark commented

2026-06-20 19:22:33 +00:00

Owner

Single-plane profile_extrude produces wrong geometry + fitter regresses to 4 spheres on `sample_plane_cyl.ply`

What happened

A side-by-side visual comparison of samples/sample_plane_cyl.ply against
the reconstructed STEP exposes two problems. Both were caught on
2026-06-20 while testing the new single-plane extrude fallback that was
introduced in commit 16404ff.

The input (7000 points, raw units = mm)

A vertical cylinder along the Z axis, ~30–40 mm diameter, ~100 mm tall,
intersected at mid-height by a horizontal plane slab parallel to the XY
plane, ~100×100 mm wide, ~1–2 mm thick.

The output (1 solid, 19 faces, 158 triangles)

A thin plate 0.05 × 0.05 × 0.01 in model units. The cylinder is
entirely missing. The plane's full ~100×100 mm footprint is
collapsed to a 50×50 mm convex hull of the fitter's plane inliers.
There is no curved surface, no hole, no trace of the cylinder.

Side-by-side

Two distinct bugs

Bug 1 — fitter regressed: cylinder → 4 spheres

The pre-16404ff report on this sample had primitives
plane + cone + sphere. The post-16404ff report has
plane + 4 spheres + profile_extrude. The cylinder is being broken
into 4 sphere fits instead of being classified as a cylinder.

This is the inverse of pc2cad#3 ("RANSAC primitive fitting unstable on
synthetic data — sphere misclassified as cylinder"). #3 is sphere
masquerading as a cylinder; #24 is the cylinder being broken into
spheres. Same underlying instability in the RANSAC search, opposite
direction. The cylinder's ~1000 axial points are being claimed greedily
by the densest 1000-point subset that happens to fit a sphere, and the
rest fragment into 3 more sphere fits.

Likely fix path (not for me to implement, the user said no local
patches): make the cylinder-vs-sphere RANSAC comparison penalize sphere
fits whose inlier distribution has high axial variance, or add a
post-RANSAC merge step that combines co-axial sphere fragments into a
cylinder. The original cone fit suggests this used to work; the new
profile_extrude emission may have changed segment assignment or
inlier-pool partitioning in a way that exposed the latent bug.

Bug 2 — single-plane extrude depth source is wrong

The new fallback in derive_profile_operations picks extrude depth
from the nearest cone/cylinder's t_max - t_min. When the fit produces
no axial primitive (because of bug 1), the fallback falls through to
"10% of the polygon's largest extent" — which gave the 0.01 m
thickness. That number is meaningless on a plane slab that should
have a depth of 1–2 mm (matching the pointcloud's ~1–2 mm slab
thickness, not 10 mm).

Likely fix path (also not for me): when no axial primitive is found,
the depth should be estimated from the pointcloud's local density /
bbox along the plane's normal direction, not from the polygon's XY
extent. The current 10%-of-extents heuristic is a stopgap that's
silently producing a 10×-too-thick plate on real input.

A second-order observation: even with a correct depth, the plane
should be reconstructed as a true bounded slab (i.e. trimmed to the
actual 100×100 mm pointcloud extent, not the convex hull of the
fitter's 3146 plane inliers, which is only ~50×50 mm because the
fitter capped inlier count). The polygon in the plane primitive is
the inlier hull, not the geometric extent. That's a separate concern
from bugs 1 and 2.

Verification commands

cd ~/Projects/pc2cad
python pc2cad_cli.py samples/sample_plane_cyl.ply \
  --out /tmp/sample_plane_cyl.step --out-format build123d
python /tmp/render_compare.py   # produces the 3 PNGs above

pc2cad#3 — sphere misclassified as cylinder (inverse direction of
bug 1 here)
pc2cad#22 — build123d export path, the upstream this test exercises
pc2cad#23 — full FreeCAD→OCP port, the architectural fix for the
rest of the export story
Commit 16404ff — the single-plane extrude fallback that exposed
both bugs

# Single-plane profile_extrude produces wrong geometry + fitter regresses to 4 spheres on `sample_plane_cyl.ply` ## What happened A side-by-side visual comparison of `samples/sample_plane_cyl.ply` against the reconstructed STEP exposes two problems. Both were caught on 2026-06-20 while testing the new single-plane extrude fallback that was introduced in commit `16404ff`. ## The input (7000 points, raw units = mm) A vertical cylinder along the Z axis, ~30–40 mm diameter, ~100 mm tall, intersected at mid-height by a horizontal plane slab parallel to the XY plane, ~100×100 mm wide, ~1–2 mm thick. ![Input point cloud](https://forgejo.tail018ac4.ts.net/attachments/040d9389-fc1c-4cdb-bbd0-6e7895518aca) ## The output (1 solid, 19 faces, 158 triangles) A thin plate 0.05 × 0.05 × 0.01 in model units. The cylinder is **entirely missing**. The plane's full ~100×100 mm footprint is collapsed to a 50×50 mm convex hull of the fitter's plane inliers. There is no curved surface, no hole, no trace of the cylinder. ![Reconstructed CAD](https://forgejo.tail018ac4.ts.net/attachments/eeaba244-514d-43d5-9246-b0fadd25941a) ## Side-by-side ![Side by side](https://forgejo.tail018ac4.ts.net/attachments/7469d826-7570-4eb2-8948-2a3550f4ba7f) ## Two distinct bugs ### Bug 1 — fitter regressed: cylinder → 4 spheres The pre-`16404ff` report on this sample had primitives `plane + cone + sphere`. The post-`16404ff` report has `plane + 4 spheres + profile_extrude`. The cylinder is being broken into 4 sphere fits instead of being classified as a cylinder. This is the inverse of pc2cad#3 ("RANSAC primitive fitting unstable on synthetic data — sphere misclassified as cylinder"). #3 is sphere masquerading as a cylinder; #24 is the cylinder being broken into spheres. Same underlying instability in the RANSAC search, opposite direction. The cylinder's ~1000 axial points are being claimed greedily by the densest 1000-point subset that happens to fit a sphere, and the rest fragment into 3 more sphere fits. Likely fix path (not for me to implement, the user said no local patches): make the cylinder-vs-sphere RANSAC comparison penalize sphere fits whose inlier distribution has high axial variance, or add a post-RANSAC merge step that combines co-axial sphere fragments into a cylinder. The original cone fit suggests this used to work; the new profile_extrude emission may have changed segment assignment or inlier-pool partitioning in a way that exposed the latent bug. ### Bug 2 — single-plane extrude depth source is wrong The new fallback in `derive_profile_operations` picks extrude depth from the nearest cone/cylinder's `t_max - t_min`. When the fit produces no axial primitive (because of bug 1), the fallback falls through to "10% of the polygon's largest extent" — which gave the 0.01 m thickness. That number is meaningless on a plane slab that should have a depth of 1–2 mm (matching the pointcloud's ~1–2 mm slab thickness, not 10 mm). Likely fix path (also not for me): when no axial primitive is found, the depth should be estimated from the pointcloud's local density / bbox along the plane's normal direction, not from the polygon's XY extent. The current 10%-of-extents heuristic is a stopgap that's silently producing a 10×-too-thick plate on real input. A second-order observation: even with a correct depth, the plane should be reconstructed as a true bounded slab (i.e. trimmed to the actual 100×100 mm pointcloud extent, not the convex hull of the fitter's 3146 plane inliers, which is only ~50×50 mm because the fitter capped inlier count). The `polygon` in the plane primitive is the inlier hull, not the geometric extent. That's a separate concern from bugs 1 and 2. ## Verification commands ```bash cd ~/Projects/pc2cad python pc2cad_cli.py samples/sample_plane_cyl.ply \ --out /tmp/sample_plane_cyl.step --out-format build123d python /tmp/render_compare.py # produces the 3 PNGs above ``` ## Related - pc2cad#3 — sphere misclassified as cylinder (inverse direction of bug 1 here) - pc2cad#22 — build123d export path, the upstream this test exercises - pc2cad#23 — full FreeCAD→OCP port, the architectural fix for the rest of the export story - Commit `16404ff` — the single-plane extrude fallback that exposed both bugs

compare_sidebyside.png

291 KiB

compare_pointcloud.png

269 KiB

compare_step.png

64 KiB

Mark added the

blocked:needs-decision

label

2026-06-20 20:37:47 +00:00

Mark closed this issue

2026-06-22 02:18:54 +00:00

Mark commented

2026-06-22 02:19:25 +00:00

Author

Owner

Closed — decomposed into three sub-issues on 2026-06-22:

#38 — bug 1: RANSAC cylinder→4-sphere regression
#39 — bug 2: single-plane extrude depth source is wrong
#40 — third concern: plane polygon is the inlier hull, not the geometric extent

Original body, attachments, and verification commands preserved here for reference. The three children carry the implementation work.

Closed — decomposed into three sub-issues on 2026-06-22: - #38 — bug 1: RANSAC cylinder→4-sphere regression - #39 — bug 2: single-plane extrude depth source is wrong - #40 — third concern: plane polygon is the inlier hull, not the geometric extent Original body, attachments, and verification commands preserved here for reference. The three children carry the implementation work.