Skip to content

Mesh Processing

mesh_processing: Triangle mesh operations: subdivision, simplification, and conversion to point clouds via uniform or Poisson-disk sampling.
Methods subdivision, simplification, mesh_to_pcd_uniform, mesh_to_pcd_poisson_disk
Input files .ply, .obj, .stl, .off
Bundled datasets to try open3d_bunny, open3d_icp_demo, kitti_raw_2011_09_26_drive_0001, medshapenet, c3vd_endoscopy_splats, hf_objaverse_glb_shard
Example Gaussian splats on the Stanford bunny
CLI pardesline run mesh_processing.<method> input.ext [--param value]

How it works

Mesh processing operates on triangle meshes \(\mathcal{M} = (V, F)\). Key operations include:

Subdivision increases mesh resolution by splitting triangles. Midpoint subdivision inserts new vertices at edge midpoints, quadrupling the face count per iteration.

Simplification reduces complexity using Quadric Error Metrics (QEM) which iteratively collapses edges minimizing \(\Delta(v) = v^T Q v\).

Mesh-to-Point-Cloud sampling converts surfaces to point sets via uniform or Poisson-disk sampling.

Methods

subdivision

Midpoint subdivision splits each edge at its midpoint creating 4 new triangles from each original. After \(k\) iterations: \(|F| \cdot 4^k\) faces.

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("mesh_processing.subdivision")  # catalog defaults
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.mesh_processing.subdivision --input ... --output ....

Defaults validated on: generated meshes in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
iterations int 1 1 to 4 Number of subdivision iterations. 1 iteration quadruples the triangle count; two or more are rarely needed before smoothing and make files large quickly.

simplification

Quadric Error Metrics decimation collapses edges in order of increasing cost \(v^T (Q_a + Q_b) v\) until the target triangle count is reached.

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("mesh_processing.simplification")  # catalog defaults
# same thing, domain vocabulary: `mesh.decimate()`
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.mesh_processing.simplification --input ... --output ....

Defaults validated on: Stanford bunny (examples/hippocampus_mask_to_mesh uses the PyVista decimation instead); generated meshes in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
target_triangles int 5000 100 to 100000 Target number of triangles. 5 000 triangles is the size at which an organ or part surface still looks right on screen and loads instantly in a viewer; raise it for metrology.
maximum_error (advanced) float 1e+30 0.0 to 1e+30 Advisory maximum quadric error allowed per edge collapse (very large = unbounded, matches default behavior). Unbounded (1e30) matches Open3D's behaviour: the decimation stops at the triangle budget, not at an error threshold, so the result is predictable in size.
boundary_weight (advanced) float 1.0 0.0 to 100.0 Weight applied to boundary edges during decimation. 1.0 treats boundary edges like interior ones; raise it on open meshes (scans with holes) so the boundary does not shrink.

mesh_to_pcd_uniform

Points are sampled uniformly, each triangle receiving samples proportional to its area \(A_i\).

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("mesh_processing.mesh_to_pcd_uniform")  # catalog defaults
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.mesh_processing.mesh_to_pcd_uniform --input ... --output ....

Defaults validated on: Stanford bunny (examples/points_to_gaussians_bunny); generated meshes in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
num_points int 5000 100 to 100000 Number of points to sample. 5 000 points keep registration and clustering fast while covering a surface; the examples use 20 000-50 000 for visual density.
use_triangle_normal (advanced) bool false Assign each sampled point the normal of its source triangle instead of an interpolated vertex normal. Off: interpolated vertex normals are smoother on organic surfaces; turn on for CAD meshes with flat faces.
seed (advanced) int -1 -1 to 2147483647 Random seed for sampling (-1 = nondeterministic). -1 (non-deterministic) is Open3D's default; set a seed when a result must be reproducible, as the examples do.

mesh_to_pcd_poisson_disk

Blue-noise sampling where no two samples are closer than minimum distance \(r\), producing more uniform coverage.

import pardesline as pl

obj = pl.load("input.ply")
result = obj.run("mesh_processing.mesh_to_pcd_poisson_disk")  # catalog defaults
# same thing, domain vocabulary: `mesh.to_points(num_points=...)`
result.provenance[-1]  # the step, every parameter used, duration, version

Output: .ply. Script: python -m pardesline.ops.mesh_processing.mesh_to_pcd_poisson_disk --input ... --output ....

Defaults validated on: Stanford bunny (examples/poisson_reconstruction_bunny, examples/icp_registration_bunny); generated meshes in tests/test_ops_smoke.py

Parameter Type Default Range What it does Why this default, when to change it
num_points int 5000 100 to 100000 Approximate number of points. 5 000 is the approximate count after elimination: Poisson-disk sampling gives evenly spaced points, which is what registration features and PointNet training want.
init_factor (advanced) int 5 1 to 20 Initial uniform-sample oversampling factor before elimination. 5x oversampling before elimination is Open3D's recommendation: enough candidates for an even blue-noise distribution without a long elimination pass.
use_triangle_normal (advanced) bool false Assign each sampled point the normal of its source triangle instead of an interpolated vertex normal. Off: interpolated normals; on for flat-faced CAD meshes.
seed (advanced) int -1 -1 to 2147483647 Random seed for sampling (-1 = nondeterministic). -1 is non-deterministic; the examples pass seed=1 for reproducible figures.

Generated from src/pardesline/catalog/data/mesh_processing.json and _defaults.json on 2026-10-10.