/sdf

Parallelized triangle mesh --> continuous signed distance field on CPU

Primary LanguageC++BSD 2-Clause "Simplified" LicenseBSD-2-Clause

Triangle mesh to signed-distance function (SDF)

Given a triangle mesh and a set of points, this library supports:

  1. Computing SDF of mesh at each point: sdf(points)
  2. Computing whether each point is inside mesh: sdf.contains(points)
  3. Computing nearest neighbor vertex index for each point: sdf.nn(points) All operations are CPU-only and parallelized.

Quickstart

Install python binding: pip install pysdf

Usage example:

from pysdf import SDF

# Load some mesh (don't necessarily need trimesh)
import trimesh
o = trimesh.load('some.obj')
f = SDF(o.vertices, o.faces); # (num_vertices, 3) and (num_faces, 3)

# Compute some SDF values (negative outside);
# takes a (num_points, 3) array, converts automatically
origin_sdf = f([0, 0, 0])
sdf_multi_point = f([[0, 0, 0],[1,1,1],[0.1,0.2,0.2]])

# Contains check
origin_contained = f.contains([0, 0, 0])

# Misc: nearest neighbor
origin_nn = f.nn([0, 0, 0])

# Misc: uniform surface point sampling
random_surface_points = f.sample_surface(10000)

# Misc: surface area
the_surface_area = f.surface_area

# All the functions also support an additional argument 'n_threads=<int>' to specify number of threads.
# by default we use min(32, n_cpus) 

To modify the vertices/faces, you can change f.vertices_mutable and f.faces_mutable, then call f.update() to update the internal data structures. You can also use f.vertices and f.faces to access vertices and faces (non-writable).

Screenshots

Robustness under self-intersections:

Reasonable result for non-watertight mesh with multiple parts:

Reasonable result for voxel grid

C++ Library Usage

sdf::SDF sdf(verts, faces); // verts (n, 3) float, faces (m, 3) float

// SDF. points (k, 3) float; return (k) float
Eigen::Vector3f sdf_at_points = sdf(points);

// Containment test, equal (but maybe faster) than sdf >= 0.
// points (k, 3) float; return (k) bool
Eigen::Matrix<bool, -1, 1> contains_points = sdf.contains(points);

// Miscellaneous: nearest neighbor. points (k, 3) float; return (k) int
Eigen::VectorXi nn_verts_idxs = sdf.nn(points);

// Miscellaneous: uniformly random points on surface (generates 10000 in this case)
Eigen::Matrix<float, -1, 3> random_surface_points = sdf.sample_surface(10000);

// Surface area
float surface_area = sdf.surface_area

Note SDF is > 0 inside and < 0 outside mesh.

Use in CMake project

  • find_package(sdf)
  • After defining a target: target_link_libraries(your_target sdf::sdf)

Robust mode

By default 'robust' mode is used. sdf::SDF sdf(verts, faces, false) to disable. The SDF computation will be slightly faster but may be incorrect if the mesh has self-intersections or incorrect winding (not CCW) on some faces.

Python

See the quickstart section for a usage example. help(pysdf) will show more unimportant miscellaneous functions you may want to use (surface normal, area, etc.).

Copying

By default, SDF(verts, faces) will copy the vertices/faces to ensure memory safety, especially since the arguments may be automatically converted. Use SDF(verts, faces, copy=False) to prevent this, if you are sure verts/faces are of types np.float32/np.uint32 respectively and will not be destroyed before the SDF instance. In this mode, vertices_mutable/faces_mutable are unavailable.

Warning about reliability

In robust mode (default) we use raytracing (parity count) to check containment. Currently the ray tracing has the same limitation as embree, that is when ray exactly hits an edge the intersection gets double counted, inverting the sign of the distance function. This is theoretically unlikely for random points but can occur either due to floating point error or if points and mesh vertices are both taken from a grid. In practice, we (1) randomize the ray tracing direction and (2) trace 3 rays along different axes and take majority to decrease the likelihood of this occurring.

In non-robust mode we use nearest surface normal to check containment. The contains check (and SDF sign) will be wrong under self-intersection or if normals are incorrectly oriented.

Dependencies

  • Eigen 3 (Python: automatically downloaded if needed)
  • Vendored (no installation needed):
    • nanoflann
    • nushoin/RTree
  • Optional:

Build + Install

mkdir build && cd build && cmake .. && make -j4 && sudo make install

Demo

A demo program can optionally be built, if meshview is installed. To use it, run ./sdf-demo BASIC_OBJ_FILE. Try the sample-obj/*.obj included in the project.

Benchmark

vs. trimesh

All benchmarks are ran on a 6-core CPU (Intel i7 8th generation, high-performance laptop). More is better.

Model vertices trimesh contains eval/s (numpy) trimesh contains eval/s (pyembree) our SDF evals / s (robust) our SDF evals / s (non-robust)
3241 29,555 237,855 5,077,725 8,187,117
49246 6,835 62,058 2,971,137 4,407,045
179282 1,301 20,157 1,672,859 1,987,869

vs. JianWenPL/multiperson (CUDA)

Here we compare to https://github.com/JiangWenPL/multiperson/tree/master/sdf. The GPU code is ran on a single GTX 1080 Ti, and CPU is a 6-core i7 5820K. I evaluate the SDF on an x-by-x-by-x grid in [-1,1]^3.

Results for SMPL model (13776 faces, 6890 vertices)

Grid resolution multiperson SDF runtime, ms our runtime, ms (robust) our runtime, ms (non-robust) speedup (robust)
32 46.70460891723633 13.006210327148438 7.317066192626953 3.59
64 236.6414031982422 62.57128715515137 51.19466781616211 3.78
128 1521.0322265625 400.36678314208984 347.3823070526123 3.80

Results for SMPL-X model (20908 faces, 10475 vertices).

Grid resolution multiperson SDF runtime, ms our runtime, ms (robust) our runtime, ms (non-robust) speedup (robust)
32 71.34893035888672 13.09061050415039 8.291006088256836 5.45
64 353.7056579589844 66.21336936950684 57.36279487609863 5.34
128 2303.649658203125 477.12063789367676 396.78120613098145 4.83

Notes:

  • This is not really a fair comparison, since the SDF computed in multiperson is more exact (computes distance to all faces) and mine is approximate (only distance to faces adjacent to nearest neighbors). Both methods are prone to numerical error but perhaps mine is more so.
  • Basically the more faces the mesh has, the more performance advantage our method has. For small meshes with very few faces the CUDA implementation should be somewhat faster.
  • My implementation is designed to evaluate on arbitrary continuous points, so it requires the meshgrid points to be generated and passed, while the implementation in multiperson assumes a grid which it generates on the fly, potentially saving some memory access time especially when resolution is high.

License

BSD 2-clause

This library relies on the Eigen (MPL2) nanoflann (BSD) and RTree (MIT) libraries.