fast_sweeping

The fast sweeping method for the computation of the signed distance function in 2D in 3D. Implemented in Rust, with Python bindings (see python directory).

Based on [1].

Usage

Add the following to your Cargo.toml

[dependencies.fast_sweeping]
git = "https://github.com/rekka/fast_sweeping.git"

At the top of your crate add:

extern crate fast_sweeping;

Depending on the dimension, use signed_distance_2d or signed_distance_3d for the Euclidean distance, or anisotropic_signed_distance_2d, anisotropic_signed_distance_3d for other norms.

Accuracy

There are two main things to consider when evaluating the accuracy of the method.

Initialization near the level set

There is no unique way of doing this. In this implementation, we simply split the squares/cubes of the regular mesh into 2 triangles/6 tetrahedra and assume that the level set function is linear on each of them. The initial distance at the vertices is then taken as the distance to the line/plane going through the triangle/tetrahedron (not to the intersection of the line/plane with the triangle/tetrahedron). If the vertex is contained in multiple triangles/tetrahedra that intersect the level set, the minimum of all distances is taken.

The main advantages of this approach are:

Simple.
Does not move flat parts of the level set (unless near other parts of the level set).
Does not move symmetric corners.

Disadvantages:

Introduces some more anisotropy.
The initial value outside of corners is not really the distance to the level set but smaller. But this does not seem to actually cause larger error in the circle test case, see examples/error. The error seems to be bigger inside a circle. It appears that within small neighborhood (dist <= 3h) of the level set the max error of order h².

For an example see examples/redistance.

Finite difference approximation

Here we use the simplest first order upwind numerical discretization as given in [1]. This gives an error of order O(|h log h|).

References

[1] Zhao, Hongkai A fast sweeping method for eikonal equations. Math. Comp. 74 (2005), no. 250, 603–627.

Future work

Support noneven distances.
Support f32.
Possible optimizations are:
- Only compute distance in a small neighborhood of the level set.
- Use multiple threads. However, this is relatively nontrivial due to the sequential nature of the Gauss-Seidel iteration.

Python bindings

See python directory.

License

MIT license: see the LICENSE file for details.