Sign in to follow this  

How is CSG done?

This topic is 1448 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

This question is one of the reasons I am much happier with programming that doesn't involve anything low-level. Since this is way higher tier math for my understanding when it comes to implementation, and there is not much info on this floating around the internet, I decided to ask here.

 

Here is the link http://en.wikipedia.org/wiki/Constructive_solid_geometry

 

I would like to know how is this calculated, and how do people store this kind of data when they make an engine (first thing that comes to mind, when looking at level editors). I find this interesting since I would need to store entire geometric data (vertices, faces, normals) and uv's along with this when I get the end result of the csg operation.

Share this post


Link to post
Share on other sites

A "simple" approach is to memorize objects as volumes (for 2 spheres you just have 3 coordinates + radius), then you sample points until you find a "solid" point after that you start searching only nearby points where there's "empyt/solid" surface and you generate vertices in thhose points (you can link vertices using something like marching cubes).

 

note that I just proposed you the dummiest and not optimized algorithm, there are for sure much more better ways to do that.

 

note that there exist Framworks for doing that already (any 3d editing tool like Maya or blender already do that at application/script level), so if you find one for you language you can use it directly and probably will result in something more optimized than anything you can do youself

Edited by DemonRad

Share this post


Link to post
Share on other sites

The short version is that there's no magic here -- You remember all those algebra questions from high school where the instructor asked you to find the point where two lines intersect? Its basically just like that, except you have more complex 3-Dimensional shapes involved. If you can describe the shape mathematically, and you have another mathematically-defined shape in the same vector space, then in principle you can plug in 2 inputs into both equations (say, X and Y), and see if the third comes out the same (Z), if so, that point is an intersection of the two shapes.

 

Of course, simply sampling points like this is inefficient, so they do things more cleverly, but the optimized forms are all based on this basic fact. For example, lets say you wanted to calculate the intersection of a cube and a sphere -- well, a sphere is about as simple as it gets, having a center and a radius, and a cube is just 6 planes that are constrained; there are well-known algorithms for finding out if and how a sphere intersects a plane: Take any plane in X-Y, and if the sphere intersects it, the center of the resulting circle is Xs,Ys,Zp (where sub-p means 'plane', and sub-s means 'sphere') -- Now, you can use the Pythagorean theorem the find the radius of the circle that's struck on the plane by the sphere -- The longest-side-term (hypotenuse of the right triangle) is given by the radius of the sphere, and one of the sides is given by the distance between the center of the sphere, and the center of the circle we calculated earlier. By the way, if the distance between the centers is longer than the radius of the sphere, you know the sphere doesn't interact with that plane. Now you have a description of exactly how the sphere interacts with that plane, and you can now apply the same constraints to the circle as the side of the cube (this is a similar process, except now you're operating inside a plane, and the former 'plane' is now a line), and if you still have some circle left, you have a description of how the sphere interacts with that side of the cube. Repeat for all 6 sides, and you're done. You know a lot about how the sphere and the cube interact and you can use that information to modify the shapes involved.

Share this post


Link to post
Share on other sites

There are basically two methods: You either represent your objects as polyhedra and implement boolean operations on polyhedra (more information [url="http://doc.cgal.org/latest/Nef_3/index.html"]here[/url]) or you simply keep the tree that represents the expressions that were used to describe the volume originally. Depending on the type of operations you want to use, one might be better than the other.

 

 

Here's a wackier implementation: https://github.com/colah/ImplicitCAD

Share this post


Link to post
Share on other sites

Actually there are 3 methods:

 

1. Polyhedra (almost every modelling program, such as Blender does that)

2. Distance fields (see for example this site, which is full of that stuff)

3. Depth and stencil buffer tricks (screen-space, very old technique, works on 1990s hardware! See for example this paper, or this library)

 

Distance fields, which let you basically do some form of raymarching through a 3D texture (without actually having a 3D texture!), are by far the coolest, though they come as somewhat unintuitive first.

Edited by samoth

Share this post


Link to post
Share on other sites
Sign in to follow this