🔒 Locked

Create a voxel representation from a 3D model

Started by JDX_John Apr 26, 2014 at 5:12 PM 6 replies 2.4k views

Original Post

Apr 26, 2014 at 5:12 PM

I want to build a voxel model from a 3D model, so that each voxel can be either true or false

At a high level this sounds rather easy - create a NxNxN array and shoot a bunch of rays through your model, as the ray reaches each voxel check if the number ray-poly collisions is odd or even.

But will this work? If I create a 256x256x256 array aligned to x/y/z axes, can I simply do:


for(x=0 to 256)
for(y=0 to 256)
fire_ray(0,0,1)

Or would I have to send rays from each plane and average them somehow? I'm concerned primarily about missing polygons parallel to the rays but it seems to me a in well-formed closed polygon model this shouldn't be a problem?

alvaro

Apr 26, 2014 at 6:57 PM

Your pseudo-code should work. If your polygons form closed manifolds, you shouldn't run into any problems. Well, getting this kind of thing to be robust in situations where several things happen at the same vertex can be tricky. But start with a naive approach and see if it works for you.

tool_2046

Apr 27, 2014 at 2:31 AM

This is probably massively out of date, but can you adjust clip planes to render to slices and use the slices to build the voxel representation? Basically fix a camera on an axis, create clip planes that clip out everything above and below a predetermined thickness, then write out to a slice?

agleed

Apr 27, 2014 at 3:28 AM

In case you're interested in better performing versions, there are faster algorithms available for this problem, or, at least, algorithms that better utilize hardware to make them faster. Even fast enough to use it on-the-fly every frame. You could check out, for example, this bit on voxelization using the hardware rasterizer:

http://www.seas.upenn.edu/~pcozzi/OpenGLInsights/OpenGLInsights-SparseVoxelization.pdf

I haven't implemented the algorithm myself, and he doesn't do true/false voxelization, but his algorithm should be easily adaptable to that. It's used as part of the sparse voxel octree global illumination algorithm that made the rounds two years or so back.

JDX_John

Apr 27, 2014 at 3:43 PM

This is probably massively out of date, but can you adjust clip planes to render to slices and use the slices to build the voxel representation? Basically fix a camera on an axis, create clip planes that clip out everything above and below a predetermined thickness, then write out to a slice?

Wouldn't that only end up rendering the edges rather than filling them? Or have I misunderstood?

tool_2046

Apr 27, 2014 at 8:00 PM

This is probably massively out of date, but can you adjust clip planes to render to slices and use the slices to build the voxel representation? Basically fix a camera on an axis, create clip planes that clip out everything above and below a predetermined thickness, then write out to a slice?

Wouldn't that only end up rendering the edges rather than filling them? Or have I misunderstood?

Yeah, you are right. You need to use the stencil buffer to get an accurate in/out voxel model. Basically, I'm thinking of the voxelization described near the bottom of section 30.2.4 here:

http://http.developer.nvidia.com/GPUGems3/gpugems3_ch30.html

rouncer

Apr 27, 2014 at 9:26 PM

that way is really slow, its actually a huge pain to do anything voxels detailed enough.

the way i recommend is you raster, not raytrace, raster a bunch of depth maps on an axis, one front faces, then one backfacing, etc, until youve got all the overlaps.

the cool thing is, if you chop the missing pixels off, you can render him with displacement map billboards, the billboards get smaller and smaller as it slowly disintigrates to nothing.

theres a few more issues, (like when discontinuities meet each other) but thats the way im making my next voxel thing,... should get a 65536x65536 overhanging procedural terrain out of it. i dont stuff around.

JDX_John

Apr 29, 2014 at 7:43 AM

Well I'm not really planning on needing to generate the voxel model every frame, it's more an offline generation issue (to begin with anyway). Once I have this different real-time uses involve using it as a regular 3D texture, and doing some sort of 2D ray-tracing within it. Possibly the latter can be replaced with GPU/shader trickery as it is going to be a performance drain but as I want to model a physical behaviour rather than fake it graphically, I may not get away with that. But this is a whole separate topic really :)

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.