2

I've coded a full octree implementation without too much optimization for 3D reconstruction, however, the tree structure contains too many pointers, and cannot support more than 256^3 voxels.

Theoretically, for a non-tree structure if I used a vector<bool> which uses ~1 bit per voxel, this would be more acceptable because the non-tree structure could support 2k^3 with 8GB memory.

However an optimized octree structure should be able do equal to or better than this, since:

  1. It shouldn't have to store every voxel, since condensation can allow compression of nearby, same-value voxels.

  2. It shouldn't use too many pointers, since pointers themselves uses a fair amount of bytes already.

  3. The octree must have a fairly low node/voxel ratio.

For a full octree the node number could be calculated as (s^3 -1) / 7. The s is the volume resolution, which is a power of 2. For example if s = 4, I'd need 1 + 8 = 9 nodes in the octree to represent a 4x4x4 grid of voxels.

Does anyone know of an octree implementation in C++ that meets these specifications?

Philip Conrad
  • 1,451
  • 1
  • 13
  • 22
tomriddle_1234
  • 3,145
  • 6
  • 41
  • 71
  • This sort of stuff tends to be done using GPU's these days, you may be able to adapt the techniques used there for improving your code, a good starting place is Cyril Crassin's PhD work: http://maverick.inria.fr/Membres/Cyril.Crassin/thesis/ – Necrolis Jan 23 '13 at 06:23
  • Why don't you just optimize youre current implementation? Leaf nodes for example, don't need a pointer to child nodes, which saves you some memory. Also you can use a sparse octree. I don't know how your octree implementation looks like, but if you want it to be memory efficient, why are you using vectors? 256^3 is a pretty low value. – Tara Feb 07 '13 at 11:13
  • not sure that I understand the question but I think you can use array tree – Mzf Jun 09 '13 at 18:46

1 Answers1

3

I think octrees are the way to go but children nodes should be built only if needed (at least one voxel is set). Moreover, compression should be used. Often adjacent voxels have same values and thus RLE compression seems to works well. This is explained in this thesis http://www.terathon.com/voxels/

Thelvyn
  • 264
  • 3
  • 8