1. Perform marching cubes over a sparse dual grid
Dual Marching Cubes
Scott Schaefer and Joe Warren
Perform marching cubes
over a sparse dual grid

2. Goals: Capture thin features
& Use fewer triangles.
Dual Marching Cubes (440 tris)
Marching Cubes (67k tris)
Dual Contouring (17k tris)

3. Process Overview: 1. Octree defines resolution
2. Grid vertex placed per octree cell at features of signed distance function
3. Dual grid edges and faces are found
4. Perform marching cubes over dual grid

4. Process Overview: 1. Octree defines resolution
2. Grid vertex placed per octree cell at features of signed distance function
3. Dual grid edges and faces are found
4. Perform marching cubes over dual grid

5. What are features of a signed distance function?
Example: Red line = isosurface

6. Features go where planes collide
Example: Grey shape = distance function

7. Thin shape will ‘grab’ vertices

8. Process to place vertices
1. Approximate surface at sample points with planes x
f(x) xi Ti
1d example with one sample point

9. Process to place vertices
2. Find vertex position to minimize the Error Quadric x
f(x) xi Ti
Vertex v at position x’

10. Value predicted from ith sample
Actual value of f(x,y,z) x
f(x) xi Ti
w=f(x’)
Ti(x’)
Vertex v at position x’

11. x
f(x) xi Ti
w=f(x’)
Ti(x’)
(w-Ti(x’))2 = Error2 from this sample
Vertex v at position x’
(Normalize by slope of Ti )

12. x
f(x) xi Ti xj Tj
Error min is here-ish

13. Features of distance field are not features of the surface!
f(x)
Note that the 1d ‘surface’ is here
f(x) x

14. Features are often near medial axis

15. Result of minimization: 1 vertex per cell
(make sure vertex is in its own cell!)

16. Process Overview: 1. Octree defines resolution
2. Grid vertex placed per octree cell at features of signed distance function
3. Dual grid edges and faces are found
4. Perform marching cubes over dual grid

17. Octree Generation Generate(Node): List samples = Sample(Node)
if (Error(samples) > thresh):
Generate(Children(Node))
Sample(Node): Fine regular sampling? Random sampling?
Error(samples): Error Quadric

18. Process Overview: 1. Octree defines resolution
2. Grid vertex placed per octree cell at features of signed distance function
3. Dual grid edges and faces are found
4. Perform marching cubes over dual grid

19. Dual grid looks like this:

20. dual grid:
Faces become Vertices
Connect Adjacent Faces’ Vertices

21. Border vertices are special
inserted where needed

22. See paper for algorithm

23. Process Overview: 1. Octree defines resolution
2. Grid vertex placed per octree cell at features of signed distance function
3. Dual grid edges and faces are found
4. Perform marching cubes over dual grid

24. Perform marching cubes

25. Perform marching cubes
Pretend you have all 8 corners for every ‘cube’ in the dual grid

26. Perform marching cubes
Should have used a finer octree

27. Problem: slivers

28. Problem: slivers
Dual grid vertices near isosurface
Isosurface w=0

29. Solution: push vertices to surface
Dual grid vertices near isosurface
Isosurface w=0

30. Result of sliver elimination
Without sliver elim
With sliver elim

31. Congratulations, you’re done.

32. Limitations? Paper shows thin features work well
But what about multiple adjacent thin features?

33. Element inversions?
Can faces of the dual grid become inverted?

34. Element inversions?
Can faces of the dual grid become inverted?

35. That’s all. Questions?