Fast Scene Voxelization and its Applications

1
Fast Scene Voxelization and its Applications
Elmar Eisemann and Xavier Décoret ARTIS-GRAVIR
/IMAG-INRIA ARTIS is a team of the GRAVIR/IMAG
laboratory, a joint effort of CNRS, INRIA, INPG
and UJF
2
A long time ago ...

• Simple games and simple pleasures...

3
Nowadays ...

• 3D
• Complex models
• Complex effects

4
Context

• Complex and dynamic scenes
• Need for
• Fast information acquisition
• Fast information evaluation
• Shader compatibility

5
Context

• Lots of effects

6
Context

• Lots of effects

7
Context

• Lots of effects

8
What is common?

• All use information from voxels!

9
What is a Voxel?

• Space discretized by cubes (voxels)
• Can contain information

10
Why would I care?

• Lots of applications
• Acceleration structure
• Reconstruction
• Visibility
• Terrain rendering
• Simplification...

11
Whats good about voxels?

• Constant complexity
• (scene independent)
• Simple structure
• Simple entities

12
Whats bad about voxels?

• Storage cost (3D)
• Not GPU adapted
• Slow on CPU
• 100K polygons
• octree 256 x 256 x 256
• 3 minutes Haumont and Warzee 2002

13
Previous Work on GPU Voxelization

• Everitt et al. 2001 using Depth Peeling
• Layers are peeled off using zBuffers

• Pro
• Representation is accurate from the viewpoint
• Contra
• One pass per depth complexity
• Result not easy to use (and several textures)

14
Goals of our approach

• Representation easy to use on GPU
• Result in one pass
• Handle depth complexity problem

15
Previous Work on GPU Voxelization

• Dong, Chen, Bao, Zhang and Peng
• Pacific Graphics 2004,
• Real-time Voxelization for Complex Models
• Same key idea for representation
• Comparison later...

16
Our contribution(s)

• Simple and compact representation
• GPU adapted (especially in future)
• Very fast
• 1.000 K polygons
• 1024 x 1024 x 128
• At approx. 70Hz (0.02 seconds)
• Three sample applications

17
The price of speed

• Less information
• basically 1 bit
• trade-off grid resolution ltgt information
• Not accurate
• not watertight
• For most applications this is not problematic
• Not hierarchical

18
Lets voxelize ...
19
Principle - The key idea

• Represent a 3D grid as a 2D image

slicemap

• Use GPU to fill grid by rasterizing

20
Principle - Voxel encoding

• Associate bitmask to distance of fragment

21
Principle - Voxel encoding

• Fragments are ORed

using OpenGL LogicOps
22
Principle - Voxel encoding

• Shader extremely simple (lt10 instructions)

23
Principle - Voxel encoding

• 32 slices for 4 channels (RGBA) with 8 bits

24
Principle - Voxel encoding

• How can we do better?

25
Technical details - Precision

• Using MRT 432 128 in a single pass

26
Principle - Grid
27
Technical details Local precision
uniform
28
Technical details Local precision

• Tighter interval using two depth maps

29
Technical details Local precision
local
30
Technical details Local precision

• Very efficient
• 1 or 2 depth renderings (FastZBuffer)
• still simple shader (2 supplementary texlookups)

31
Limitations
32
Limitations

• Max resolution 4096 x 4096 x 128
• Only rasterizable objects
• Different representation along Z-axis
• but can be beneficial

33
Comparison to Dong et al.

• Same encoding in bits
• 3 slicemaps to solve hole problem
• limits resolution or several render passes
• needs fusion into final voxel texture
• No local precision

34
Applications

• Transmittance shadow maps
• Refraction
• Shadow volume culling and clamping

35
Transmittance Shadow Maps - TSM
36
TSM - Related Works

• Deep shadow maps Lokovic and Veach 00
• Opacity maps Kim and Neumann 01
• Nalu Nguyen and Donnelly 05

37
TSM - Related Works

• Deep shadow maps Lokovic and Veach 00
• Opacity maps Kim and Neumann 01
• Nalu Nguyen and Donnelly 05

38
Deep shadow maps

• Shadows for complex scenes
• 1D attenuation function per pixel
• Evaluation like shadow map

39
Deep shadow maps

• Pro
• nice results
• Contra
• far from real-time
• no physical meaning

Images courtesy of Lokovich and Veach, Deep
Shadow Maps, Proc. of Siggraph 2000
40
TSM - Related Works

• Opacity maps Kim and Neumann 01
• Store float opacity in layers (interpolation
  simple)
• Creation of one layer uses whole geometry
• For GPU evaluation (original algo used CPU
  evaluation)
• Currently gt16 layers possible perfomance
  dropKoster04
• Expensive for complex scenes and several layers

41
Transmittance shadow maps Key idea

• Semi-transparent surfaces

Attenuation nb of intersected
surfaces
42
Transmittance shadow maps Key idea

• Transparent objects slicemap

Attenuation nb of intersected
voxels
43
Transmittance shadow maps - Algo

• Input Object with semi-transparent material
• 2 render passes
• Light
• Viewer

163.000 tris
44
Transmittance shadow maps - Algo

• Near/far depth maps of semi-transparent parts and
  local slicemap

depth maps
slicemap
45
Transmittance shadow maps - Shader
46
Transmittance shadow maps - Shader

• Value truncation by position

just a mod
S(p)
47
Transmittance shadow maps - Shader

• Bit count

48
Transmittance shadow maps
Results
49
Transmittance shadow maps

• Trade-off Precision vs. information

50
Deep shadow maps in real time
51
Deep shadow maps in real time

• 3 phenomena

supersampling
partial occlusion (small occluders)
semi-transparent (glass)
volumetric (fog)
52
Deep shadow maps in real time

• 3 phenomena

supersampling
TSM
partial occlusion (small occluders)
semi-transparent (glass)
volumetric (fog)
53
Deep shadow maps in real time

• 3 phenomena

supersampling
TSM
Now...
partial occlusion (small occluders)
semi-transparent (glass)
volumetric (fog)
54
Applications

• Transmittance shadow maps
• Refraction
• Shadow volume culling and clamping

55
Volumetric effects Related Works

• Inspired by Wyman 05
• Refract ray twice
• front face normals and back face normals
• Traversed depth approximated
• precalculation depth map difference

56
Volumetric effects Approach

• Use volume info for refraction / scattering as
  well as attenuation

Lokovic and Veach
57
Volumetric effects - Algorithm

• Suppose closed object
• Goal traversed volume/create filled voxels

Depth map difference works
58
Volumetric effects - Algorithm

• Suppose closed object
• Goal traversed volume/create filled voxels

Depth map difference fails
59
Volumetric effects Naive approach

• Single slicemap from eye and flood fill

60
Volumetric effects Naive approach

• Single slicemap from eye and flood fill
• Problems
• On GPU flood fill is too slow
• Evaluation texture is impossible
• More than 4 Gbyte for 32 slices
• Several faces in one voxel (ambiguity)

61
Volumetric effects Key Idea

• Create two slicemaps
• Front and back facing surfaces
• This can be done in a single pass (MRT)

62
Volumetric effects Key Idea

• Two slicemaps allow incremental volume
  computation via in/out states

63
Volumetric effects - Ambiguity
Even ambiguity can often be solved!
64
Volumetric effects Evaluation

• Evaluation texture for a single channel
• info to resolve ambiguity
• Texture lookups are independent
• Details in paper

65
Volumetric effects

• Results

gt200 fps on a Geforce6800
66
Applications

• Transmittance shadow maps
• Refraction
• Shadow volume culling and clamping

67
Shadow volume culling and clamping

• CC ShadowVolumes
• Lloyd, Wendt, Govindaraju, Manocha 2004
• Culling and Clamping of SV Crow 77
• CC Shadow volumes
• Caster culling

- 1
1
1
68
CC Shadow volumes

• Shadow volume clampingReceiver culling

69
Shadow volume culling and clamping

• Our approach uses the slicemap to perform shadow
  clamping (modified depth function)

70
Shadow volume culling and clamping

• Our approach uses the slicemap to perform shadow
  clamping (modified depth function)

Shadow volumes
CC shadow volumes
71
Shadow volume culling and clamping

• Slicemap with modified depth function

72
Shadow volume culling and clamping

• Our approach uses the slicemap to perform shadow
  clamping (modified depth function)

73
Shadow volume culling and clamping

• How to determine volume extent
• Currently in a shader, 1 pass for 1 slice

74
Shadow volume culling and clamping

• Proposed hardware extension
• OR in area of occluder would be sufficient
• (1 pass for 128 slices)

75
Shadow volume culling and clamping

• Slight Optimizations to improve culling and
  clamping

76
Conclusion

• Fast voxelization of highly complex scenes
• Grid of high resolution 4096 x 4096 x 128
• Completely GPU based
• Input of almost arbitrary nature/topology
• Several applications possible

77
And the future ...?

• 32 bit framebuffers imply 512 slices in 1 pass
• Direct X 10 implies bit operations

78

• Thank you very much
• for your attention!