Smart HardwareAccelerated Volume Rendering

1
Smart Hardware-Accelerated Volume Rendering

• Stefan Roettger
• Stefan Guthe
• Daniel Weiskopf
• Wolfgang Strasser
• Thomas Ertl

2
Overview

• Current state of the art in direct volume
  rendering
• What can be improved?
• Rendering of segmented data
• Hardware-accelerated raycasting

3
Direct Volume Rendering

• 3D slicing approach (Akeley 87)
• Pre-integration (Max VolVis 90, Roettger VIS
  00, Engel HWW 01)
• Pre-integrated material properties (Meissner GI
  02)
• Hardware-accelerated pre-integration (Roettger
  VolVis 02, Guthe HWW 02)
• Multi-Dimensional TF (Kniss VIS 01)
• Volume clipping (Weiskopf VIS 02)

4
What is missing?

• From a medical point of view
• Pre-integration is difficult to apply to
  segmented medical data
• Pre-integration quality is still not good enough
• 8 bit frame buffer produces artifacts on consumer
  graphics hardware

5
Pre-Integration

• Ray integral depends on three variables Sf, Sb,
  and l, where l is assumed to be constant
• Pre-compute a table for all combinations of Sf
  and Sb and store it in a 2D dependent texture

6
Volume Clipping

• Use additional scalar clip volume C(x,y,z)
• Iso surface for C0.5 defines clip geometry
• Adjust Sf, Sb, and l according to clip volume
  (naive approach set l0)
• for the case Cflt0.5ltCb
• w Cb-0.5/Cb-Cf
• Sf (1-w)SbwSf
• l lw ? ? ?w

C0.5
7
Pre-Integration Segmentation

• Segmentation with two materials is easy
• Define second transfer function TF2
• In the pixel shader
• Make a lookup in TF1
• for the blue area
• Blend with the lookup in TF2
• for the grey area

C0.5
8
Quality Comparison
naive clipping
clipped Bonsai
correct adjust- ment
9
Undersampling Quality
Slicing artifacts
10
Undersampling Quality
Slicing artifacts
11
Sampling Quality
Slicing artifacts
12
Sampling Quality
Interpolation artifacts
13
Supersampling Quality
Still minor interpolation artifacts
14
Supersampling Quality
Almost correct
15
Drawback of Pre-Integration

• Linear interpolation assumed in slab
• But in fact the interpolation is trilinear
• Inside the slab one may cross a voxel boundary
• Lighting is also a non-linear operation
• Conclusion For superior quality we need at least
  2-times, better 4-times oversampling!

16
Ray Casting

• Supersampling is slow, but fortunately we do not
  need to supersample everywhere
• Define importance volume which tells where to
  sample more precisely
• Depends on 2nd deriv. of scalar volume and 1st
  deriv. of TF
• Perform adaptive ray casting on the graphics
  hardware

17
Hardware-Accelerated Ray Casting

• Implemented on the ATI Radeon 9700 with multiple
  floating point render targets
• Need to process all pixels at once
• Cannot exploit ray coherence
• Early ray termination by early Z-test
• Exploit hierarchical Z-buffer compression
• Adaptive sampling includes space leaping
• Stop if all pixels are terminated (asynchronous
  occlusion query)

18
Hardware-Accelerated Ray Casting

• Store ray parameter to determine actual position
• Complete PS 2.0 code given in the paper

19
Quality Comparison
4-times oversampling 8 bit frame buffer
HW ray casting full floating point
20
Performance

• Same performance as 4 times over-sampling with
  alpha Direct 9 drivers (about 2 seconds per
  frame)
• But already much better quality
• With latest drivers we achieve 2-5 frames per
  second due to greatly improved performance of
  occlusion query (12 ms vs. 100 ms)

21
ANFSCD The Bonsai
Note Raw data of all three Bonsai is available
on my homepage
22
Conclusions

• We have shown how to combine volume
  clipping/segmentation with pre-integrated volume
  rendering
• With respect to quality HW ray casting is
  superior to the traditional slicing approach and
  with latest drivers is also faster
• By reducing the number of adaptive samples frame
  rates can be pushed even higher while maintaining
  good quality
• Now switching to the Live Demo

23
Fin

• Thanks for your attention!