1. Hardware-Accelerated Adaptive EWA Volume Splatting Wei Chen ZJU Liu Ren CMU Matthias Zwicker MIT Hanspeter Pfister MERL

2. 2 Volume Splatting Object-order method 3D reconstruction kernel centered at each voxel (elliptical Gaussian) Voxel contribution = 2D footprint (color, opacity) Weighted footprints accumulated into image Voxel kernels Screen 2D footprints = splats

3. 3 Speed Quality Software Texture splats Axis-aligned EWA Swan Image-aligned Fast splats OpenGL ex Westover1989 Crawfis 1993 Swan 1997 Mueller 1999 Huang 2000 Zwicker 2001 Xue 2003 Related Work Our work

4. 4 Outline EWA volume splatting Adaptive EWA splatting GPU implementation Results and conclusions

5. 5 EWA Volume Splatting Compensate aliasing artifacts due to perspective projection EWA Filter = low-pass filter warped reconstruction filter W Volume Low-Pass Filter EWA volume resampling filter Projection Convolution

6. 6 EWA Volume Splatting (512x512x3) Reconstruction filter only: 6.25 fps EWA filter: 4.97 fps Low-pass filter only: 6.14 fps EWA filter: 3.79 fps

7. 7 Analysis of EWA Filter Warped recon- struction kernel Low-pass filter Resampling filter Minification Magnification

8. 8 Analysis of EWA Filter Shape of EWA Splat is dependent on distance from the view plane r k Reconstruction filter radius u 2 Distance to the view plane r h Low-pass filter radius EWA splat Note that

9. 9 Adaptive EWA Filtering Warped recon- struction kernel Low-pass filter Resampling filter if u 2 > A use low-pass filter if u 2 < B use reconstruction filter if A<u 2 < B use EWA filter

10. 10 Patch Processing Process a 8 x 8 patch of voxels at a time Filter selection based on four corners of each patch (choose smallest) Traversal order Patch Distance

11. 11 Adaptive EWA Volume Splatting (512x512x3) Adaptive EWA filter: 6.88 fps EWA filter: 4.83 fps Adaptive EWA filter: 1.84 fps EWA filter: 1.75 fps

12. 12 Outline EWA volume splatting Adaptive EWA splatting  GPU implementation Results and conclusions

13. 13 Object-Space EWA Splatting Object-space EWA splatting with texture mapping [Ren et al. Eurographics 2002] Texture (unit Gaussian) Unit quad (0,0)(1,0) (0,1) (1,1) EWA Splat (elliptical Gaussian) Projection Texture mapping Textured quad Vertex shader computation

14. 14 Proxy Geometry Template Rectilinear volumes: use one proxy geometry template for all slices in each direction Store vertex indices in AGP memory... Regularity Voxel geometry Proxy geometry template Quad geometry

15. 15 Vertex Compression Compress each vertex to 32 bits Decompression on-the-fly in programmable hardware To store vertex information of 256x256x256 volume in video memory Without compression 2,048 MBytes  With compression 12 MBytes Retained-mode hardware acceleration feasible

16. 16 Retained vs. Immediate Mode Data#Total Splats #Rendered Splats Immediate Mode Retained Mode Bonsai83886082748660.53 fps7.53 fps Engine72089602475771.40 fps10.28 fps Lobster54613445559761.19 fps10.60 fps Head1363148829552420.12 fps2.86 fps Factor of ~10 improvement

17. 17 Interactive Classification: Opacity Culling Hardware-accelerated list-based traversal For each slice For each 32 x 32 patch of voxels (smaller indices) Indices of proxy geometry organized into iso-value lists using bucket sort; CPU merges lists on-line Render only iso-value lists with visible voxels Patch 0 128 256

18. 18 Interactive Classification: Opacity Culling DataList-based opacity culling Standard opacity culling Head2.80 fps0.3 fps Engine10.18 fps0.8 fps Bonsai7.23 fps0.8 fps Lobster10.30 fps1.1 fps Includes changes to TF every frame Factor of ~10 improvement

19. 19 Deferred Shading Volume texture access is only possible in fragment programs * However, per fragment shading is expensive Solution: deferred shading in two passes * Newer GPUs allow texture access in vertex programs

20. 20 Deferred Shading Pass one: 3D texture access, classification and illumination in vertex shader, render one pixel per voxel Pass two: reuse the pixel data from the first pass to shade the 2D footprint Performance gain: 5%-10% speedup Pass one Pass two Final result

21. 21 Experiments P4 2.4 GHz ATI 9800 Pro with 256 MB RAM Direct3D 9.0b with VS 2.0 and PS 2.0 TypeEWAReconstruction OnlyLow-pass Only Regular704526 Rectilinear744526 Vertex shader instructions

22. 22 Sheet-buffer Composition Axis-aligned traversal, addition in sheet buffers, then blending front-to-back 0.80 fps 3.00 fps 3.45 fps

23. 23 UNC Head: 208x256x225 2.86 fps #Rendered splats: 2,955,242 8.5M splats / sec

24. 24 Bonsai: 256x256x128 7.53 fps #Rendered splats: 274,866 2M splats / sec

25. 25 Engine: 256x256x110 10.28 fps #Rendered splats: 247,577 2.5M splats / sec

26. 26 Lobster: 301x324x56 10.60 fps #Rendered splats: 555,976 5.9M splats / sec

27. Video

28. 28 Our Contributions Adaptive EWA computation Volume data compression Retained-mode hardware acceleration Interactive opacity culling Deferred two-pass shading

29. 29 Future Work Image-aligned EWA volume splatting Irregular volume splatting Pointsprites in OpenGL Floating point textures Vertex texture for classification

30. 30 Acknowledgementshttp://graphics.cs.cmu.edu/projects/adpewa/index.html Jessica Hodgins (CMU) Markus Gross (ETH)