1. A Hardware-Assisted Hybrid Rendering Technique for Interactive Volume Visualization Brett Wilson Kwan-Liu Ma University of California, Davis Patrick S. McCormick Los Alamos National Laboratory

2. 2 Overview Problems with large-scale volume visualization Hybrid rendering –Hybrid data generation –Storage –Rendering Results Future work

3. 3 Large-scale volume visualization Data: Resolutions are 512 3 (128MB) and higher Commodity PC: 1GB RAM, 128MB video memory Want to display large data on these small computers

4. 4 Combine volume and point rendering Large, slowly varying regions –Hardware volume rendering Small areas of high detail –Point-based rendering Combine the efficiency of both rendering techniques

5. 5 Selected previous work Hardware volume rendering [Cabral 1994] [Wilson 1994] –Multi-resolution [LaMar 1999] [Weiler 2000] –Parallel [Kniss 2001] [Lum 2001] [Lum 2002] Splatting [Westover 1989] –Extensions [Mao 1996] [Mueller 1999] –EWA Volume Splatting [Zwicker 2001][Ren 2002]

6. 6 Generating hybrid data Low-resolution volume Original volume Generated points Region of high error Generating low-resolution volume data Generating points for regions of high error

7. 7 Point selection Error evaluated on a regular grid –Usually same resolution as original data Points generated where error is above a given threshold –Also allows goal-oriented generation

8. 8 Hybrid data generation overview Original data Low-res volume dataPoint data (interpolation) (difference & threshold) Result data

9. 9 Rendering opaque features Original vol.HybridLow-res vol.Point data += Points enhance boundary of opaque feature Transparent area can’t be made more transparent

10. 10 Rendering transparent features Original vol.HybridLow-res vol.Point data += Points enhance boundary of opaque surroundings Transparent feature can’t be made more transparent

11. 11 Dealing with overestimation errors Most important features usually drawn as opaque –Effect minimized Pick a low-resolution volume that is always more transparent than the original –Limits transfer functions –Requires a lot of points

12. 12 Rendering volumetric data Texture-mapped polygons rendered back-to-front Eye Result is the illusion of volume

13. 13 Computing volume appearance Color/opacity –Paletted texture lookup for transfer function value Lighting –Paletted texture lookup for specular/diffuse Register combiners Result

14. 14 Rendering hybrid data Slices of points are interleaved with polygons Eye Each slice of points is loaded into a display list.

15. 15 Computing point appearance Color/opacity –Paletted texture lookup for transfer function value Lighting –Paletted texture lookup for specular/diffuse Scale by error value –Map into transfer-function space with 2D texture lookup Register combiners Result

16. 16 Hybrid data storage Volumetric data –1 byte value –1 byte normal Point data –3 to 6 byte position (depending on grid resolution) –1 byte original value –1 byte normal –1 byte error Space: 512 3 = 256 3 + 26 M points

17. 17 Results Simulation –Argon bubble simulation, [Lawrence Berkeley National Lab] Medical –MRI of a human chest, [Kubota Co., Japan] Mechanical –Furby ® (mechanical toy) CT scan, [Los Alamos/Hytec] Test machine: –1GHz Pentium III Xeon, 1GB RAM, 128MB GeForce 4 Ti 4600

18. 18 Results: 512 3 Argon bubble simulation Area of focusFull frame (one of a time-varying simulation)

19. 19 Results: 512 3 Argon bubble simulation Original: 512 3 268MB Low-res volume: 256 3 33MB Hybrid: 256 3 + 361K points 37MB (1/64 error threshold )

20. 20 Results: 512 3 Chest MRI

21. 21 128 3 +11M Hybrid 128 3 + 4M points 1/12 error threshold, 512 3 grid 40MB 0.29 s/frame 128 3 volume 4MB 0.01 s/frame 512 3 volume 268MB 0.35 s/frame

22. 22 Hybrid 256 3 + 7M points 1/64 error threshold, 512 3 grid 97MB 0.47 s/frame 256 3 volume 33MB 0.04 s/frame 512 3 volume 268MB 0.35 s/frame

23. 23 Results: 512  512  2048 Furby ® Mechanical data –Many sharp edges –Very high dynamic range –Very high resolution Full size = 1 GB (including normals) Non-square voxels

24. 24 Results: 512  512  2048 Furby ® 256 3 33MB 0.07 s/frame 256 3 + 3M points (512 3, 1/16 error) 59MB 0.36 s/frame

25. 25 Results: 512  512  2048 Furby ® 256 3 + 3M points (512 3, 1/16 error) 59MB 0.36 s/frame 256 3 + 4.7M points (1024 3, 1/8 error) 71MB 0.58 s/frame

26. 26 Future enhancements Non-cubic error sampling interval –View dependent Automatic parameter selection Optimize point size/opacity Incremental point loading and rendering

27. 27 Conclusions New way to reduce data size for previewing –Preserves fine details Allows very large data to be viewed on small computers –Acceptable performance Effective for simulation, medical, and mechanical data

28. 28 Acknowledgements Los Alamos National Laboratory DOE SciDAC NSF contract ACI 9983641 (PECASE Award) LSSDSV contract ACI 9982251 Data –Argon bubble: Center for Computational Sciences and Engineering at the Lawrence Berkeley National Laboratory –Chest MRI: Dr. H. Miyachi at Kubota Co., Japan –Furby ® : Anthony Davis at Hytec Inc. and Bill Ward of the Los Alamos National Laboratory