Compression Domain Volume Rendering

1
Compression Domain Volume Rendering

• Jens Schneider
• Rüdiger Westermann
• Technical University Munich

2
Motivation

• Need to deal with data of increasing size
• Large-scale
• Multi-dimensional
• Multi-parameter
• Increasing problems
• Compression
• Representation
• Rendering
• We will adress all three problems!

3
Talk Outline

• The Approach Vector Quantization
• Contributions
• Quality and speed
• Hierachical encoding
• PCA-Split
• Progressive encoding of time-resolved data
• Multi-dimensional data
• Vectors of arbitrary length
• Rendering from compressed data
• GPU-based decoding and rendering
• Per-fragment evaluation
• Interactive framerates

4
Talk Outline

• The Application Volume Rendering
• Large-scale volumetric data sets
• Time-varying sequences

1.4 GB / 20 fps
16 MB / 14 fps 0.78 MB / 11 fps
70 MB / 24 fps
5
Vector Quantization
6
Vector Quantization

• LBG-Algorithm
• Linde, Buzo and Gray 1980
• Iterative refinement of a previous Codebook
• Sensitive to quality of first Codebook
• Usually computationally expensive
• Speed-Up possible (and necessary)
• Partial searches
• Fast searches
• Better initial Codebook (i.e. PCA-Splits)
• LBG-Algorithm can be fast!

7
Vector Quantization

• The PCA-Split
• Lensch et.al. 2001 BRDF Compression
• Covariance analysis to find optimal splitting
  plane
• Cut a cluster of input vectors in two by this
  plane.
• Plane is given by centroid of current set and
  largest Eigenvector ( normal) of the
  Auto-Covariance Matrix

8
Vector Quantization

• LBG as PCA post-processing
• Increases fidelity
• Leads to stable Voronoi-Regions
• Only a few steps are necessary
• Great speed-up compared to LBG only!
• A series of LBG steps, codebook from last slide

9
Example

• Full-color confocal microscopy scan, 5122x32xRGB?

10
Hierarchical Vector Quantization
Laplace Decomposition
11
Hierarchical Vector Quantization

12
Hierarchical Vector Quantization

• Output
• One RGB Index-Volume
• Two Codebooks

RGB Index-Volume ? 3D Texture Codebooks ? 2D
?-Textures
13
Example

• Visible Human (Male), RGB slice 2048x1216
• Compression took 10.0 seconds, PSNR 34.72dB

Original (7.1MB)
Compressed (285KB)
14
Timings

• Reference System P4 2.8GHz, 1GB memory
• VHP Slice, 2048x1216 RGB 10.0 sec
• Engine 2562x128 CT-Scan 19.0 sec
• Skull 2563 CT-Scan 50.6 sec
• Vortex Sequence, 1283x100 13 (5) min
• Shockwave Sequence, 2563x89 29 (13) min

15
Rendering

• GPU-based decoding
• Indices stored in 3D RGB-texture (3/64th original
  size)
• Decode index per block ? dependent fetch
• Decode adress per block ? 43 adress texture

16
Rendering

• Render 3D index and adress texture
• Nearest neighbor interpolation for both
• GL_REPEAT for adress texture
• Per-fragment decoding
• Decode detail components and dependent fetch
• Add the details to average component (Red
  channel)
• Lookup result in 1D RGB? transfer function
• Problem
• Complex fragment shader slows down rendering

17
Rendering

• Solution Deferred Fragment Processing
• Avoid decoding in empty regions. Empty means
• a) ? -Transfer function maps 0 ? 0.
• Check on CPU
• Switch between two possible rendering modes
• b) Average value is 0 (Red channel)
• Check in a first, simple fragment program
• Fragments depth value is set accordingly
• Second pass discard (early Z-Test) or render
  fragment
• Full decoding only performed in second pass

18
2562x128 Engine CT Scan

• 19.0 seconds, PSNR 36.17dB (P4 2.8GHz)

Original (8MB) 19 fps
Compressed (402KB) 12 fps
19
2563 Skull CT Scan

• 50.6 seconds, PSNR 35.35dB (P4 2.8GHz)

Original (16MB) 14 fps
Compressed (780KB) 11 fps
20
Time-resolved Sequences

• Exploit temporal coherences during compression
• Group of Frames (GOF)
• First frame in a GOF
• PCA-Split followed by LBG-Refinement
• Other frames
• LBG-refinement of last Index-Volume and Codebook
• Result
• Great speed-up (factor 2 to 3)
• Very large GOFs possible (64 frames)
• Virtually same fidelity as frame-by-frame

21
1283x100 Vortex-Simulation

• 5 minutes, PSNR 34.43dB (P4 2.8 GHz)

Original (200MB) - 28 fps
Compressed (11MB) - 16 fps
22
2563x89 Shockwave-Sequence

• 13 minutes, PSNR 51.36dB (P4 2.8 GHz)

Original (1.4GB) - 20 fps
Compressed (70MB) - 24 fps
23
Conclusions

• Compression ratios of approx. 201
• Interactive rendering possible
• Easy random access to each frame
• Wide variety of data sets handled
• Currently only nearest neighbor interpolation
• Mainly limited by performance / instruction
  count.
• Tri-linear interpolation can be done on newer
  GPUs!

24
Online Demo

• Demo-machine with ATi 9800 Pro
• kindly provided by ATi

25
Thank You!

• Questions ?