Multiresolution Volume Rendering of Large TimeVarying Data Using Videobased Compression

1
Multi-resolution Volume Rendering of Large
Time-Varying Data Using Video-based Compression

• Student Chia-Lin Ko
• Advisor Prof. Jung-Hong Chuang
• Prof. Tsai-Pei Wang

2
Outline

• Introduction
• Related work
• Pre-processing
• Run-time rendering
• Result
• Conclusion
• Future work

3
Introduction

• Volume rendering
• Static volume v.s. time-varying volume

4
Introduction

• Texture-based volume rendering
• Problem large volume data

5
Introduction

• Motivation
• Break the hierarchical decompression dependency
  in the conventional hierarchical wavelet
  representation methods
• Provide interactive playback

6
Introduction

• Contribution
• Present a new framework that combines the
  multi-resolution hierarchical representation with
  video-based compression to manage and render
  large scale time-varying data.
• LOD selection by region-of-interest(ROI)
• Caching and pre-loading
• Efficient decompression along time aixs

7
Related Work

• Static volume data
• Multi-resolution Rendering
• Wavelet tree
• Time-varying volume data
• WTSP tree
• MPEG compression on time-varying data

8
Related WorkMulti-resolution Rendering

• Construct a multi-resolution hierarchical
  representation
• Adapt the data resolution to render the
  interesting or important regions with higher
  accuracy

LaMar E., Hamann B., Joy K. I. Multiresolution
Techniques for Interactive Texture-Based Volume
Visualization. In IEEE Visualization '99 Weiler
M., Westermann R., Hansen C., Zimmermank K., Ertl
T. Level-Of-Detail Volume Rendering via 3D
Textures. In Proceedings of IEEE Volume
Visualization and Graphics Sympsium 2000 Boada
I., Navazo I., Scopigno R. Multiresolution
Volume Visualization with a Texture-Based Octree.
The Visual Computer 17, 3 (2001)
9
Related WorkMulti-resolution Rendering

• Advantage
• More efficient rendering of large volume data
• Problems
• Require frequent Disk I/O when the data size is
  even larger than main memory.

10
Related WorkWavelet tree

• Hierarchical wavelet representation

Guthe S., Wand M., Gonser J., Straßer W.
Interactive Rendering of Large Volume Data Sets.
In IEEE Visualization '02 (2002)
11
Related WorkWavelet tree

• Advantage
• Reduce storage space and the number of disk I/O
• Disadvantage
• Decompressing overhead

12
Related WorkWTSP Tree

• Construct a wavelet-tree at each time step, then
  apply 1D wavelet transform along the time axis

Wang C., Shen H. W. A Framework for Rendering
Large Time-Varying Data Using Wavelet-Based
Time-Space Partitioning (WTSP) Tree. Technical
Report OSU-CISRC-1/04-TR05, 8 pp., January 2004
13
Related WorkWTSP tree

• Advantage
• Flexible spatio-temporal multi-resolution data
  browsing
• Problem
• Updating along the time axis requires the
  traversal of the binary time tree
• Additional decompression overhead
• Getting worse when the number of time steps is
  increasing.

14
Related WorkMPEG compression on time-varying data

• Directly Borrow the idea of MPEG compression to
  compress time-varying volume data set.
• They are not designed for handling large
  time-varying data set
• Guthe and Straßer Real-Time Decompression and
  Visualization of Animated Volume Data 2001
• Sohn et al. Feature Based Volumetric Video
  Compression for Interactive Playback 2002

15
Pre-processing
16
Basic video compression
17
Pre-processing overview

• 1. Classify each time step as I-frame or P-frame
• 2. Subdivide into blocks
• 3. Apply Hierarchical wavelet transform
• 4. Compress and store
• (a) I-frame wavelet tree
• (b) P-frame motion compensation

18
(No Transcript)
19
I-frame compression

• Wavelet Haar wavelet transform with lifting
  scheme
• The high-pass filtered coefficients are then
  encoded using run-length encoding combined with a
  fixed Huffman encoder

20
P-frame compression
21
P-frame compression

• 1. Subdivide a block node into micro-blocks of
  size
• Li x Lj x Lk
• 2. For each micro-block
• (a). Apply motion-compensation-based prediction
  from its
• corresponding spatial node in the previous
  frame.
• (b). Prediction from its neighboring voxels
• 3. Encode difference data motion vectors using
  run-length encoding combined with a fixed Huffman
  encoder.

22
Run-Time Rendering

• LOD selection
• Run-time decompression
• Rendering of blocks
• Caching
• Pre-loading

23
LOD selection

• The LOD selection is dependent on the
  user-controlled region of interest (ROI)
• Two types of ROI
• Spatial - to decide the spatial resolution
• Temporal - to decide the temporal updating
  frequency

24
Spatial ROI

• let the regions inside the ROI have the highest
  resolution

25
Temporal ROI
The blocks that are inside the temporal ROI
bounding box will be updated at every frame
26
Run-time decompression

• I-frame
• Traverse the wavelet tree and reconstruct data
  blocks

27
Run-time decompression

• P-frame
• Decode the difference data and motion vectors to
  recover the data blocks of P-frame.

28
Rendering of blocks

• Use texture-based volume rendering
• Enforcing a back-to-front traversal order of the
  octree.
• For each block, a 3D texture is created and
  loaded into the texture memory.
• Place view-aligned slices into the block and
  render these slices in back-to-front order.
• Alpha blending delivers the volume integrals
  along viewing rays for all pixels on the screen.

29
Rendering of blocks
30
Caching

• Caching helps us to reuse previously
  reconstructed data blocks
• save the time for loading and reconstructing
  these data blocks

31
Pre-loading

• The decompression workload of I-frame is usually
  much larger than the workload of P-frame
• An obvious delay is observed in the playback when
  meeting an I-frame.
• To distribute the workload more evenly, at
  P-frame we can pre-load the data blocks of the
  next I-frame in advance. This will make the
  playback smoother.

32
Result

• The algorithm was implemented in C and OpenGL.
• All benchmarks were performed on a 2.4GHz Intel
  core 2 processor with 2GB main memory , and an
  nVidia GeForce 8800 GTX graphics card with 768MB
  video memory.

33
Resultdata set

• The time-varying data sets used in our testing is
  Turbulent Combustion Simulation data set from
  Institute of Ultra-Scale Visualization (IUSV)

34
(No Transcript)
35
Resultpre-processing

• In Jet_chi data set, we chose the block size to
  be 128x256x32 and each block is of size 2MB. It
  will construct a 3-level hierarchy octree with 73
  nodes.
• In Jet_merg data set, we chose the block size to
  be 64x128x32, and each block is of size 0.5 MB.
  It will construct a 5-level hierarchy octree with
  4681 nodes.

36
Resultpre-processing

• We test the compression ratio with different
  number of I-frames in Jet_chi data set.
• We set all compression parameters the same to
  compress the Jet_merg data set with WTSP tree
  method and our algorithm.

37
ResultRun-time rendering

• Test decompression efficiency with Jet_merg data
  set.
• Test the decompression time and disk loading
  bandwidth under two different situation
• The decompression time includes
• disk I/O to fetch compressed data
• decoding the compressed bit streams
• reconstruction of data blocks

38
Situation 1
39
Situation 2
40
Result Playback frame rate
41

• S1

S1
S2
S1
S2
42
S3
S3
43
ResultVideo demo

• Browsing
• Interactive playback

44
Conclusion

• A new framework is proposed
• combines the multi-resolution hierarchical
  representation with video-based compression
• efficient reconstruction of data in time axis.
• interactive playback is possible.

45
Future work

• Adjust the spatial and temporal LOD by data
  distortion
• Adapt the current advanced video compression
  techniques to improve the compression and
  decompression efficiency.

46
Thank You