CS690 Vis Papers

1
CS690 Vis PapersView Selection for Volume
RenderingA Feature-Driven Approach
toLocating Optimal Viewpoints for Volume
VisualizationImportance-Driven Volume
Rendering

• Joshua New

2
Vis Paper I
View Selection for Volume Rendering Udeepta
D. Bordoloi, Han-Wei Shen The Ohio State
University
3
Vis Paper I

• Problem Find best N viewpoints (polygonal
  literature but techniques not suited to volume
  rendering)
• Assumptions
• Data is centered at the origin
• Camera always looking at origin from a fixed dist
• Absorption/Emission optical model
• Three Viewpoint Characteristics
• View Goodness (more important voxels are highly
  visible)
• View Likelihood (voxel visibilities similar
  within a threshold)
• View Stability (max view change within threshold
  of camera shift)

4
Vis Paper I

• Similarity/Contrast with the next paper
• Both based on entropy function
• Entropy based on voxels rather than weighted
  averages of isosurfaces or interval volumes
• Other paper claims this is 2D and theirs is 3D
  (think about whether or not you agree)
• Viewpoint Evaluation
• Volume ray-integration
• Information Theory

5
Vis Paper I

• Information Theory
• Random set of J symbols ai with probability pi
• Information of symbol
• Information of sequence
• Entropy, avg information,maximized when ai1/J

6
Vis Paper I

• The Math
• Noteworthiness Wi of voxel vi
• Voxel probability fi from frequency of histogram
  bin
• Visual probability of voxel vi
• Entropy, average information, of view V

7
Vis Paper I

• The Math Works

8
Vis Paper I

• Handling Different Views
• Problem Need visibilities to be same regardless
  of viewing angle (from center of voxel)
• Ray-casters calculate opacity along the ray
• Texture-based renderers from frame-buffer pixel
  locs
• Solution GPU acceleration (fragment shader)
• Align obj to axis most perpendicular to viewing
  plane
• Data stored with floating point p-buffer
• Iteration P/frame clear, camera aligned w/
  current slice, previous slice rendered relative
  shear, frag combines opacities to
  frame-buffer/texture
• Entropy equations now calculated

9
Vis Paper I

• Best and Worst Views (uniformly sampled sphere)

10
Vis Paper I

• Handling Different Views (cont.)
• View (Dis)Similarity
• Kullback-Leibler (KL) Distance (for 2
  distributions p and p)
• Jansen-Shannon Divergence
• Partition
• Partition view sphere based on view similarity
  based on JSD (user interactable)

11
Vis Paper I

• Different Views Handled
• View with highest entropy represents a
  partition
• Problem Partition boundary problem (close
  viewpts)
• Solve by greedily taking best entropy view until
  sufficiently far from other partitions best view

12
Vis Paper I

• Handling time-sequences
• Assume Markov Process
• Time-dependent noteworthiness
• Parameter k highlights change
• Visual probability
• Entropy of view V over all time

13
Vis Paper I

• Optimizations
• Noteworthiness
• Dont calculate if close to 0 (low opacity)
• View Similarity
• Jansen-Shannon Divergence rewritten in terms of
  entropy and use previously calculated terms
• Time Dependence
• Dont calculate if change between time is close
  to 0

14
Vis Paper I

• Results (128 views, 2Ghz P4, 8xAGP GeForce5600)
• Speed 1283 data visibility 128 views 42s (3 fps)
• Tooth data 128x128x80

15
Vis Paper I

• Results
• Vortex data 1283x14

16
Vis Paper I

• Shockwave data 2563x50

17
Vis Paper II
A Feature-Driven Approach to Locating Optimal
Viewpoints for Volume Visualization Shigeo
Takahashi, Issei Fujishiro, Yuriko Takeshima,
Tomoyuki Nishita The University of Tokyo
Tohoku University
18
Vis Paper II

• Method
• Based on (Shannon) entropy as well, but always
  normalize
• Max entropy
• Uniformly sample view sphere
• Subdivide icosahedron twice withLoop subdivision
  rule
• Icosahedron is 20-sided polygon(DD players
  companion)

19
Vis Paper II

• Nice interface (visual partitions)

20
Vis Paper II

• Isosurfaces
• Uniformly sample data range (pick isovalue)
• Multiple (32) isosurfaces per view
• Weight entropy of each isosurface based on the
  opacity transfer function of the isovalue
• Shortcomings
• Unjustified sampling
• Oblivious connected components bw isovalues
• Neglects thickness of the volume

21
Vis Paper II

• Interval Volumes
• Subvolume from integrating a connected component
  to an isosurface within some range
• Interval Volume Decomposer (IVD) creates
  level-set graph from changing number of
  isosurface components
• Apply same equations forisosurface entropy to
  calcinterval volume entropy

22
Vis Paper II

• Weighted Unweighted Isosurfaces IVs

23
Vis Paper II

• Entropy based on transfer function

24
Vis Paper II

• Neighboring Interval Volumes
• Decomposed IV is a link in a level-set graph,
  emphasize certain IV combinations
• Each IV entropy computed separately so have to
  account for overlap

25
Vis Paper II

• Results

26
Vis Paper II

• Results

27
Vis Paper II

• User Study
• 32 grads in CGVis, 42 viewpts (Users
  color-coding)

28
Vis Paper III
Importance-Driven Volume Rendering Ivan
Viola, Armin Kanitsar, Meister Eduard
Groller Institute of Computer Graphics and
Algorithms Vienna University of Technology,
Austria (some pictures excerpted from Violas
PhD Thesis)
29
History

• CS594FT book (Computer Vision)

30
History
31
History
32
History

• VOI Visualization Exploded Views

33
History

• VOI Visualization Exploded Views

Contextual Zoom Demo
34
Vis Paper III

• VOI Visualization

Planar Cut-away
Opacity TF
Multi-Volume Peel
35
Vis Paper III

• Motivation
• Larger datasets but relatively small VOI
• Provide focuscontext view of data
• Clinical Use
• Time-consuming to change transfer functions when
  view changes to highlight VOI
• Assumptions
• Dataset is pre-segmented with importance
  dimension assigned to each voxel

36
Vis Paper III

• Method (introduce view position to pipeline)
• Use ray-casting for rendering (not real time)
• Object importance constant scalar value
• Importance compositing
• Level of sparseness inverse screen footprint

37
Vis Paper III

• Teaser

38
Vis Paper III

• Importance Compositing
• Maximum Intensity Projection (MIP)
• Selecting highest intensity value along a ray
  (best for sparse/contrasted data)
• Max Importance Projection (MImP) (volume
  cut-away)
• Fully transparent between camera position and
  object of highest importance (use object
  grayscale instead of binary)
• Use general cylinder as clipping object and sweep
  across footprint of most important object
• Countersink from scaling as gets closer to camera
• Change starting point of rays intersecting the
  clipping frustrum

39
Vis Paper III

• MImP illustration

40
Vis Paper III

• Importance Compositing
• Average Importance

41
Vis Paper III

• Importance Compositing
• Visibility-preserving (MImP)

42
Vis Paper III

• Sparseness
• Change saturation to highlight (occlusion problem)

43
Vis Paper III

• Sparseness
• Smoothly interpolate RGBA

44
Vis Paper III

• Sparseness
• Screen Door Transparency

45
Vis Paper III

• Sparseness
• Volume Thinning (reduce to few isosurfaces based
  on gradient magnitude and curvature magnitude)

46
Vis Paper III
47
Vis Paper III
48
Vis Paper III

• DEMO VIDEOS
• (http//www.cg.tuwien.ac.at/research/vis/adapt/200
  4_idvr/)