TimeCritical Volume Rendering

1
Time-Critical Volume Rendering

• Han-Wei Shen and Xinyue Li
• Department of Computer and Information Science
• The Ohio State University

2
Volume Rendering
3
Current Challenge

• Big data problem
• Large data size ( 512M) makes interactive
  volume rendering difficult
  
• High computational cost
• High main memory / texture memory
  requirement
  
• Large-scale data has become common for both
  scientific and medical applications

4
Possible Solutions

• Parallel software/hardware algorithms
• Hierarchical data representations
• trade image quality for speed

5
Hierarchical Volume Representation
Error measure
Error measure
Raw Data
6
Quality vs. Speed

• Run-time tradeoff is controlled by the error
  tolerance
  
• Usually specified by the user at run time
• A good error tolerance requires extensive
  knowledge about the data
  
• Difficult to predict the execution time
  associated with a particular error tolerance
  
• View dependent
• Transfer function dependent
• Machine dependent

7
Automatic Control

• Is it possible to control the error tolerance
  automatically?
  
• Goals
• Time-critical computation automatically adapt
  to the user-specified performance goal
  
• Adaptive rendering - render regions of different
  importance using different error tolerances
  
• Algorithm and machine independent

8
Previous Work

• Adaptive level of detail selection
• Two approaches and their challenges
• Reactive difficult to
• Maintain a constant frame rate
• Adjust error threshold appropriately
• Predictive difficult to
• Predict the volume rendering execution time

9
New Approach

• An intra-frame predictive-reactive algorithm
• Break the rendering task into small subtasks
• Monitor the performance of each rendering
  subtask
  
• Predict the rendering time for each subtask based
  on the past experience
  
• Adjust the error tolerance whenever necessary

10
Algorithm Overview
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
11
Algorithm Overview
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
12
Algorithm Overview
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
13
Algorithm Overview
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
14
Algorithm Overview
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
15
Algorithm Overview
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
16
Key Issues

• How to budget the rendering time for each
  subvolume?
  
• How to predict the rendering time for the next
  subvolume?
  
• How to adjust error tolerance?

17
Budget Rendering Time
Render subvolume
Subdivide volume into subvolumes
Budget rendering time for next subvolume
Predict rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
18
Budget Rendering Time

• Based on importance value
• Each subvolume Vi is assigned an
• importance value Ii
• Sum up all the importance values to I
• Given the total rendering time T, then each
  subvolume is allocated
  
• Ti T x Ii / I
• Importance value distance to gaze direction,
  transparency, standard deviation, etc

19
Rendering Time Prediction
Render subvolume
Subdivide volume into subvolumes
Predict rendering time for next subvolume
Budget rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
20
Rendering Time Prediction

• Based on several factors
• Number of voxels (of various resolutions)
• Number of slice planes (for 3D texture mapping)
• Projection area
• Extrapolating the rendering time from past
  experience to predict the next subvolume

21
Adjust Error Tolerance
Render subvolume
Subdivide volume into subvolumes
Predict rendering time for next subvolume
Budget rendering time for next subvolume
Split time budget among subvolumes
Pick an initial guess of error tolerance for t
he first subvolume
Adjust error tolerance
22
Adjust Error Tolerance

• Based on the difference between budgeted Time Ti
  and predicted rendering time Tp
  
• Use Fuzzy Logic Control

Fuzzy Control System
New error tolerance
Tp - Ti
23
Fuzzy Logic Control Method

• Why fuzzy logic control?
• No explicit mathematical model between the system
  input and output is required to achieve the
  control goal
  
• Specifying the control rules is simple using
  linguistic rules which have the form If then
  
  
• General rule
• If the rendering time difference is high
  (low), then the change to the error tolerance
  should be high(low)
  

24
Fuzzy Logic Control
More details can be found in our recent paper
Adaptive Volume Rendering using Fuzzy Logic
Control, Joint Eurographics-IEEE TCVG Symposium
on Visualization, 2001
25
Preliminary Results

• Two small data sets
• Delta wing data set 112 x 128 x 51
• Brain data set 128 x 128 x 72
• We tested the control of two rendering methods
• 3D texture mapping (SGI Octane 4MB Texture
  Memory)
  
• Software ray casting (300MHz R12000, 512 MB RAM)

26
Sample Images (1)
Brain data set
27
Sample Images (2)
Delta wing data set
28
Hardware Rendering Control (1)
Delta wing 10 frames/second goal
Various viewing and
scaling
Fixed threshold
Time-critical control
29
Hardware Rendering Control (2)
Brain 10 frames/second goal
Various viewing and scaling
Fixed threshold
Time-critical control
30
Software Rendering Control (1)
Delta 6 seconds/frame goal
Various viewing and scaling
Fixed threshold
Time-critical control
31
Software Rendering Control (2)
Brain 8 seconds/frame goal
Various viewing and scaling
Fixed threshold
Time-critical control
32
Future Work

• Test large scale data sets (10 x texture/main
  memory space)
  
• Develop and evaluate the performance prediction
  model
  
• Develop and evaluate the error tolerance
  adjustment model
  
• Develop different importance functions
• Apply to parallel and remote software/hardware
  rendering algorithms