VisibilityDriven View Cell Construction

1
Visibility-Driven View Cell Construction

• Oliver Mattausch, Jirí Bittner,
• Michael Wimmer

Institute of Computer Graphics and
Algorithms Vienna University of Technology
2
Motivation
BSP Teller 92
Our method
Render Time
Low
High
3
Introduction

• Visibility preprocessing
• Partition view space into view cells
• Compute PVS for each view cell
• Runtime
• Locate the view cell
• Render PVS
• View cell construction got little attention
• But as important as PVS computation

4
Our Goal

• Create good view cells for general 3D scenes
• What are good view cells?
• Fast rendering
• Low memory consumption

5
Example 1
Render Time
Low
High
6
Example 2
On the groundA few buildings visible
Above the roofsMany buildings visible
gt Should belong to separate view cells
7
Previous Work

• Cells and Portals
• KD tree Airey90
• BSP tree Teller92
• Breaking the walls Lerner03
• Watershed Haumont03
• Implicit construction (during visibility
  computation)
• Adaptive sampling Nirenstein03
• Visibility Octree Saona-Vázquez99
• 5D sampling Gotsman99

8
View Cell Construction

• Main idea minimize expected rendering time

c expected rendering time (cost to minimize)
S set of view cells
p probability of view point inside a view
cell (volume of view cell)
r render time estimate Wimmer03
9
View Cell Construction

• Three steps
• Visibility sampling
• View space subdivision
• View space merging

?
?
1.
2.
3.
10
View Cell Construction

• Three steps
• Visibility sampling
• View space subdivision
• View space merging

?
?
1.
2.
3.
11
Visibility Sampling

• Fast stochastic sampling
• Casts rays from potential viewpoints towards
  scene objects
• Coarse estimate of visibility (1M samples)

View space scene bounding box
View space distant region
12
Visibility Sampling

• Compute maximal free segments
• Optionally detect empty view space

Viewpoint inside the wall
Visibility samples
Invalid
Valid
Valid
13
View Cell Construction

• Three steps
• Visibility sampling
• View space subdivision
• View space merging

?
?
1.
2.
3.
14
View Space Subdivision

• Binary Space Partitioning (BSP)
• Geometry-aligned splits
• Axis-aligned splits
• Greedy optimization select next best split
• 2 stage method
• Compute locally best split within each leaf
• Take globally best split (priority queue)

15
Splitting Plane Selection

• Minimize render cost
• PVS computed from visibility samples

16
Splitting Plane Selection Example
Best plane
C(Left) 6 5
C(Left) 18 7
C(Left) 26 8
C(Left) 28 8
C(Left) 2 4
C(Left) 14 5
C(Right) 28 8
C(Right) 24 7
C(Right) 12 5
C(Right) 4 4
C(Right) 2 2
C(Right) 16 5
C(Left Right) 186
C(Left Right) 224
C(Left Right) 228
C(Left Right) 232
C(Left Right) 178
C(Left Right) 160
17
View Space Subdivision Properties

• Subdivision is progressive
• Intuitive global termination criteria
• Memory limit
• Minimal render cost decrease

18
View Space Subdivision

• Result Set of elementary cells

5000 view cells
19
View Cell Construction

• Three steps
• Visibility sampling
• View space subdivision
• View space merging

?
?
1.
2.
3.
20
View Space Merging

• Bottom-up refinement of the subdivision
• Merge view cells while minimizing the render cost
  increase
• Merge candidates neighboring pairs of view cell

5000 view cells
200 view cells
21
View Space Merging

• View cell hierarchy Merge history tree
• Scalable view cells
• Extracting optimal set of view cells

22
Results Render Cost

• Building (5 floors, 8000 objects)

BSP Teller92, KD-VT Nirenstein04,
KD-cycling-axis Subdivision our method (only
subdivision) Merge our method (subdivision
merge)
Render cost evaluation using 8M samples
23
Results Render Cost

• Vienna (11000 objects, 1M polygons)

BSP Teller92, KD-VT Nirenstein04,
KD-cycling-axis Subdivision our method (only
subdivision) Merge our method (subdivision
merge)
Render cost evaluation (using 10M samples)
24
Results Render Cost Histogram

• Vienna (11000 objects, 8M polygons)

25
Results Timings

• Timings (using unoptimized implementation)
• 1.5 M samples
• 50000 view cells
• 3.4 GHZ Pentium 4

26
Conclusions

• New method for view cell construction
• Handles all types of scenes
• Main ideas
• Visibility sampling
• Render time minimization
• View cell hierarchy
• Lower render cost
• Better distribution of the render cost
• Scalable view cells

27
View Cell Construction

• Thank you for your attention!

28
Results Sampling Density

• Influence of sampling density (Building interior)
• Visibility samples 50K, 200K, 1M

Initial subdivision
Final render cost evaluation (using 8M samples)
Conclusion Coarse sampling is sufficient!
29
View Space Subdivision

• Problem with greedy optimization
• Can run into local minimum using pure render cost
  decrease
• Solution also take render cost of node into
  account

Cut through a scene The leaves of the
subdivision are shown in random color No render
cost decrease for upper hierarchy levels gt one
half of the view space (shown in green) is not
subdivided
30
Results Merge Step

• Another application for the merge stepMake
  depth-first approach progressive

31
Results Geometry-Aligned Splits

• Influence of geometry-aligned split
  planes(Rotated floor of building interior)