Visibility Culling using Hierarchical Occlusion Maps

1
Visibility Culling using Hierarchical Occlusion
Maps

• Hansong Zhang, Dinesh Manocha, Tom Hudson,
  Kenneth E. Hoff III
• Presented by
• Chris Wassenius

2
Outline

• Introduction
• Related Work
• Method
• Results
• Conclusion / Future work

3
Introduction

• GPU power has increased year after year
• So has the size of models...
• Need acceleration algorithms
• Visibility culling
• Level of detail
• texturing

4
Introduction
5
Related Work

• Z-buffer
• BSP (Binary Space Partitioning) Trees
• creation is time consuming
• not dynamic
• PVS (potentially visible set)
• works well for specific models
• doesn't work so well with arbitrary models

6
Related Work

• Object space algorithms (Coorg and Teller, and
  Hudson et al. 1996)
• convex objects only
• can't combine occluders
• Hierarchical Z-buffer algorithm
• Octree and Z-pyramid
• good, but expensive

7
Method

• Features
• Generality
• No restriction on types of occluders
• Occluder Fusion
• combines a forest of small and/or disjoint
  occluders
• Significant Culling

8
Method

• Features (continued)
• Portability
• Efficiency
• algorithm only takes a few milliseconds per frame
• significant speedup in interactive walkthroughs
  of models
• Approximate Visibility Culling
• able to cull small visible holes in occluders

9
Method

• Basic Idea
• Select objects of the model as occluders
• Create hierarchical occlusion maps (HOM)
• Render objects in the model based on
• overlap test with HOM
• depth test

10
Method

• Construction of the Occlusion Map Hierarchy
• Selects occluders from the occluder database
  (preprocessing step)
• traverses the bounding volume hierarchy of the
  occluder database
• selects subset, utilizes temporal coherence
• Occluders are rendered in pure white
• Builds hierarchy by averaging pixels

11
Method

• Occlusion Map
• object is projected to screen, area of projection
  is made opaque
• each pixel records opacity of a rectangular block
  in screen space
• opacity the ratio of the sum of the opaque areas
  in the block to the total area in the block

12
Method

• Occlusion Map Hierarchy
• Recursively average 2 x 2 blocks of pixels
• case use GPU and CPU
• special case of bilinear interpolation
• Stop at some minimal resolution (e.g. 4 x 4)

13
Method

• Overlap Test (to see if an object is occluded)
• Check opacity of the pixels it overlaps in the
  HOM
• Exact overlap test is too expensive
• Use screen spaced bounding rectangle for
  projection

14
Method

• Overlap test (continued)
• uses HOM to accelerate test
• Begins at the level of the hierarchy where the
  size of a pixel is close to the size of the
  bounding rectangle
• Examines each pixel that overlaps rectangle
• If each pixel is completely opaque, object is
  overlapped by occluders
• Else recursively descends to higher resolution
  level
• If all pixels in rectangle are opaque then object
  is overlapped by occluders
• Else algorithm renders object

15
Method

• High level opacity estimation
• If low resolution map pixel has a low opacity
  level, descendants most likely have low opacity
  levels
• If low resolution map pixel has a high opacity
  level, descendants most likely have high opacity
  levels.

16
Method

• Opacity Threshold
• Value at which a pixel is considered completely
  opaque
• Different threshold for each level in hierarchy
• In effect, specifies the size of allowable holes

17
Method

• Approximate Visibility Culling

18
Method

• Depth Test vs. Overlap Test

19
Method

• Depth Estimation Buffer
• partition screen space and use separate Z-plane
  for each region
• estimate depth and position of occluders by
  projected bounding box
• take furthest z-value of projected rectangle for
  each occluder
• for each partition, set distance to the furthest
  occluder

20
Method

• Depth Test (for a potential occluded objection)
• Again, use projected rectangle as aproximation
• Check each partition of the depth estimation
  buffer that is covered by the rectangle
• if any partition is greater than (further than)
  rectangle's depth
• object is rendered
• otherwise,
• object is not rendered

21
Method

• Occluder selection to form Occluder Database
  (preprocessing step)
• Size
• small objects typically don't serve as good
  occluders
• Redundancy
• Complexity

22
Method

• Dynamic Selection of occluders at run-time
• Selects based on distance from view point, size,
  and temporal coherence
• limits the amount of selected occluders (can vary
  per frame)

23
Results
24
Results

• 635,252 polygons
• 82.7 culled

25
Conclusion and Future Work

• Pros
• Good visibility culling algorithm for large
  models with large depths
• Works well with arbitrary models
• Cons
• Fairly large overhead

26
Conclusion and Future Work

• Integrate LOD
• Occlusion preserving simplification