RunTime Management for LOD

1
Run-Time Management for LOD
2
Run-Time Issues

• LOD Framework
• discrete, continuous, view-dependent, imposters
• Selection Criteria
• distance, size, environment, gaze
• Popping Prevention
• hysteresis, geomorphs, alpha-blending
• Frame Rate Management
• free v fixed frame rate
• Resource Management
• out of core, network streaming

3
Selection Criteria

• LOD Framework
• discrete, continuous, view-dependent, imposters
• Selection Criteria
• distance, size, environment, gaze
• Popping Prevention
• hysteresis, geomorphs, alpha-blending
• Frame Rate Management
• free v fixed frame rate
• Resource Management
• out of core, network streaming

4
Discrete (aka Static) LOD

• Model each object at several resolutions
• Very simple run-time just switch models
• Can compile models into display lists
• Can use imposters for lower LODs
• Cannot vary resolution over a model
• May not scale well to large object nos.

5
Hierarchical LOD

• Group Discrete LODs into a hierarchy
• e.g., neighborhood ? city block ? building
• Better scalability for large structured models

Images from University of Stuttgart, Institute
for Photogrammetry
6
Continuous LOD

• Sequence of edge collapse / vertex splits
• aka Progressive Meshes (PM)
• Run-time manages verts/polys not models
• Allows fine control over changes in LOD
• Supports progressive transmission

7
Continuous LOD Example

• Progressive meshes
• Hoppe (Microsoft)

8
View-Dependent LOD

• Selective refinement of continuous LOD
• more detail toward viewpoint (for terrain)
• more detail following a path (line of sight)
• David will talk more about this later...

Images from Hoppe (1997)
9
Selection Criteria

• LOD Framework
• discrete, continuous, view-dependent, imposters
• Selection Criteria
• distance, size, environment, gaze
• Popping Prevention
• hysteresis, geomorphs, alpha-blending
• Frame Rate Management
• free v fixed frame rate
• Resource Management
• out of core, network streaming

10
Distance LOD

• Select resolution based upon the distance between
  an element and the viewpoint, i.e. coarser
  resolution for distant geometry.
• Simple to calculate (3-D Euclidean distance)
• Scale dependent
• Resolution dependent
• Field of View dependent

11
Size LOD

• Select resolution based upon the projected screen
  size (or area) of an element. Objects appear
  smaller as they move further away.
• Requires 3-D to 2-D projection
• Scale invariant
• Resolution invariant
• Field of View invariant
• Bounding spheres or ellipsoids normally used
  instead of boxes as more efficient to calculate
  projected extent

12
Environmental Conditions

• Slacken LOD thresholds through use of fog, haze,
  clouds, smoke, etc.
• Only useful if these make sense!

Images from www.benchmark.pl
13
Silhouette Preservation

• Human visual system sensitive to edges
• So, put more detail around silhouettes
• Test if view-to-face vector and face normal are
  orthogonal (use cone of normals)

From Luebke Erikson (1997)
14
Occlusion LOD

• Reduce LOD for parts of a model that are obscured
  by other objects in the current view frustum
• Of course, can also use cull entire objects that
  are occluded too

Images from http//www.cs.unc.edu/geom/HSLOD/
15
Specular Highlights

• Lower LODs affect lighting
• Particularly apparent at specular highlights
• Put more detail at these points to maintain
  fidelity
• Xia Varshney (1996)
• Techniques exist to do this for textured lit
  models
• Luebke et al.

16
Gaze-Directed

• Velocity
• Retina less sensitive to rapidly moving objects
• Eccentricity
• Less sensitivity in our peripheral field
• Depth-of-Field
• Objects blur outside eyes fusional distance
• However require eye-tracking, or head-tracking,
  or make gaze assumptions

17
Gaze-Directed Example

• Gaze-directed Adaptive Rendering
• Ohshima, Yamamoto, and Tamura (Canon Inc.)

18
Attention-Directed

• Work out where the user is likely to be looking
  using models of visual attention
• Image-based (e.g., motion spatial freq)
• Model-based

• Or, control where the user looks through the
  dramatic content of your scenes

Image from H. Yee (2001)
19
Shadow LOD

• Use different shadow techniques
• Use lower LOD models for silhouette texture
• Use pre-generated generic texture
• Multiple passes used softer shadows do fewer
• Reduce off-screen rendering buffer size for
  multipass

12.0 Hz
10.1 Hz
3.0 Hz
Images by OpenWorlds Inc
20
Shader LOD

• Procedural shaders that can adjust their detail
• Bump map to texture map
• Texture map or noise function to single color
• Multi textures to one texture
• BRDF approximations

Image from http//www.sgi.com/software/shader/whit
epaper.pdf
21
Simulation LOD Example

• Modeling Hair Using LOD Representations
• Ward, Lin, Lee, Fischer, and Macri (UNC)

22
Popping Prevention

• LOD Framework
• discrete, continuous, view-dependent, imposters
• Selection Criteria
• distance, size, environment, gaze
• Popping Prevention
• hysteresis, geomorphs, alpha-blending
• Frame Rate Management
• free v fixed frame rate
• Resource Management
• out of core, network streaming

23
Hysteresis

• A lag in the switch between LODs
• Reduce scintillating effect of models
  continuously switching at threshold
• 10 of switch range is good enough

24
Priority Schemes

• Assign priorities to each object
• Reduce high priority objects less
• Account for semantic importance of certain
  objects, e.g., walls, players, etc.

Image from Funkhouser (1993)
25
Alpha-Blending

• Blend between LODs in image space
• Means rendering 2 LODs in fade range
• Can fade over time or distance range
• Render two images as opaque then blend, or may
  get self-occlusions

Image from OpenGL Performer Getting Started Guide
26
Geomorphs

• Blend between LODs in geometry space

• Morph vertices vu and vt toward vs
• Specify frames or time to morph over
• Support for geomorphs in Unreal engine

27
Geomorphs Example

• Progressive Meshes
• Hoppe (Microsoft)

28
Frame Rate Management

• LOD Framework
• discrete, continuous, view-dependent, imposters
• Selection Criteria
• distance, size, environment, gaze
• Popping Prevention
• hysteresis, geomorphs, alpha-blending
• Frame Rate Management
• free v fixed frame rate
• Resource Management
• out of core, network streaming

29
Frame Rate Management

• High, constant frame rates are important for
  interactive 3D games LOD can help
• Adjust LODs further to maintain budget Hz
• Reactive Fixed Frame Rate
• Adjust priorities based on previous frame rate
• Predictive Fixed Frame Rate
• Adjust priorities to meet budget on this frame

30
Reactive Fixed Frame Rate

• Adjust LODs based on previous Hz
• calculate frame rate for previous frame
• compare to desired frame rate for this frame
• adjust LODs to be displayed accordingly
• Does not guarantee fixed frame rate
• Will over/undershoot when there are large changes
  in complexity between frames
• Used by OpenGL Performer

31
Predictive Fixed Frame Rate

• Funkhouser Séquin - SIGGRAPH 1993
• Predictive fixed frame rate for Discrete LOD
• Achieves a bounded fixed frame rate ( ?)
• Based on a constrained optimization of
• Maximize ?s Benefit(Object, Lod, Algorithm)
• Subject to ?s Cost(Object, Lod, Algorithm) ?
  TargetFrameRate
• Equiv. to NP Complete Knapsack problem
• But a simple greedy algorithm can get within 50
  of the optimal solution

32
Predictive Fixed Frame Rate
Benefit(O,L,R) Size(O) Accuracy(O,L,R)
Importance(O) Focus(O) Motion(O)
Hysteresis(O,L,R)

• Benefit contribution to model perception
• Size larger objects contribute more to image
• Accuracy no of verts/polys, shading model, etc.
• Priority account for inherent importance
• Eccentricity distance from center of display
• Velocity ratio of apparent speed to avg polygon
  size
• Hysteresis use state from previous frame
• Assign LOD benefits by hand

33
Predictive Fixed Frame Rate

• Cost assumes a 2-stage pipelined rendering
• per-primitive (coordinate transforms, lighting,
  clipping)
• per-pixel (rasterization, Z-buffering,
  alpha-blend, texturing)
• One of these will be the bottleneck, therefore
  cost is the largest of both
• Required profiling the target platform first to
  instantiate the C1, C2, and C3 constants

34
Predictive Fixed Frame Rate
Hierachical LOD
Discrete LOD
Automatic Hierarchical LOD Optimization in
Animation Mason and Blake (U. Cape Town)
35
Predictive Fixed Frame Rate

• Gobbetti Bouvier (1999) extend the Cost Benefit
  model to Continuous LOD
• Maximise Benefit(W, S(r))
• Subject to Cost(W, S(r)) ? TargetFrameRate
• Defined for a set of multiresolution objects S at
  resolutions r (0..1)
• W viewing configuration (camera lights)

36
Predictive Fixed Frame Rate

• Cost time to render scene of objects at
  resolutions r given viewing params W.
• Includes terms to model
• Graphics initialization phase (e.g., clearing the
  Z-buffer and setting up initial state)
• Sequential rendering of each object
• Finalization phase (e.g., buffer swapping)

cost(W, S(r)) Tinit Tfinal ? Tisetup tmax
. r
37
Summary

• We covered
• Various run-time LOD frameworks
• Lots of ways to modulate LOD not just distance
• Techniques to cover up popping effects
• Fixed frame rate systems
• Later you will learn more about
• view-dependent simplification
• resource management (e.g., out-of-core)

38
End of Presentation
End of Presentation
39
Occlusion LOD Example

• Modeling Hair Using LOD Representations
• Ward, Lin, Lee, Fischer, and Macri (UNC)

40
Image-Based LOD

• Render each LOD off-screen and analyze images to
  decide which parts of the model to simplify

• Guides simplification based upon the visual
  effect of the reduction rather than some
  geometric metric
• Uses a sphere of cameras to capture multiple
  viewpoints
• Deals with surface properties and textures
• Uses RMS error of luminances to compute image
  distances (fast but not perceptually based)
• Not real-time yet (several secs to mins or hours)