Are Points the Ultimate Modeling Primitive

1
Are Points the Ultimate Modeling Primitive?

• Turner Whitted, Microsoft Research
• with
• Marc Levoy, Stanford University
• John Snyder, Microsoft Research

2
Outline

• Part I
• Looking back at an early chain of research
  projects (with Marc Levoy)
• Part II
• Snapshot of present day experiments with a point
  rendering pipeline (with John Snyder)

3
Rendering points an old idea

• E. Catmull, A subdivision algorithm for computer
  display of curved surfaces, PhD dissertation,
  1974
• C. Csuri, et al, Towards an interactive high
  visual complexity animation system, Proc.
  SIGGRAPH 79.
• W. Reeves, Particle systems a technique for
  modeling a class of fuzzy objects, Proc.
  SIGGRAPH 83.

4
2D inverse mapping
Anti-aliasing with
read
accumulate

• Super-sample (uniformly or adaptively)
• Low pass filter read source multiple times
• Re-sample

Store entire super-sampled source image
5
2D forward mapping
Crow, Franklin C., The use of grayscale for
improved raster display of vectors and
characters, Proceedings of SIGGRAPH 78.

• Choose source sample
• Look up pre-computed contribution to destination
  region
• Blend at destination (with visibility if needed)

6
2D forward mapping
Using small pre-computed textures as source
Whitted 83
7
Textures/points
Marc Levoy 25 Sept. 1984

• . . . open up new possibilities for complex
  models.

8
textures are simply a container for points
Marc Levoy 24 Sept. 1984
9
Points as primitives

• Goals
• Ability to manage geometric complexity
• Algorithmic simplicity
• Interactivity

10
PSF-clouds forerunner of splats
Marc Levoy 30 Sept. 1984
11
Rate control

• Primary concern is insuring coverage

The Use of Points as a Display Primitive Marc
Levoy and Turner Whitted Technical Report
85-022, Computer Science Department, University
of North Carolina at Chapel Hill, January, 1985.
12
Progressive display
Levoy 85
13
Surfaces from points

• Forward mapped rendering
• with anti-aliasing
• Complete flexibilty
• Shading
• Texturing
• Displacement mapping

Levoy 85
14
Simplicity one representation
15
Splatted volume rendering

• Primary goal was interactivity
• devise algorithm to exploit parallelism
• Blind faith in the sampling theorem

Lee Westover, Interactive Volume
Rendering,Proceedings of the 1989 Chapel Hill
workshop on Volume visualization, 1989. (first
published use of term splatting)
16
Signal processing for splatting

• Answer question what is the image contribution
  of a single projected point (extend forward
  mapping to volumes)
• Slightly flawed model kernels overlap in z

Westover 89.
17
Typical splatting application
Ozone concentrations in the northeastern US
(1990) Data courtesy of the US Environmental
Protection Agency, National Center for
Atmospheric Research, and Numerical Design Limited
18
Timeline

• 1983 - 2D brushes
• 1984 - incubating 3D PBR (Levoy)
• 1985 - points as primitives (Levoy)
• 1989, 1990 - splatting for volumes (Westover)
• 1991 - hierarchical splatting (Laur Hanrahan)
• 1994 Commanche PC game with range-sorted
  rectangular splats
• 1998 hierarchical coverage (Grossman Dally)
• 2000 - QSplat, Surfels, WarpEngine
• 
  and beyond

19
Jump forward a few years
20
Point-based graphics today

• Modeling points
• Densely acquired
• Adaptively sampled, hierarchical
• Augmented with surface properties
• Rendering splats
• Oriented
• Trimmed
• Interpolated

21
Something we tried

• Rendering primitive
• Ellipsoids

373776 Ellipsoids John Snyder, Kirk Olynyk 2001
22
Goals and approaches

• Fidelity (capture all detail)
• Adaptively sample at the source
• Quality (render it without artifacts)
• Careful filtering at the back end
• Integration
• Use a single primitive representation for 2D, 3D,
  and even imagery
• Simplicity
• Use simplest possible primitive representation

23
Render with points

• Adaptively sample and interpolate at the source
• Retain flexibility at the source
• High frequency content is not uniformly
  distributed over the source domain
• Anisotropy is free
• Accumulate, not overwrite, at the destination
• Every sample contributes, more samples means more
  high frequency content
• Piecewise constant reconstruction is OK
• Current GPUs dont do this

24
Experiments with Points as rendering primitives
John Snyder, Kirk Olynyk, Tom Blank Microsoft
Research
Walk through source samples coherently
Interpolate depending on view transform
Map and write into supersampled image bins
Read supersampled bins, reconstruct, filter,
re-sample, in real time
25
Experimental Testbed
26
Point sampling vs mip-mapping
27
Points vs mip-mapping (cont.)
anisotropic filtering requires no special
processing
28
Point sampling geometry images

• Convenient way to store a collection of points1
  with coherent access (Gu, Gortler, Hoppe 2003)
• Puts samples where detail is
• Easy to interpolate

1see slide no. 8
29
Surface sampling rate

• Is not correct for viewing
• not adaptive in image space

point sample density plot
30
Surface sampling rate

• Is not correct for viewing
• not adaptive in image space

shaded rendering
31
Point pipeline
Very high bandwidth!
struct binPoint int x, y, z0, z1, r, g, b,
a, w binBINDEPTH
32
Additional sample rate control

• Gated mapping
• points that are nearly identical contribute no
  additional detail.
• weights summed and points merged before writing
  into sample memory.

33
Bandwidth sanity check

• Measured rates for one scene1
• Brute force gt 8.39, lt 98.0 GB/s
• Gated mapping gt 5.14, lt 63.4 GB/s
• Additional rate control strategy
• Higher order reconstruction
• Hierarchical source data.

1extrapolated up to 1600x1200 _at_72 Hz refresh
34
Summary

• Points for rendering offer
• simplicity
• flexibility
• and all the quality one can use
• for all the bandwidth one can afford

35
Acknowledgements

• Adaptive sampling
• Kirk Olynyk, MSR
• Tom Blank, MSR
• Volume splatting
• Lee Westover, NVIDIA

36
(No Transcript)