1 / 33

Sampling

Sampling. Attila Gyulassy Image Synthesis. Overview. Overview Problem Statement Random Number Generators Quasi-Random Number Generation Uniform sampling of Disks, Triangles, Spheres Stratified Sampling Importance Sampling of General Functions. Problem Statement. What is sampling?

liam
Download Presentation

Sampling

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Sampling Attila Gyulassy Image Synthesis

  2. Overview • Overview • Problem Statement • Random Number Generators • Quasi-Random Number Generation • Uniform sampling of Disks, Triangles, Spheres • Stratified Sampling • Importance Sampling of General Functions

  3. Problem Statement • What is sampling? • Want to Take a function f and recreate it using only certain values • e.g. data points used in interpolation • where to pick those points? • Sometimes don’t know f but can evaluate it • would like to choose data points used to reconstruct function in an optimal way

  4. Problem Statement (ctd) • Monte Carlo Integration • 2 ways to improve • improve estimation method • carefully selecting samples****** • Use filtering to recreate original function • covered next time • important to know necessary sampling frequency

  5. Example

  6. Overview of Sampling • Over some domain • Sometimes parametrizable • Some sample density • Random / Regular

  7. Random Numbers • Would like to get uniformly distributed random numbers over a range [a,b] • Problems • large open spaces • slow convergence • nondeterministic

  8. RNG methods • Non-linear additive feedback • Linear congruence methods • Mersenne Twister algorithm • many more...

  9. Quasi-Monte Carlo • Use deterministic roughly uniform aperiodic distribution through domain • I.e. pseudo-random numbers • Want low discrepancy • small = evenly distributed • large = clustering • causes clumping and sparse regions • Want high speed

  10. Quasi-Random Generators

  11. Halton Sequence • N-dimensional points xi xm = (2(m), 3(m),…, PN-1(m), PN(m)) PI = ith prime number (2,3,5,7,…) r(m) is the radical-inverse function of m to the base r. The value is obtained by writing m in base r and then reflecting the digits around the decimal point. 2610 = 110102 reflecting 0.010112 = 11/2710

  12. Halton Sequence • N-dimensional points xi xm = (2(m), 3(m),…, PN-1(m), PN(m)) PI = ith prime number (2,3,5,7,…) m = a0r0 + a0r1 + a0r2 + a0r3 + ... r(m) = a0r-1 + a0r-2 + a0r-3 + a0r-4 + ...

  13. Halton Sequence • Starting at (1,1,…,1) better than starting at (0,0,…,0) 1 = 1.0 => 0.1 = 1/2 2 = 10.0 => 0.01 = 1/4 3 = 11.0 => 0.11 = 3/4 4 = 100.0 => 0.001 = 1/8 5 = 101.0 => 0.101 = 5/8 6 = 110.0 => 0.011 = 3/8 7 = 111.0 => 0.111 = 7/8 Notice even distribution

  14. Hammersley Sequence • Similar to Halton xm = (m/N,2(m), 3(m),…, PN-1(m)) PI = ith prime number (2,3,5,7,…) m = a0r0 + a0r1 + a0r2 + a0r3 + ... r(m) = a0r-1 + a0r-2 + a0r-3 + a0r-4 + ... Where N is number of total samples

  15. Hammersley Sequence

  16. Hammersley Sequence

  17. Poisson Random Numbers • Generate random numbers according to the Poisson distribution function This turns out to be the same as just “throwing darts”**

  18. Result of RNGs • Basically, now we have random numbers in [0,1] • what do we do with these? • How does this relate to sampling?

  19. Uniform Sampling of a Disk • Want Subdivision into equal area regions

  20. Uniform Sampling Over a Sphere • demo

  21. Uniform Sampling - Disk vs Sphere • Sampling of disk and projecting onto hemisphere = sampling on 1/2 of sphere

  22. Uniform Sampling of Triangles • Compute probability density function for triangles

  23. Uniform Sampling of Triangles • The u and v are not independent

  24. Stratified Sampling • Alternative to uniform • break domain into strata • fills in gaps faster

  25. Importance Sampling • Basic Idea • sample at important locations to decrease variance

  26. Importance Sampling ctd. • As seen last time, use a probability density function f to pick samples • properties

  27. Importance Sampling ctd. • Then, our approximation becomes (here g(x) is prob. Dens. Funciton, not f(x))

  28. Importance Sampling ctd. • How do we pick f? • want to minimize variance • where G is integral of original function g(x) • … after much math we get • which is great!! Except, G is what we are trying to find G2 f(x) = |g(x)| / G

  29. Importance Sampling ctd. • If we don’t know G, how can we pick f • If we apply a filter to g, so integral is of form Then if the filter is clamped [0,1] the filter itself becomes a reasonable estimate for f Problems with this method?

  30. Importance Sampling ctd. • Remember f(x) = |g(x)| / G gives least variance • motivation for adaptive sampling • build f from first few samples

  31. Conclusion • Multiple ways to generate “random” numbers • have to pick best method for each application • Many sampling techniques, with pros and cons • uniform • stratified • importance • adaptive

  32. References • http://www.math.iastate.edu/reu/2001/voronoi/halton_sequence.html • http://www.cse.cuhk.edu.hk/~ttwong/papers/udpoint/udpoint.pdf • http://www.fz-juelich.de/nic-series/volume10/janke1.pdf • http://graphics.stanford.edu/courses/cs348b-00/lectures/lecture13/montecarlo.1.pdf • Principles of Digital Image Synthesis, Glassner

More Related