Kevin Steele
CS 6650
Programming Assignment #1: Monte Carlo
Integration, Density Estimation, Metropolis
Part 0: Analytic
Integration of function
All images were computed using 32-bit floats with no gamma correction. I originally generated them using doubles, and those images were minutely better. However, the computation speed gain using floats more than made up for the slightly poorer image quality. |
Animations These animations show how the function behaves under a
phase shift. Each frame shows the function phase shifted by 2pi/90 from the
previous frame (90 frames in each animation). The stratified MCI solution
exhibits animated aliasing, and the pseudo-bspline MCI solution substitutes
the aliasing for animated noise. The animations are Quicktime format, and are best viewed
as looped movies. MCI Stratified
100 Samples per Pixel MCI BSpline 100
Samples per Pixel MCI BSpline
2000 Samples per Pixel |
|
Part 1: MC
Integration, 16 random samples
|
Difference Image from Analytic Solution (intensity
scale factor: 2.2)
|
Difference Image from Analytic Solution (intensity scale factor: 10)
|
Part 2: MC Integration, 16 stratified samples
|
Difference Image (intensity scale factor: 6.71)
|
Difference Image (intensity scale factor: 10)
|
Part 3: MC Integration, 100 random samples
|
Difference Image (intensity scale factor: 5.1)
|
Difference Image (intensity scale factor: 10)
|
Part 4: MC Integration, 100 stratified samples
|
Difference Image (intensity scale factor: 31.88)
|
Difference Image (intensity scale factor: 10)
|
Part 5: MC Inegration, 100 random samples, bspline filter
|
|
|
|
Part 6: Density Estimation, 100 random samples per pixel on average
The scaling for image intensity (power) was done using the analytic integral of the function, rather than an MCI approximation. This improves the accuracy of this particular image, but is not practical in the general case. |
Difference Image (intensity scale factor: 8.5)
|
Difference Image (intensity scale factor: 10)
|
|
Part 7: Density Estimation, 100 stratified samples per pixel on average
|
Difference Image (intensity scale factor: 42.5)
|
Difference Image (intensity scale factor: 10)
|
Part 8: Metropolis, 100 random samples per pixel on average
The area of mutation for each sample in this image was the entire image. Each sample point had an equal probability to transport to any position in the image. This allowed the samples to converge much faster, though such a large proportional area may not be practical or possible in the general case. |
Difference Image (intensity scale factor: 9.11)
|
Difference Image (intensity scale factor: 10)
|