Fourier Depth of Field

Siggraph 2009

Cyril Soler, Kartic Subr,

Frédo Durand,
Nicolas Holzschuch &
François Sillion
Rendered

FinalDOF renderer
Depth of field is expensive

http://developer.amd.com/media/gpu_assets/Scheuermann_DepthOfField.pdf
 Oversampling
● Typically we'll oversample a pixel, in a “circle of
confusion”, near it's location.
In focus Out of focus
 Algorithm for estimating defocus

P = {uniformly distributed image samples}

NA // number of aperture samples
for each pixel x in P
L ← SampleLens(NA)
for each sample y in L
Sum ← Sum + EstimatedRadiance(x, y)
Image (x) = Sum / NA
We're doing more work to display less
data?
 Less data at lower frequencies

● Nyquist Limit
● “If a function x(t) contains no frequencies higher than B hertz,
it is completely determined by giving its ordinates at a series of
points spaced 1/(2B) seconds apart.”

http://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem
 Summary
Bandwidth Estimation

Sample generation
over image and lens

Estimate radiance
rays through image and lens samples

Reconstruct image
from scattered radiance estimates

http://artis.inrialpes.fr/~Kartic.Subr/Files/Pres/FourierDOFPres.ppt
 Radiance function

● Given a point
● The radiance function is
the intensity of light in all
directions
Example of light path
Local light field propagation

Central ray

[Durand05]
● We're not talking about the spectrum of light
transmitted.
● We're talking about measuring the local changes in
the radiance function...
Fourier Transform → frequency space
 Emission

● Spacial frequencies
● Angular frequencies
 Transport

● Angular shear in frequency space

 Occluders

● Convolution (in frequency space) with frequency of

occluder. (Product in ray-space).
Durand 2005 goes into (much) more depth

Durand 2005, A Frequency Analysis of Light Transport

 Old Algorithm for estimating defocus

P = {uniformly distributed image samples}

NA // number of aperture samples
for each pixel x in P
L ← SampleLens(NA)
for each sample y in L
Sum ← Sum + EstimatedRadiance(x, y)
Image (x) = Sum / NA
 Our adaptive sampling

(P, A) ← BandwidthEstimation()

P = {uniformly
{bandwidthdistributed
dependent image
image samples}
samples} – 1 to 10% final samples
N
AA= {aperture variance estimate}
for each pixel x in P
LA←
N SampleLens(N
proportional A)
to A(x)
for each sample y in L
Sum ← Sum + EstimatedRadiance(x, y)
Image (x) = Sum / NA

Reconstruct (Image, P)
 Image space sampling Density

● W,H image dimensions

● fh, fv field of view
th
● Max energy - from angular bandwidth (use 98
percentile to avoid outliers)

● (From Nyquist limit)

 Generating samples from

● Fast Hierarchical Importance Sampling with Blue

Noise Properties - Sigg04
● Penrose tilings
 Reconstruction from sparse samples
● “weighted average of a constant number of
neighboring samples”
● “adaptively varying the radius of contribution of
each pixel”
● “In practice, we use a Gaussian weighting term with
a variance that is proportional to the square root of
the local density”
Results: Computation time (seconds)

Bandwidth 90 45 60
estimation

Raytracing 4500 3150 7401

Image
reconstruction 10 3 8
Summary of phenomena
Reference
 Defocus
 Reflectance
 Occlusion

Image-space bandwidth Aperture variance

Reflection
 Comments?
● Final number of points is the integral of the sample
density, rather than a given value
● Only ~20 times faster?
● Does it only process the spectra after the last
bounce? (or is it gathered before?)
● Uses a conservative bandwidth estimate – lots of
room to tighten bounds