Anton S. Kaplanyan

Karlsruhe Institute of Technology, Germany

Adaptive Progressive Photon Mapping

Adaptive PPM Original PPM

2

Progressive Photon Mapping in Essence

Pixel estimate using eye and light subpaths

Generate full path by joining subpathsEye subpathimportance

Photonradiance

𝛾𝑖+1

Kernel-regularized connection of subpaths

𝑊 𝑁 𝛾𝑖

3

Reformulation of Photon Mapping

PPM = recursive (online) estimator [Yamato71]

Rearrange the sum to see that

Kernelestimation

Pathcontribution

4

Radius Shrinkage

Shrink radius (bandwidth) for th photon map

User-defined parameters and

Problem:

Optimal value of and are unknown

Usually globally constant / k-NN defined

5

Box scene(reference)

User Parameters Example

6

User Parameters Example

Larger 𝛼

Larg

er

𝒓 𝒓

𝒓 𝒓Differenceimage

7

Radius Shrinkage Parameters

𝑟0

𝑟0

𝑟0

…𝛼

Optimal Convergence of Progressive Photon Mapping

10

Optimal Asymptotic Convergence Rate

𝑟0

𝑟0

𝑟0

…𝛼

11

𝛼  𝐨𝐩𝐭

Optimal Convergence Rate

Variance and bias depend on [KZ11]

Optimal rate is with Asymptotic convergence

Unbiased Monte Carlo is faster:

12

Convergence Rate of Kernel Estimation

Convergence rate for dimensions

Suffers from curse of dimensionality

Adding a dimension reduces the rate!Shutter time kernel estimation – not recommended

Wavelength kernel estimation – not recommended

Volumetric photon mapping

Adaptive Bandwidth Selection

14

Optimal Asymptotic Convergence Rate

𝑟0

𝑟0

𝑟0

…𝛼

15

Adaptive Bandwidth Selection

might not yield minimal

Minimize with respect to Achieve variance ↔ bias tradeoff

Select optimal using past samples

16

Estimation ErrorMean Squared Error [Hachisuka et al. 2010]

17

Estimation Error

Variance is two-fold: Path measurement contribution

Kernel estimation

18

Estimation Error

Measurement variance is higher

19

Estimation Error

So, MSE has noise (path variance) and bias

Variance Bias

20

Adaptive Bandwidth Selection

Both variance and bias depend on

Where is a pixel Laplacian

Laplacian is unknown

21

Estimating Pixel Laplacian

consists of Laplacians at all shading pointsWeighted per-vertex Laplacians

22

𝑥 𝑥+h𝑢𝑥− h𝑢

∆ 𝐿𝑢=𝐿𝑥+ h𝑢 +𝐿𝑥− h𝑢 −2𝐿𝑥

h2

Estimating Per-Vertex Laplacian

Estimate per-vertex Laplacian at a point

Recursive finite differences [Ngen11]

Yet another recursive estimator

Another shrinking bandwidth

Robust estimation on discontinuities

23

Adaptive Bandwidth Selection

Estimate all unknownsPath variance

Pixel Laplacian

Minimize MSE as MSE(r)

Lower initial error Keeps noise-bias balance

Data-driven bandwidth selector

24

Progressive Photon Mapping Adaptive PPM

20 seconds!

Results

25

Progressive Photon Mapping Adaptive PPM

3 seconds!

Results

26

Conclusion

Optimal asymptotic convergence rateAsymptotically slower than unbiased methods

Not always optimal in finite time

Adaptive bandwidth selectionBased on previous samples

Balances variance-bias

Speeds up convergence

Attractive for interactive preview

Thank you for your attention.