Acceleration of Monte Carlo Path Tracing in General Environments GGG speech

1
Acceleration of Monte Carlo PathTracing in
General EnvironmentsGGG speech
Frederic Pérez1Ignacio Martín1François X.
Sillion2Xavier Pueyo1
2
Outline

• Importance Sampling
• Previous work
• Link Probabilities
• Results
• Conclusions
• Future Work
• Current Work
• The new SIR

3
Importance Sampling

• Variance reduction technique
• Use the best samples to evaluate an integral
  Probability Density Function (PDF) prop. to
  kernel
• Reflectance equation
• Importance sampling
• Use of approximate for PDF ?
  Usually by means of a 1st pass

4
Previous Work

• Chen et al. 91
• Progressive Multi-Pass
• Jensen 95
• Importance Driven Path Tracing using the Photon
  Map
• Szirmay-Kalos et al. 98
• Importance Driven Quasi-Random walk
• Stürzlinger 96, Ureña Torres 97
• Final Gathering

5
Shortcomings

• Fixed subdivision of hemisphere
• Cannot adapt to irradiance changes
• Missing samples
• Sources does not fit solid angles
• Fixed subdivision of object space first pass

6
Shortcomings

• Smallprojections

? Wasting samples
? Bad PDF
7
Link Probabilities Overview

• 1st Pass Radiance Clustering
• Produces Line-Space hierarchy Links
• Coarse solution not for rendering
• 2nd Pass Monte Carlo Path Tracing
• Improved Importance Sampling with Link
  Probabilities
• Use links arriving at leaves as a representation
  of the irradiance

8
Link Probabilities Basic Algorithm

• Importance Sampling in a random walk step

for each bounce at point x
x
9
Link Probabilities PDF per leaf?

• Using a PDF per leaf ? Artifacts for 1st
  bounce

? Solution Recompute PDF per pixel
10
Link Probabilities

• Link overlap

S2
S1
y
? Check if links sender is hit
11
Link Probabilities

• PDF accuracy and visibility

S1
l1
S2
l2
R
Finite elements pass ? two links arriving at
leaf R
12
Link Probabilities

• PDF accuracy and visibility

Solution Adaptive PDFs
13
Results

• Simple test scene

14
Results

• Room with indirect illumination

15
Results
16 samples
128 samples
64 samples
NEE
806s
1613s
201s
LP
1273s
2434s
409s
16
Conclusions

• Enhancement of Path Tracing by Importance
  Sampling based on the estimated irradiances
  computed with a Radiance Clustering 1st pass
• Quality of the images doesnt strongly depend on
  the lighting conditions

17
Future Work

• Extension for isotropic participating media
• Account for links arriving to leaf clusters with
  media and sample directions as for surfaces
• Addition of glossy surfaces and anisotropic
  participating media
• Account for BRDF/phase function
• Use of Link Probabilities in BPT
• Eye paths as presented
• Light paths similarly using importance
• Comparison to Metropolis Light transport

18
Acknowledgments ( the paper)

• SIMULGEN ESPRIT project 35772
• http//iiia.udg.es/Simulgen
• Generalitat de Catalunyas AIRE
• CICYTs TIC98-0586-C03-02
• The paper
• http//ima.udg.es/frederic/publications.html

19
Current Work

• New code Getting rid of BRIGHT SIR
• Finishing the implementation over the old code
  estimated to be harder than re-starting
• The old software was hard to maintain
• Strange bugs...
• Avoid code duplication, scene duplication ? work
  with a more compact and efficient system? new
  SIR

20
Current Work New Two Pass Method

• New code using the new SIR
• New 1st pass particle tracing based
• Same requirements as previous 1st pass
• Obtain quickly a coarse solution to guide a 2nd
  pass
• Maintaining the 2nd pass using the Link
  Probabilities
• Reimplementing and finishing it

21
Current Work New 1st Pass

• Extended Hierarchical Monte Carlo Radiosity ?
  Link hierarchy
• Bekaert et al. Hierarchical Monte Carlo
  Radiosity. In Rendering Techniques 98
• Adaptively refined mesh and (implicit) link
  structure ? but storing links for the 2nd pass
• Extended for anisotropic participating media and
  glossy surfaces ? by means of Illumination Samples

22
Current Work Snapshot

• Implementation
• As in Bekaerts (RenderPark)
• Breadth-first MCR ? Push-pull only after drawing
  all samples
• Local Monte Carlo
• Currently almost finished for diffuse surfaces
• To do
• Finish it for diffuse surfaces
• Implement participating media
• Reimplement second pass

23
Current Work First Results
24
The new SIR

• http//ima.udg.es/sir/
• Using GNU configure/build
• Self documented using doxygen
• Reference guide style
• Dependency graphs
• tkcvs web access commit ? news!
• Based on triangles (so far)
• GTK as GUI
• GTS to triangulate (instead of CDT)