Photon-driven Irradiance Cache

1
Photon-driven Irradiance Cache

• J. Brouillat P. Gautron K. Bouatouch
• INRIA Rennes University of Rennes1

2
Motivations
? Photon-Driven Irradiance Cache

• Global Illumination in scenes with complex light
  interactions
• Indirect diffuse component is very important for
  the overall aspect of final rendering
• Multiples bounces, any lighting, large scenes
  (architectural design, walkthrough)
• Interactive or view-independent
• Fast preview of good quality

3
Previous work
? Photon-Driven Irradiance Cache

• Mesh-based
• Radiosity Lightmaps Goral et al. 84 compute
  a solution for the whole scene at once, can be
  interactively displayed, Discontinuity meshing
  Lischinski 92
• Point-based
• Irradiance Cache Ward et al. 88 Irradiance is
  sparsely sampled over surfaces
• Radiance cache splatting Gautron et al. 05
  interactive display of an irradiance cache
• Photon mapping Jensen 96 divides light flux
  into photons
• Complex lighting, very large scenes Christensen
  et al. 04
• Final gathering combined with irradiance cache

4
Previous work
? Photon-Driven Irradiance Cache

• Point-based (continued)
• Instant radiosity Keller 97 approximate
  indirect lighting using several virtual
  point-light sources (VPLs)
• Many-light sources problem Lightcuts Walter et
  al. 05, Matrix Row-Column Sampling Hasan et al.
  07
• Incremental Instant Radiosity Laine et al. 07
  interactive, limited to one bounce interactions
• Metropolis Instant Radiosity Segovia et al. 07
  reduce the number of VPLs used, view dependent
• Direct-to-indirect transfer Hasan et al. 06
  samples surface with points, precompute
  contribution of these points to final image
• Dynamic lights, fixed scene/viewpoint
• High-quality interactive relighting , large
  precomputation time (hours)
• Meshless Hierarchical basis Lehtinen et al. 08
• Dynamic lights and viewpoint, long precomputation
  time.

5
Our approach
? Photon-Driven Irradiance Cache ? Our algorithm

• Point-based short precomputation time view
  independency
• How combination of photon maps and irradiance
  cache
• Photon map
• Elegant method, widely used
• Any kind of light sources, multiple bounces,
  highly dense environments, large scenes
• Irradiance caching
• Interactive display, compact representation, no
  surface parameterization
• Compute a view-independent irradiance cache from
  a photon map
• Advantages
• Fast view-independent preview of photon map
  simulations
• Uses GPUs (splatting)
• Easy to implement

6
More precisely
? Photon-Driven Irradiance Cache ? Our algorithm

• Photon Mapping
• () View-independent lighting simulation
  (scattering photons)
• (-) Rendering is either noisy or costly (final
  gathering), and view-dependent
• Irradiance Cache
• () Interactive rendering of the cache for static
  scene, high quality
• (-) View dependent new records have to be
  created when walking through unexplored parts of
  the scene. Records cover the whole scene many
  viewpoints
• Our approach
• We first compute a coarse irradiance cache from a
  photon map
• Covers the whole scene view independent
• No additional rays to be cast
• Interactive fast preview of the photon map
  solution
• This coarse cache can be refined using final
  gathering
• Reduce the number of rays cast, using visibility
  information from the photon paths

7
Irradiance cache data structure
? Photon-Driven Irradiance Cache ? Irradiance
cache

• Irradiance cache set of records
• A record
• Position (vector)
• Harmonic mean distance, DH, from record to other
  objects of the scene (scalar)
• Irradiance value (vector)
• Irradiance gradient (vector)
• Deduce these values from a photon map

8
Algorithm overview
? Photon-Driven Irradiance Cache

• 1st pass
• Building a classical photon map
• 2nd pass, from the photon map
• Deduce position of irradiance records and
  harmonic mean distance
• Compute coarse irradiance values for each record
• Coarse irradiance cache can be used for
  interactive (pre-)vizualisation
• 3rd pass
• Refine irradiance values associated with each
  record using final gathering, compute gradients
• Can be done in parallel with previzualisation
• Refine first the visible records

9
First Pass Photon map
? Photon-Driven Irradiance Cache ? First pass

• Scattering the photon across the scene
• We fix a minimal number of bounces, then use
  Russian Roulette

direct lighting
photon map 250k photons
10
Second Pass Irradiance record positions
? Photon-Driven Irradiance Cache ? Second pass

• We want the cache to cover most part of the scene
  (avoid cache misses)
• Records need to be created at each point where
  incident irradiance is not null
• Photon represents incident light
• We use photon positions as candidates for
  records
• As we process each photon position, we have to
  decide whether to create a new record or not.

11
Irradiance record positions
? Photon-Driven Irradiance Cache ? Second pass

• We start with an empty cache
• The record A is created at photon a position
• This record have a zone of influence
• Inside this zone, the irradiance stored in the
  record A will be used for interpolation.
• Its data structure is augmented with a new field
  giving a list of all the photons within its zone
  of influence (kd-tree query)

1
12
Irradiance record positions
? Photon-Driven Irradiance Cache ? Second pass

• Photon b lies inside the zone of influence of a
  previous created record
• No need to create a new record at b position
• How do we know it ?
• In classical irradiance caching, the records are
  stored in an octree
• Octree query provide the neighborhood of b
• Millions of photons too costly
• Our approach avoids any octree query
• How new field in the photon data structure
• wc cumulative contribution of neighbour records,
  updated at the creation of each record

2
13
Irradiance record positions
? Photon-Driven Irradiance Cache ? Second pass

• Decision on creating a new record or not relies
  on wc
• For photon c, wc is less than a given threshold
  not enough records around it
• Then a new record C is created at photon position
• wc field of all the photons lying in the zone of
  influence of C is updated

3
14
Harmonic mean distance estimation
? Photon-Driven Irradiance Cache ? Second pass

• Irradiance records need to carry DH, the harmonic
  mean distance to objects of the scene
• DH is used to determine the size of the zone of
  influence of the record
• Bad estimation leads to artifacts during rendering

15
Harmonic mean distance estimation
? Photon-Driven Irradiance Cache ? Second pass
1

• Usually done by casting rays towards the scene
• We want to deduce DH from the photon maps without
  additional ray casting

16
Harmonic mean distance estimation
? Photon-Driven Irradiance Cache ? Second pass
2

• Photons store information about the distance
  between subsequent hit points
• Photons close to a record K give us information
  about visibility around K

17
Harmonic mean distance estimation
? Photon-Driven Irradiance Cache ? Second pass

• By reprojection, we estimate visibility around K
  without casting rays
• This may miss some occluders, and introduce
  errors
• We use only photons close to K
• inside the zone of influence
• Neighbour clamping Krivanek et al. 06
• Reduction of DH based on neighbour records
  (borders,)

3
18
Estimating irradiance from photon map
? Photon-Driven Irradiance Cache ? Second pass

• We estimate incoming irradiance at record
  positions with density estimation
• Usually, search for the n-nearest neighbours
  photons in a kd-tree
• Automatic adaptation with actual density of
  photon at P
• n can be seen as a trade off between noise and
  blur in the final image

n ?
P
Irradiance at P sum of flux carried by n
photons divided by area A
19
Estimating irradiance from photon map
? Photon-Driven Irradiance Cache ? Second pass

• Records with large zone of influence may have
  high visual importance on the final rendered
  images
• Blur is visually more pleasant than noise, for
  previews
• We want n vary depending on the size of the zone
  of influence
• To this end, we use a range-search approach
• Search for the nearest photon lying in the zone
  of influence

r
P
r a.DH
20
Coarse irradiance cache
? Photon-Driven Irradiance Cache ? Second pass ?
Results

• 250 k photons, 4 bounces
• PM 0.8 sec
• Cache Generation 1,20 sec
• Irradiance values
• Still some noise
• no gradients
• Interactive visualization
• Fast preview of global illumination solution from
  a photon map

21
Third Pass Refining the cache
? Photon-Driven Irradiance Cache ? Third pass

• We would like to refine the coarse cache
• More accurate estimation of irradiance for each
  record
• Compute irradiance gradients
• We use gradients from Krivanek et al. 05
• dividing the hemisphere into cells and by casting
  rays (final gathering)
• We want to save rays thanks to information
  contained in the photon map

22
Refining the cache
? Photon-Driven Irradiance Cache ? Third pass

• Like DH estimation, reuse visibility information
  from photons
• Photons are linked along a path
• Instantly retrieve the photon corresponding to
  last hit point
• Assign each last-hit-point-photon to a cell of
  the hemisphere.

1
23
Refining the cache
? Photon-Driven Irradiance Cache ? Third pass

• Photons have been assigned a cumulative
  irradiance value at the end of the 2nd pass
• If a cell has been assigned a photon
• Fast estimation of incident radiance
• No ray cast

2
24
Refining the cache
? Photon-Driven Irradiance Cache ? Third pass

• Cast rays through non already assigned cells
• Incident radiance along the ray
• search for the nearest photon at intersection
  point
• Compute irradiance and gradients
• Possible to refine DH
• Consequence creation of new records

3
25
Results
? Photon-Driven Irradiance Cache ? Third pass ?
Results

• Cornell Box (32 poly.)
• 250 k photons, 4 bounces
• Total time 6.7 sec
• PM 0.8 sec
• Cache Generation 1,20 sec
• Refinement 4.6 sec
• 39 saved rays (compared to brute force final
  gathering per record)
• visualization 40 fps

Results gathered on a Pentium4 3.8 Ghz computer,
2GB RAM, NVidia GeForce 7800 GTX 512MB
26
Results
? Photon-Driven Irradiance Cache ? Third pass ?
Results

• Sibenik Cathedral (80k poly.)
• 4 M photons, 9 bounces
• Total time 204 sec
• PM 30 sec
• Cache Generation 12 sec
• Refinement 162 sec
• 46 saved rays
• Visualization 5 fps

Results gathered on a Pentium4 3.8 Ghz computer,
2GB RAM, NVidia GeForce 7800 GTX 512MB
27
Fast preview comparison
? Photon-Driven Irradiance Cache ? Comparisons

• End of 2nd pass
• View-independent 28.3 11.9 sec
• View-dependent 28.3 1.5 sec

• End of 3rd pass
• View-independent 28.3 174 sec
• View-dependent 28.3 26 sec

28
Refined cache comparison
? Photon-Driven Irradiance Cache ? Comparisons

• Stochastic ray tracing
• 900 rays / pixel 1300 sec
• View dependent

• Photon-driven Irradiance Cache
• 7 sec, view independent
• Interactive visualization

29
Refined cache comparison
? Photon-Driven Irradiance Cache ? Comparisons

• Stochastic ray tracing
• 900 rays / pixel 3hour
• View dependent

• Photon-driven Irradiance Cache
• 204 sec, view independent
• Interactive visualization

30
Conclusion
? Photon-Driven Irradiance Cache

• Combination of photon maps and irradiance cache
• Compute an irradiance cache from a photon map
• Fast view-independent preview of photon map
  simulations
• High-quality after refinement
• Easy to implement
• Future works
• Parallelization of irradiance cache creation
• Glossy surfaces, using a radiance cache Krivanek
  et al. 05
• Out-of-core aspects (very large scenes)

31
Questions?
? Photon-Driven Irradiance Cache
Fin
32
Our approach
? Photon-Driven Irradiance Cache ? Our algorithm

• Point-based short precomputation time view
  independency
• How combination of photon maps and irradiance
  cache
• Photon map
• Elegant method, widely used
• Any kind of light sources, multiple bounces,
  highly dense environments, large scenes
• Irradiance caching
• Interactive display, compact representation, no
  surface parameterization
• Compute a view-independent irradiance cache from
  a photon map
• Advantages
• Fast view-independent preview of photon map
  simulations
• Uses GPUs (splatting)
• Easy to implement

33
More precisely
? Photon-Driven Irradiance Cache ? Our algorithm

• Photon Mapping
• () View-independent lighting simulation
  (scattering photons)
• (-) Rendering is either noisy or costly (final
  gathering), and view-dependent
• Irradiance Cache
• () Interactive rendering of the cache for static
  scene, high quality
• (-) View dependent new records have to be
  created when walking through unexplored parts of
  the scene. Records cover the whole scene many
  viewpoints
• Our approach
• We first compute a coarse irradiance cache from a
  photon map
• Covers the whole scene view independent
• No additional rays to be cast
• Interactive fast preview of the photon map
  solution
• This coarse cache can be refined using final
  gathering
• Reduce the number of rays cast, using visibility
  information from the photon paths

34
Estimating irradiance from photon map
? Photon-Driven Irradiance Cache ? Second pass

• Boundary bias error in estimation of area A
• Over estimation around geometry boundaries
• Under estimation on curved surfaces
• Light leaks
• Since were doing much less density estimations,
  we can use techniques to compute a more accurate
  approximation of A
• Convex hull Jensen01, OctoBoxes, Geometry
  feelers Tobler et al. 06,

• over-estimation of A

convex hull
35
Direct visualization of photon map
? Photon-Driven Irradiance Cache ? Second pass ?
Comparison

• ? Our method
• ? 250k photons
• ? 2 sec
• ? View independent
• ? Interactive display
• ? 12 DE / record

• ? Photon Mapping
• ? 250k photons
• ? n 32
• ? 4 sec
• ? View dependent
• ? 1 DE / pixel

• ? Photon Mapping
• ? 50k photons
• ? n 128
• ? 4 sec
• ? View dependent
• ? 1 DE / pixel

• ? Christensen 99 caching
• ? 50k photons
• ? n 128
• ? 1,25 2.5 sec
• ? View dependent
• ? 1 DE / photon