An Efficient Progressive Refinement Strategy for Hierarchical Radiosity

1
An Efficient Progressive Refinement Strategy for
Hierarchical Radiosity

• Nicolas Holzschuch, François Sillion and George
  Drettakis

iMAGIS/IMAG, Grenoble France
A joint research project of IMAG and INRIA
2
Motivation

• Hierarchical radiosity is a significant step in
  radiosity algorithms
• creates links between patches and refines them
• linear in the number of elements created
• Proceeds top-down
• First establish links between input surfaces
• Then refine these links where needed

3
Motivation (2)

• Initial linking step quadratic in the number of
  polygons
• Many top-level links will never carry significant
  energy
• Subdivision is often too high

4
Proposed improvements

• Delaying initial linking of input surfaces
• Reducing the number of links

5
Our test scenes
6
Initial Linking

• Proportion of top-level links with BF lt ?

7
Subdivision
8
Previous work

• Hierarchical radiosity (Hanrahan, 90- 91)
• Link refinement based on radiance and form-factor
• proceed from top to bottom
• multigridding
• Importance-driven hierarchical radiosity (Smits,
  Arvo Salesin, 92)
• Links refined using importance and influence on
  the final image

9
Previous work (2)

• Hierarchical radiosity and discontinuity meshing
  (Lischinski, Tampieri Greenberg 93)
• First refine patches using a discontinuity mesh,
  then re-refine using radiosity and form-factor
• Structured sampling (Drettakis Fiume 93)
• Adapt mesh to illumination structure

10
Delaying initial linking

• Delay top-level linking between input surfaces
  until strictly necessary
• First iteration results achieved more rapidly
• Spread computation over several iterations
• Avoid part of initial linking computation gain
  on total computation time

11
Classification of pairs

• Initially, all pairs of polygons are un-classified

Un-classified

• Important pairs progressively become
  classified.

• We compute visibility tests only for
  classified pairs.

Classified
Un-classified
Visible
Partial
Invisible
12
Linking algorithm

• First record all polygon pairs as un-classified.
• As soon as a pair qualifies for linking (BF gt
  ?), compute visibility and link it accordingly.
• The remainder of the algorithm is not modified.

13
Energy Balance

• Partially linked polygons do not emit all their
  energy
• Un-radiated energy affects the energy balance
• quantify the importance of this lost energy
• compare it with the overall precision of the
  algorithm.
• Unit sum of form-factors allows estimation of
  lost energy

14
Energy Balance
15
Reducing the number of links

• Perform an a posteriori test to determine whether
  refinement was required

16
Reducing the number of links

• We cancel the refinement if the following four
  expressions are true

17
Results first iteration
18
Results ten iterations
19
Subdivision
20
Conclusion

• Delaying top-level linking between input surfaces
• storage costs are reduced
• we obtain first results earlier
• still quadratic in the number of input surfaces
• Reducing the number of links
• improved subdivision criterion
• limits un-necessary subdivisions
• Future work
• Simplify already subdivided meshes