BiScale Radiance Transfer

1
Bi-Scale Radiance Transfer

• Peter-Pike Sloan
• Xinguo Liu
• Heung-Yeung Shum
• John Snyder
• Microsoft

2
Radiance Transfer Techniques

• Bidirectional Texture Functions (BTFs)
• local transport
• directional light basis
• fine spatial scale (meso-scale)

3
Radiance Transfer Techniques

• Precomputed Radiance Transfer (PRT)
• global transport
• low frequency light basis
• coarse spatial scale (macro-scale)

4
Radiance Transfer Techniques

• Bi-Scale Radiance Transfer
• local global transport
• low frequency light basis
• two spatial scales

5
Our Goal

• add meso-scale texture to PRT
• extend BTF to handle
• global transport
• soft (area) lighting
• dynamic light/view in real-time

6
Bidirectional Texture Functions (BTFs)
b(xp,vp,s)
Dana99

• 6D scalar function
• xp texture index at point p
• s light direction
• vp view direction at p
• get exit radiance e by integrating over
  directions s

7
Masking (View Dependence)
xp
different viewing angles image different surface
points BTF models parallax/occlusion
8
Precomputed Radiance Transfer (PRT)
l
p
object
9
Precomputed Radiance Transfer (PRT)
l
l'
p
10
Precomputed Radiance Transfer (PRT)
e(p,vp)
Kautz02
l'
vp
p
11
Related Interactive Rendering Techniques
12
Problems with BTFs and PRT

• BTF hard to add global transport and soft
  lighting
• already unwieldy 6D function
• expensive integration for area lighting
• PRT hard to add meso-scale effects
• lengthy simulation meso-scale many surface
  points
• huge storage PRT matrix (625D) at each surface
  point
• cant render in real-time at meso-scale
  resolutions

13
Radiance Transfer Texture (RTT)

• 4D vector-valued function (25-vector)
• computed (BTF ? RTT) by integration

14
RTT Demo
15
Radiance Transfer Texture (RTT)

• just a spatially-varying BRDF vector from PRT

• tabulated over small patch
• provides inexpensive area lighting
• but still no global transport

16
Bi-scale Radiance Transfer
17
Bi-Scale Radiance Transfer

• l vector source radiance spherical function
• Mp 25x25 transfer matrix at point p (source ?
  transferred incident)
• q(xp) ID map (2D ? 2D, maps RTT patch over
  surface)
• b(x,v) RTT (4D ? 25D, tabulated over small
  spatial patch)

applies macro-scale transferred radiance to
meso-scale RTT
18
Preprocessing

• compute transfer matrix Mp at mesh vertices
  Sloan02
• generate BTF over small patch b(xp,vp,d) Liu01
• convert BTF to RTT by integration b(xp,vp)
• build ID map between surface and RTT patch q(xp)

19
ID Map q(xp)

• RTT texel 25x3 components, 8x8 views
• impractical to store/synthesize unique RTT texel
  per surface point (2k x 2k)
• instead synthesize index into a small patch (64 x
  64)
• 2D rather than 4,800D
• but cant interpolate indices

20
ID Map
21
Preprocessing - ID map
RTT
mesh
22
Preprocessing - Examples
646488
646488
646488
128128 (view-independent)
bunny model 10k vertices ID map 20482048
23
Run-Time Rendering
Vertex Info
Lighting
To Pixel Shader
CPU
Vertex Shader
24
Run-Time Rendering Pixel Shader
RTT
Pixel Shader
Exit Radiance
25
Results View Dependent Effects (Masking)
26
Results Demo
ATI Radeon 9800
27
Performance Statistics

• computing PRT
• 8 minutes for 10k vertices (25x25 transfer
  matrices)
• generating synthetic RTT
• 6-27 hours for 64x64 patch
• building ID map
• 20 minutes for creating atlas
• 4 hours for texture synthesis
• run time performance
• 14.5 frames per second
• 2.2Ghz Pentium IV with ATI Radeon 9700.

28
Limitations of Bi-Scale Radiance Transfer

• static geometry
• distant, low-frequency lighting
• coarse view dependence
• shallow meso-scale texture
• no highly specular materials

29
Contributions

• propose a bi-scale representation for radiance
  transfer
• parameterize BTF by spherical harmonics
• RTT (Radiance Transfer Texture)
• for fast area lighting
• generalize PRT via RTT rather than BRDF
• practical, meso-scale transport effects
• access RTT using an id map
• high spatial resolution using small patch

30
Future Work

• better spatial filtering of RTT
• robust sampling of PRT surface signal
• combine with CPCA
• handling meso-scale silhouettes
• general LOD control

31
Questions?
Thanks to Dan Ling, John Hart, Jingdan Zhang,
Paul Debevec, Stanford, and MPI.
32
Previous Work