Inverse Global Illumination:Recovering Reflectance Models of Real Scenes from Photographs Computer Science Division University of California at Berkeley Yizhou Yu, Paul Debevec, Jitendra Malik & Tim Hawkins
Image-based Modeling and Rendering • 1st Generation---- vary viewpoint but not lighting • Recover geometry ( explicit or implicit ) • Acquire photographs • Facade, Plenoptic Modeling, View Morphing, Lumigraph, Layered Depth Images, (Light Field Rendering) etc.
Image-based Modeling and Rendering • Photographs arenot Reflectance Maps ! • 2nd Generation---- vary viewpoint and lighting for non-diffuse scenes • Recover geometry • Recover reflectance properties • Render using light transport simulation Illumination Radiance Reflectance
Previous Work • BRDF Measurement in the Laboratory • [ Ward 92 ], [Dana, Ginneken, Nayar & Koenderink 97] • Isolated Objects under Direct Illumination • [ Sato, Wheeler & Ikeuchi 97 ] • Isolated Objects under General Illumination • [ Yu & Malik 98], [ Debevec 98]
The Problem • General case of multiple objects under mutual illumination has not been studied.
Global Illumination Reflectance Properties Radiance Images Geometry Illumination
Inverse Global Illumination Reflectance Properties Radiance Images Geometry Illumination
Input Radiance Images [ Debevec & Malik 97] http://www.cs.berkeley.edu/~debevec/HDR
Synthesized Images Original Lighting Novel Lighting
Outline • Diffuse surfaces under mutual illumination • Non-diffuse surfaces under direct illumination • Non-diffuse surfaces under mutual illumination
Source Target Lambertian Surfaces under Mutual Illumination • Bi, Bj, Ei measured • Form-factor Fij known • Solve for diffuse albedo
Parametric BRDF Model [ Ward 92 ] N H Isotropic Kernel ( 3 parameters) Anisotropic Kernel ( 5 parameters)
Non-diffuse Surfaces underDirect Illumination N P2 H P1 P2 P1
Non-diffuse Surfaces under Mutual Illumination • LPiAj is not known. ( unlike diffuse case, where LPiAj = LCkAj ) Source Aj LPiAj LCkAj Pi Target LCvPi Ck Cv
Solution: iteratively estimate specular component. • Initialize • Repeat • Estimate BRDF parameters for each surface • Update and
Estimation of Specular Difference S • Estimate specular component of by Monte Carlo ray-tracing using current guess of reflectance parameters. • Similarly for • Difference gives S Aj LPiAj LPiAj Pi LCkAj Ck LCkAj LCvPi Cv
Recovering Diffuse Albedo Maps • Specular properties assumed uniform across each surface, but diffuse albedo allowed to vary.
Results • A simulated cubical room
Results for the Simulated Case Diffuse Albedo Specular Roughness
Results • A real conference room
Real vs. Synthetic for Original Lighting Real Synthetic
Diffuse Albedo Maps of Identical Posters in Different Positions Poster A Poster B Poster C
Inverting Color Bleed Input Photograph Output Albedo Map
Real vs. Synthetic for Novel Lighting Real Synthetic
Acknowledgments • Thanks to David Culler and the Berkeley NOW project, Tal Garfinkel, Gregory Ward Larson, Carlo Sequin. • Supported by ONR BMDO, the California MICRO program, Philips Corporation, Interval Research Corporation and Microsoft Graduate Fellowship.
Conclusions • A digital camera can undertake all the data acquisition tasks involved. • Both specular and high resolution diffuse reflectance properties can be recovered from photographs. • Reflectance recovery can re-render non-diffuse real scenes under novel illumination as well as from novel viewpoints.