Analysis of Planar Light Fields from Homogeneous Convex Curved Surfaces Under Distant Illumination

Transcription

1 Analysis of Planar Light Fields from Homogeneous Convex Curved Surfaces Under Distant Illumination Ravi Ramamoorthi and Pat Hanrahan

2 Motivation Forward Rendering (Computer Graphics) Complex Lighting (Environment Maps)

3 Motivation Inverse Rendering (Computer Vision and Graphics) Estimate BRDF, Lighting, both BRDF and Lighting Theoretically Possible? Practically Feasible?

4 Reflection Equation L θ i θ o B INCOMING LIGHT B(x, θ o) = π/2 π/2 BRDF ρ( θ i, θ o) OUTGOING LIGHT L(x, θ i)ρ(θ i, θ o) cos(θ i) dθ i

5 Reflection Equation L θ i θ o B INCOMING LIGHT B(x, θ o) = π/2 π/2 BRDF ρ( θ i, θ o) ˆρ(θ i, θ o) = ρ(θ i, θ o) cos(θ i) OUTGOING LIGHT L(x, θ i)ρ(θ i, θ o) cos(θ i) dθ i

6 Reflection Equation L θ i θ o B INCOMING LIGHT B(x, θ o) = π/2 π/2 BRDF ρ( θ i, θ o) ˆρ(θ i, θ o) = ρ(θ i, θ o) cos(θ i) B(x, θ o) = π/2 π/2 OUTGOING LIGHT L(x, θ i)ρ(θ i, θ o) cos(θ i) dθ i L(x, θ i)ˆρ(θ i, θ o) dθ i

7 Approach: Reflection is Convolution L θ i α θ i B(α, θ o) = BRDF ρ( θ i, θ o ) π/2 π/2 θ α o B θ o L(α, θ i)ˆρ(θ i, θ o) dθ i

8 Approach: Reflection is Convolution L θ i α θ i B(α, θ o) = BRDF ρ( θ i, θ o ) π/2 π/2 L(α, θ i) = L(θ i ) = L(α + θ i) θ α o B θ o L(α, θ i)ˆρ(θ i, θ o) dθ i

9 Approach: Reflection is Convolution L θ i α θ i B(α, θ o) = BRDF ρ( θ i, θ o ) π/2 π/2 L(α, θ i) = L(θ i ) = L(α + θ i) B(α, θ o) = π/2 π/2 θ α o B θ o L(α, θ i)ˆρ(θ i, θ o) dθ i L(α + θ i)ˆρ(θ i, θ o) dθ i CONVOLUTION : B = L ˆρ

10 Related Work Graphics: Prefiltering Environment Maps Qualitative Observation that Reflection is Convolution Miller & Hoffman 84, Greene 86 Cabral Max Springmeyer 87, Cabral Olano Nemec 99 Vision, Perception D Zmura 91: Reflection as Operator in Frequency Space Basri & Jacobs: Lambertian Reflection as Convolution

11 Related Work Graphics: Prefiltering Environment Maps Qualitative Observation that Reflection is Convolution Miller & Hoffman 84, Greene 86 Cabral Max Springmeyer 87, Cabral Olano Nemec 99 Vision, Perception D Zmura 91: Reflection as Operator in Frequency Space Basri & Jacobs: Lambertian Reflection as Convolution Our Contribution: Formal Analysis in General 2D case Key insights extend to 3D (more recent work)

12 Fourier Analysis L(θ i ) = p ˆρ(θ i, θ o) = p B(α, θ o) = p L p e Ipθ i ˆρ p,q e Ipθ ie Iqθ o q B p,q e Ipα e Iqθ o q

13 Fourier Analysis L(θ i ) = p L p e Ipθ i ˆρ(θ i, θ o) = p B(α, θ o) = p q q ˆρ p,q e Ipθ ie Iqθ o B p,q e Ipα e Iqθ o B p,q = 2πL pˆρ p,q

14 Fourier Analysis L(θ i ) = p L p e Ipθ i ˆρ(θ i, θ o) = p B(α, θ o) = p q q ˆρ p,q e Ipθ ie Iqθ o B p,q e Ipα e Iqθ o B p,q = 2πL pˆρ p,q Note: Can fix output direction: B p (θ o) = 2πL pˆρ p (θ o)

15 Insights Reflected Light Field is Convolution of Lighting, BRDF Convolution Theorem Product of Fourier Coefficients Signal Processing: Filter Lighting using BRDF Filter Lighting Input Signal BRDF Filter Inverse Rendering is Deconvolution

16 Example: Directional Source at θ i = 0 L(θ i ) = δ(θ i ) L p = 1 2π B p,q = ˆρ p,q Reflected Light Field corresponds directly to BRDF Impulse Response of BRDF filter

17 Example: Mirror BRDF ˆρ(θ i, θ o) = δ(θ i + θ o) ˆρ p,q = δ p,q 2π B p,q = δ p,q L p Reflected Light Field corresponds directly to Lighting Gazing Sphere

18 Example: Lambertian BRDF Transfer function is Clamped Cosine No output dependence, drop index q B p = 2πL pˆρ p Lambertian BRDF is Low-Pass filter Incident Reflected

19 Properties: Lambertian BRDF Filter ˆρ 2p = ( 1)(p+1) 2π (4p 2 1) 0.16 Magnitude of Fourier Coefficient > Frequency (p) > Good approximation using only terms with p 2

20 Phong, Microfacet BRDFs Rough surfaces blur highlights Microfacet BRDF is Gaussian Hence, Fourier Spectrum also Gaussian Similar results for Phong (analytic formulae in paper) 1 Exp = 128 Exp = 8 Magnitude of Fourier Coefficient > Frequency(p) >

21 Inverse Rendering Known Lighting Unknown Known X Miller & Hoffman 84 D Zmura 91 Marschner&Greenberg97 BRDF Sato et al. 97, Yu et al. 99, Unknown Dana et al. 99 Debevec et al. 00, Marschner et al. 00,...? Sato et al. 99 (shadows) Often estimate Textured BRDFs (3rd axis of table)

22 Inverse Rendering General Complex Illumination? Most inverse-brdf methods use point source Outdoor methods: Sato&Ikeuchi94, Yu&Malik98 Well-Posedness, Conditioning? Well Posed if unique solution Well Conditioned if robust to noisy data Factorization of BRDF,Lighting (find both)? Sato et al. 99 use shadows

23 Inverse Lighting L p = 1 2π B p,q ˆρ p,q Well posed unless ˆρ p,q vanishes for all q for some p. Well conditioned when Fourier spectrum decays slowly. Need high frequencies in BRDF (sharp specularities) Ill-conditioned for diffuse BRDFs (low-pass filter) Mirror Lambertian

24 BRDF estimation ˆρ p,q = 1 B p,q 2π L p Well Posed if all terms in Fourier expansion L p nonzero. Well Conditioned when Fourier expansion decays slowly. Need high frequencies in lighting (sharp features) Ill-conditioned for soft lighting (low-frequency) Directional Source Area Source (same BRDF)

25 Light Field Factorization Up to a global scale, Light Field can be factored Can simultaneously estimate Lighting, BRDF Number of Knowns (B) > Number of Unknowns (L, ρ) (B 2D) > (L 1D + ρ 1/2(2D)) Explicit Formula in paper

26 3D Fourier Series Spherical Harmonics Y lm (θ, φ) Representation Matrices of SO(3) D l mm (α, β)

27 3D Fourier Series Spherical Harmonics Y lm (θ, φ) Representation Matrices of SO(3) D l mm (α, β) L(θ i, φ i ) = l,m L lm Y lm (θ i, φ i ) ˆρ(θ i, φ i, θ o, φ o) = ˆρ lq,pq Ylq(θ i, φ i)y pq (θ o, φ o) l,p,q B(α, β, θ o, φ o) = ( B lmpq D l mq (α, β)y pq (θ o, φ o) ) l,m,p,q

28 3D Fourier Series Spherical Harmonics Y lm (θ, φ) Representation Matrices of SO(3) D l mm (α, β) L(θ i, φ i ) = l,m L lm Y lm (θ i, φ i ) ˆρ(θ i, φ i, θ o, φ o) = ˆρ lq,pq Ylq(θ i, φ i)y pq (θ o, φ o) l,p,q B(α, β, θ o, φ o) = ( B lmpq D l mq (α, β)y pq (θ o, φ o) ) l,m,p,q 2D: B pq = 2πL pˆρ p,q 3D: B lmpq = L lmˆρ lq,pq

29 Implications Lambertian BRDF 2D: Only first 2 Fourier coefficients important 3D: First 2 orders of spherical harmonics 99% energy Only the first 9 coefficients are important Similar results independently derived by Basri & Jacobs Formally, recovery of radiance from irradiance ill-posed See On the relationship between Radiance and Irradiance: Determining the illumination from images of a convex Lambertian object (submitted) Phong & Microfacet BRDFs Gaussian Filters. Results similar to 2D

30 Practical Issues (in 3D) Frequency spectra from Incomplete Irregular Data Concavities: Self-Shadowing and Interreflection Textures: Spatially Varying BRDFs

31 Practical Issues (in 3D) Frequency spectra from Incomplete Irregular Data Concavities: Self-Shadowing and Interreflection Textures: Spatially Varying BRDFs Issues can be addressed; can derive practical algorithms Use Dual Angular and Frequency-space Representations Associativity of Convolution See A Signal Processing Framework for Inverse Rendering (submitted)

32 Experiment: Cat Sculpture 3 photographs of cat sculpture of known geometry Microfacet BRDF under complex unknown lighting Lighting also estimated Then use recovered BRDF for new view, new lighting

33 Results: Cat Sculpture Images below show new view, new lighting REAL PHOTOGRAPH Numerical values verified to within 5% RENDERED IMAGE

34 Implications for Perception Assume Lambertian BRDF, no shadows Perception: Separate Reflectance, Illumination Low frequency lighting, High frequency texture Theory formally: lighting only low-frequency effects Find high-frequency texture independent of lighting But ambiguity regarding low-frequency texture, lighting

35 Conclusion Reflection as convolution Fourier analysis gives many insights Extends to 3D and results in practical inverse algorithms Signal-Processing: A useful paradigm for Forward and Inverse Rendering