Inverse Volume Rendering with Material Dictionaries Ioannis Gkioulekas1 Shuang Zhao2 Todd Zickler1 1Harvard 1 Kavita Bala2 Anat Levin3 2Cornell 3Weizmann Most materials are translucent 2 food skin jewelry architecture Photo credit: Bei Xiao, Ted Adelson We know how to render them Monte-Carlo rendering ? material parameters 3 rendered image Veach 1997, Dutré et al. 2006 We show how to measure them inverse rendering material parameters rendered image captured photograph 4 Our contributions 1. exact inverse volume rendering material • with arbitrary phase functions! 2. validation with calibration materials known parameters thin thick nondilutable solids 3. database of broad range of materials 5 Why is inverse rendering so hard? random walk of photons inside volume radiative transfer • volume light transport has very complex dependence material parameters material sample 6 thin thick nondilutable solids Light transport approximations random walk of single-bounce photons inside random walk volume Previous approach: single-scattering Narasimhan et al. 2006 thin 7 nondilutable thick solids Light transport approximations isotropic random walk of distribution of photons inside photons volume Previous approach: diffusion … … … … Jensen et al. 2001 parameter ambiguity material 1 material 2 8 ≈ ≠ thin nondilutable Papas et al. 2013 thick solids Inverse rendering without approximations random walk of exact inversion photons inside of random walk volume thin 9 nondilutable thick solids Our approach appearance matching i. material representation ii. operator-theoretic analysis iii. stochastic optimization 10 Background random walk of photons inside medium θ m = (σt σs p(θ)) 11 extinction coefficient σt scattering coefficient σs phase function p(θ) Phase function parameterization Henyey-Greenstein lobes Previous approach: single-parameter families Chen et al. 2006 Donner et al. 2008 Fuchs et al. 2007 g ∈ −1,1 Goesele et al. 2004 Gu et al. 2008 Hawkins et al. 2005 not general enough Gkioulekas et al. 2013 Holroyd et al. 2011 Jensen et al. 2001 McCormick et al. 1981 Narasimhan et al. 2006 Papas et al. 2013 Pine et al. 1990 Prahl et al. 1993 12 Wang et al. 2008 Dictionary parameterization tent phase functions dictionary of materials phase functions DD=={m {p11,, m p22,, …, …, pmQQ} } pD11 1 2 3 4 5 6 7 8 9 10 π6 5π4 π3 π π π 7 8 π π10 π11 9 p • arbitrary materials phase functions π2 π1 m p = Σ q πq m pqq • similarly for σt and σs σt = Σq πq σt,q σs = Σq πq σs,q 13 Our approach appearance matching m = Σq π q mq i. material representation ii. operator-theoretic analysis iii. stochastic optimization 14 Operator-theoretic analysis random walk of photons inside medium discretized random walk paths • propagation step τ τ τ τ m = (σt σs p(θ)) 15 Operator-theoretic analysis radiance at all medium points and directions discretized random walk paths • propagation step τ Ln+1(x, θ) = K Ln(x, θ) • total radiance radiance radiance after n after n+1 L = n Ln = (I - K)-1 Linput steps steps Σ L(x, θ) m = (σt σs p(θ)) • rendering operator R L(x, θ) = R Linput(x, θ) dictionary representation: m = Σq πq mq K(π) = Σq πq Kq R(π)= (I - Σ q πq Kq)-1 16 Our approach appearance matching m = Σq π q mq i. material representation R(π)= (I - Σ q πq Kq)-1 ii. operator-theoretic analysis iii. stochastic optimization 17 Stochastic optimization appearance matching min ǁ photo - render(π) ǁ2 π analytic operator expression for gradient! 𝜕loss π 𝜕πq = render(π) · single-stepq · render(π) R(π) Kq R(π) • gradient descent optimization for inverse rendering 18 Stochastic optimization • exact gradient descent for k = 1, …, N, πk end 19 = πk - 1 - ak 𝜕loss π 𝜕πq 𝜋𝑘−1 N = a few hundreds * exact several CPU hours = intractable Stochastic optimization Monte-Carlo rendering to compute 102 samples noisy + fast 20 104 samples 𝜕loss π 𝜕πq 𝜋𝑘−1 106 samples accurate + slow Stochastic optimization • exact gradient descent for k = 1, …, N, πk = πk - 1 - ak 𝜕loss π 𝜕πq 𝜋𝑘−1 end N = a few hundreds * exact several CPU hours = intractable • stochastic gradient descent for k = 1, …, N, πk end 21 = πk - 1 - 𝜕loss π k a 𝜕πq 𝜋𝑘−1 N = a few hundreds * noisy few CPU seconds = solvable Theory wrap-up appearance matching min ǁ photo - render(π) ǁ2 π m = Σq π q mq i. material representation R(π)= (I - Σ q πq Kq)-1 ii. operator-theoretic analysis noisy 22 𝜕loss π 𝜕πq iii. stochastic optimization Our contributions 1. exact inverse volume rendering material • with arbitrary phase functions! 2. validation with calibration materials known parameters thin thick nondilutable solids 3. database of broad range of materials 23 Measurements appearance matching min ǁ photo - render(π) ǁ2 π multiple lighting multiple viewpoints 24 Acquisition setup material sample frontlighting camera 25 backlighting Acquisition setup material sample backlighting sample frontlighting material frontlighting camera backlighting top rotation stage camera 26 stage bottom rotationbottom rotation top rotation stage stage Validation calibration materials medium material particle material % Mie theory known parameters size very precise dispersions (NIST Traceable Standards) polystyrene monodispersions 27 aluminum oxide polydispersions Frisvad et al. 2007 Parameter accuracy comparison of ground-truth and measured parameters p(θ) θ -π polystyrene 1 polystyrene 2 all parameters estimated within 4% error 28 0 π polystyrene 3 aluminum oxide ground-truth measured Henyey-Greenstein fit Matching novel measurements comparison of captured and rendered images captured rendered rendered with HG polystyrene 3 images under unseen geometries predicted within 5% RMS error 29 profiles ground-truth measured Henyey-Greenstein fit Our contributions 1. exact inverse volume rendering material • with arbitrary phase functions! 2. validation with calibration materials known parameters thin thick nondilutable solids 3. database of broad range of materials 30 Measured materials hand cream shampoo olive oil robitussin curacao mixed soap whole milk mustard 31 material R whole milk 100.92 reduced milk 57.291 mustard 16.447 wine shampoo 8.111 hand cream 20.82 coffee liquid clay 37.544 milk soap 7,625 mixed soap 3.923 glycerine soap 0.201 robitussin 0.009 coffee 0.054 olive oil 0.041 blue curacao 0.01 red wine 0.015 extinction absorption G B R G B 105.345 102.768 0.013 0.013 0.041 62.46 63.757 0.007 0.007 0.024 milk 0.061 soap 0.451 18.536 6.457 0.057 9.919 10.575 0.178 0.328 0.439 liquid clay 32.353 41.798 0.011 0.011 0.012 reduced 48.25 67.949 0.004milk 0.004 0.005 8.004 8.557 0.003 0.004 0.015 4.018 4.351 0.003 0.005 0.013 0.202 0.221 0.001 0.001 0.002 0.001 0.001 0.012 0.197 0.234 0.051 0.049 0.275 0.309 0.406 0.039 0.012 0.062 0.047 0.353 0.012 0.021 0.083 0.048 0.011 0.013 0.011 0.122 0.351 0.402 thin nondilutable first moment R G B 0.954 0.963 0.946 0.954 0.957 0.942 0.155 0.173 0.351 0.907 0.882 0.874 0.188 0.247 0.265 0.312 0.442 0.512 0.164 0.167 0.155 0.33 0.322 0.316 0.955 0.949 0.943 0.906 0.977 0.98 0.911 0.899 0.906 0.946 0.954 0.975 0.955 0.973 0.979 0.947 0.975 0.977 thick solids Measured phase functions whole milk reduced milk mustard shampoo hand cream liquid clay milk soap mixed soap glycerine soap robitussin p(θ) θ coffee 33 olive oil curacao wine -π 0 π measured Henyey-Greenstein fit Synthetic images mixed soap glycerine soap olive oil curacao rendered image 34 whole milk Synthetic images chromaticity 35 Synthetic images mixed soap glycerine soap olive oil curacao rendered image 36 whole milk Effect of phase function measured phase function Henyey-Greenstein fit chromaticity rendered image p(θ) mixed soap θ 37 -π 0 π measured Henyey-Greenstein fit Discussion • more interesting materials: more general solids, heterogeneous volumes, fluorescing materials • other setups: alternative lighting (basis, adaptive, highfrequency), geometries, or imaging (transient imaging) • faster capture and convergence: trade-offs between accuracy, generality, mobility, and usability 38 Take-home messages 1. exact inverse volume rendering material • with arbitrary phase functions! 2. validation with calibration materials known parameters thin thick nondilutable solids 3. database of broad range of materials 39 Acknowledgements • Henry Sarkas (Nanophase) • Wenzel Jakob (Mitsuba) Funding: • • • • • • National Science Foundation European Research Council Binational Science Foundation Feinberg Foundation Intel Amazon Database of measured materials: http://tinyurl.com/sa2013-inverse 40 Error surface appearance matching 41 min ǁ photo - render(π) ǁ2 π Light generation MEMS light switch RGB combiner 42 green (535 nm) laser blue (480 nm) laser red (635 nm) laser