Acquisition DISCO of Translucent Object
Michael Goesele, Hendrick Lensch, Jochen Lang, Christian Fuchs, Hans-Peter Seidel
Past Work Acquisition
BRDF Acquisition cannot model subsurface scattering Image Acquisition Captured with fixed viewpoints [Debevec 01], fixed lighting [Wood 00], or both [Levoy 96] [Jensen 01] allows arbitrary modeling of only homogenous materials
Past Work Rendering
Rendering with BSSRDFs
Expensive and slow
Dipole Approximation
Physically correct for homogenous, infinite half-space
Most methods not based on measured data
Goals
Capture the exact behavior of real translucent objects
Heterogenous properties (eg. cracks, hollow objects, volumetric variations)
Allow for modeling in arbitrary lighting with arbitrary viewpoints Integration into rendering systems ([Lensch 03])
DISCO method
Digital Imaging of Subsurface sCattering Objects Want to measure Rd(xi; xo) per color channel for all xi and xo.
DISCO (cont.)
Illuminate at a single surface point xi with known incident radiance L (xi; i), observe L (xo; o) Invert to find Rd(xi; xo)
DISCO (cont. again)
Finite width beams that can enter at large angle of incidence i
Omit samples with threshold
i some
Data larger
than Storage and Access
Raw data is several hundred gigabytes High sampling density around incident, more coarse sampling further away
Global response matrix F of throughput factors; interpolate vertices data does not fill
Throughput Interpolation
Some vertices are never lit Case Fk,c, k far from point of illumination c
Interpolate iteratively from neighboring vertices
Case Fk,c, k close to point of illumination c
Use distanced-weighted avg. response of neighbors
Interpolation (cont.)
Local Response
Store filter kernals a la [Lensch 03] Most laser peaks between discrete texels
Shift peak to all 4 neighboring texels weighted by m(d) = c1 e 1d + c1 e 2d , d is distance to peak location
Interpolate filter kernels with same method as throughput factors
Rendering
Direct port into Lensch et al.'s approach Can substitute for dipole approximations, Monte Carlo/Photon Mapping evaluations
Pretty Pictures! (and results)
Pretty Pictures! (and results, take 2)
Pretty Pictures! (and results, take 3)
Conclusions
Surface coverage limited by occlusion Additional imagery vs. Acquisition time Could try to plan acquisition images
Pre-determine lit surface positions