280 likes | 471 Views
The Natural-Constraint Representation of the Path Space for Efficient Light Transport Simulation Anton S. Kaplanyan 1,2 and Johannes Hanika 1 and Carsten Dachsbacher 1 1 Karlsruhe Institute of Technology, 2 Lightrig. Motivation. [Wikipedia Commons]. [Wikipedia Commons].
E N D
The Natural-Constraint Representation of the Path Space for Efficient Light Transport Simulation Anton S. Kaplanyan1,2 and Johannes Hanika1and Carsten Dachsbacher1 1Karlsruhe Institute of Technology, 2Lightrig
Motivation [Wikipedia Commons] [Wikipedia Commons] Rendered with manifold exploration [JakobMarschner12] Jewelry Rendered with out method
Intuition: Constraints in a Lens System • Two endpoints with specular constraints
Idea: Deviate from Specular Path • Rough scattering: use soft constraints
Outline • Introduction • Prior work • New generalized coordinates • Properties of the domain • Overview: Metropolis light transport • New half-vector mutation • Results • Conclusion Theory Practice
Rendering with Light Transport • Generate image by computing • Flux incident on the sensor • Sample all possible paths • Stochastic integration Path space
Path Integral Framework • Path integral • Measurement contribution is in area measure • Becomes singular with specular BSDFs Emission BSDFs Absorption Geometric terms
Prior Work: Specular Paths • Specular paths are hard constraints • Obey Fermat principle [Alhazen1021] • Ray transfer matrices used in optics [Gauss1840] • Pencil tracing in graphics [Shinya87] End point Start point Optical system
Prior Work: Specular Paths • Rendering specular paths with geometric knowledge • Solving for constraints [MitchellHanrahan92] • First and second order analysis [ChenArvo00] • Predictor-corrector perturbations [JakobMarschner12] • Our work is inspired by manifold exploration [Mitchell and Hanrahan 1992] [Chen and Arvo 2000] [Jakob and Marschner 2012]
Generalized Coordinates • Generalized coordinates: two endpoints + halfway vectors • Specular surfaces: halfway vectors coincide with surface normal • If halfway vector surface normal Path deviates from the specular path
Deviating the Half Vectors • A path in generalized coordinates
Path Contribution with New Formulation • Path measurement in new domain • Only one geometric term left (avoid singularities) • Transfer matrices are extended to scattering Area measure Emission Scattering distributions Camera responsivity Transfer matrix + Geometric term = Generalized G Half vector domain
Simplified Measurement Projected solid angle Surface area measure [Jakob13] [Walter et al.07] Half vector domain
Decomposition of Path Integral • Decorrelated islands • Set of 2D integrals • Easy-to-predict spectrum • Mostly changes local BSDF • Well-studied sampling
Mutation Strategy for Metropolis Light Transport Practical Rendering
Metropolis Light Transport • Take a path and perturb it [VeachGuibas97] • Specialized mutation strategies • Manifold exploration (ME) [JakobMarschner12]
Half-vector Space Mutation • Mutation 1. Perturb half-vectors 2. Find a new path • Similar machinery to ME (see ME paper) • Specular chains: fall back to ME • Jump over geometric parts • Take prediction as a proposal
Importance-Sample All BSDFs • Query avg. BSDF roughness • Approximate with Beckmann lobe • Sample as ~2D Gaussian • Known optimal step sizes from MCMC
Results: Kitchen HSLT VMLT MEMLT PSSMLT HSLT
Results: Necklace VMLT HSLT MEPT MEMLT HSLT
Conclusion • Convenient domain for paths on surfaces • Generalized coordinates • Beneficial properties of path integral • Sampling in generalized coordinates • Robust light transport (especially glossy and specular) • Importance sampling of all BSDFs along a path • Practical stratification for MLT
Limitations and Future Work • Geometric smoothness • Level of detail for displaced geometry • Rare events: needle in a haystack • Probability-1 search • Participating media • More dimensions, new soft constraints
Thank You Questions?
Backup: Stratification for MCMC • Expected change on the image from changes in half vectors Without stratification With stratification
Backup: Spectral Sampling • How to distribute step sizes among half vectors? • Spectral sampling for MC [SubrKautz2013] • Convex combination based on bandwidth (see paper) Without redistribution With redistribution