470 likes | 503 Views
Real-time Global Illumination with precomputed probe. Baek Jun Kim KAIST (Korea Advanced Institute of Science and Technology). Review. Ambient Occlusion Rendering with Neural Networks (20184230 김세훈 ) 1) Deep shading (2D frame based learning) 2) DRAO (3D geometry based learning).
E N D
Real-time Global Illumination with precomputed probe Baek Jun Kim KAIST (Korea Advanced Institute of Science and Technology)
Review • Ambient Occlusion Rendering with Neural Networks (20184230 김세훈) • 1) Deep shading (2D frame based learning) • 2) DRAO (3D geometry based learning)
Global Illumination • Direct illumination only • Global illumination (indirect illumination)
Ray Tracing • Ray Tracing • To implement Global Illumination • In Real-time? • Too costly to compute at interactive rates • Even several hours for a image • How to realize Real-time Ray Tracing • Precomputation • … and others
Papers • Real-Time Global Illumination using Precomputed Light Field Probes • Real-time Global Illumination by Precomputed Local Reconstruction from Sparse Radiance Probes
Real-Time Global Illumination using Precomputed Light Field Probes Morgan McGuire, Michael Mara, Derek Nowrouzezahrai, and David Luebke. In I3D 2017. 11.
Summary Precomputing Probe Radiance, normal, and distance map Intersection testing with distance map Ray casting using Probe Shading with Probes 1) Spatial-Temporal Radiance Denoising (BRDF factorization) 2) Irradiance with Pre-Filtered Visibility (Geometry aware)
Real-Time GI using Precomputed Light Field Probes • A scene computed in 42ms (212fps) • Let me show its previous works briefly: • Image Based Lighting • Voxel Cone Tracing
Image-based Lighting • Cube map (Environment mapping) • In this paper, every probe get its cube map.
Voxel Cone Tracing • Using Sparse Voxel Octree • Optimized in modern GPU • Hierarchical
Precomputed Light Field Probes • Rasterization to render cube map • Octahedral parameterization for efficiency • Two-level MIPmap hierarchy (heuristic)
Light Field Probe Ray Tracing • How to use precomputed probe? • We can trace a ray! • Procedure • Selecting a probe • Single-probe tracing routine • Selecting fall-back probes
Selecting a Probe • The probe whose center lies closest to the ray Ray Probe
Intersection test • Intersection test by comparing depth map value in the probe and radial distance from the probe’s center Ray Depth map in the probe Surface Probe
Intersection test (2) • If depth map value is smaller than radial distance, the ray has either hit a surface or pass behind the surface Ray Depth map in the probe Surface Probe
Intersection test (3) • By using normal map in probe, we can distinguish between HIT and behind the surface Ray Depth map in the probe Surface Probe
Selecting fall-back probes • If unable to find a definitive hit or miss, pick another probe
Shading with Light Field Probes • Now we can trace a ray with probes, then? • 3 Global Illumination Algorithm • Compute indirect lighting using MC & Importance Sampling • Unbiased, but slow • We have to reduce the number of ray samples • Spatial-Temporal Radiance Denoising • Irradiance with Pre-Filtered Visibility
Spatial-Temporal Radiance Denoising • Factorize BRDF into Lambertian term and Glossy Term • Indirect illumination pass traces one ray sampled from each of them and writes the result to each buffer • For Lambertian reflection, we compute incident irradiance instead of radiance
Spatial-Temporal Radiance Denoising (2) • Lambertian reflections is spatially slower-varying • Use of wider filter • Glossy reflection is spatially faster-varying • Use of narrower filter
Irradiance with Pre-Filtered Visibility • How to compute irradiance? • Precompute irradiance map • Interpolate irradiance map at eight probes within the cube of the intersection • Reduce number of texture operation → really fast • Limitation: Lighting and darkness leaking
Irradiance with Pre-Filtered Visibility (2) • Geometry-aware approach • While precomputing irradiance map, we also compute radial distance map • At runtime, we use the radial distance map for interpolating irradiance
Summary Precomputing Probe Radiance, normal, and distance map Intersection testing with distance map Ray casting using Probe Shading with Probes 1) Spatial-Temporal Radiance Denoising (BRDF factorization) 2) Irradiance with Pre-Filtered Visibility (Geometry aware)
Real-Time Global Illumination using Precomputed Light Field Probes Ari Silvennoinen and Jaakko Lehtinen. ACM Transactions on Graphics, November 2017.
Summary • To get indirect illumination, combined two methods • Precomputed Light Transport • Interpolation • Factorize Global Transport into (1) Local Transport and (2) Probe Transport
Real-time GI by Precomputed Local Reconstruction from Sparse Radiance Probes • GI = Direct illumination + Indirect illumination • How to get indirect illumination?
How to get indirect illumination? • In this paper, • Precomputed light transport • Direct-to-Indirect transfer [Hašan et al. 2006] • Interpolation method • Interpolated irradiance and bandlimited radiance [E.g. Ward 1992]
Direct-to-Indirect transfer [Hašan et al. 2006] • Precomputed radiance transfer method • Precompute transfer matrix that map direct illumination into indirect illumination
Interpolation method • Interpolation from radiance probes • are spatial interpolation weight ω p2 p1 x
Limitations • a) Direction mismatch • b) Visibility mismatch
Visibility-Aware Interpolation • Ψ for direction • Vi(ω) for Visibility
Probe Radiance • High-resolution angular sampling of the probes’ radiance is expensive • Replace it by Spherical Harmonics form • λ: probe radiance vector • Y: SH basis function (basis func. index j) Spherical harmonics is a kind of Laplace transform of Sphere
Local transport Operator • Local transport Operator transforms the probe radiance vector λ to interpolated incident radiance at x as a function of the continuous direction ω
Precomputed Transport • transport coefficients ϕij
Factorized transport • Factorize global transport into local transport and probe transport • Interpolate with probes
Local Transport • Precomputed local transport matrix: accounts for near field effects between the receiver and the mutually visible surfaces seen by the probes • Vi(ω) makes matrix
Probe Transport • Captures radiance in a sparse set of probes at runtime
Implementation - Precomputation • Probe locations and radii: • Every surface point is under the support of at least one probe • Every surface point is visible from at least one probe • Use greedy algorithm • Cluster Receivers into a vector
Implementation - Runtime • Computing Probe Transport • Direct illumination light map is updated every frame
Summary • To get indirect illumination, combined two methods • Precomputed Light Transport • Interpolation • Factorize Global Transport into (1) Local Transport and (2) Probe Transport
Pros and Cons • Real-Time Global Illumination using Precomputed Light Field Probes • Dynamic Camera • Static and Dynamic Environment • Real-time Global Illumination by Precomputed Local Reconstruction from Sparse Radiance Probes • Only Static Environment • Dynamic Light Source
Quiz • In Real-Time Global Illumination using Precomputed Light Field Probes, which map of the probe used to test intersection during Light Field Probe Ray Tracing? (1) Shadow map (2) Distance map (3) Irradiance map (4) Radiance map • In Real-time Global Illumination by Precomputed Local Reconstruction from Sparse Radiance Probes, global transport is factorized into 2 parts: precomputed ( ) transport and ( ) transport which is computed at runtime (1) probe, local (2) local, probe (3) SH, SVD (4) SVD, SH