Advanced Computer Graphics CS 563: VPL based RT GI Techniques - PowerPoint PPT Presentation

Advanced Computer Graphics CS 563: VPL ‐ based RT ‐ GI Techniques William DiSanto Computer Science Dept. Worcester Polytechnic Institute (WPI)

Instant Radiosity  1997 Keller  Fast and physically accurate  Models diffuse lighting

Virtual Point Light  Hemispherical light with cosine falloff  Positioned by emitting particles from light source  Some way to end bounces

Derivation of Instant Radiocity  Rendering Equation  Operator R: Express as sum of indirect illumination

Derivation of Instant Radiocity  Express rendering equation as an integration over the observable light from first hit surface  Express surface lighting term by sampling direct and indirect paths

Rendering Operator: Reflect >= 1  Express Rendering Equation as sampling along variable length paths  Function V could be visibility from AO (inviting error)  Compute Indirect: over Path Lengths (j), Pixels, Paths, and Light Sources

Derivation of Instant Radiocity  Radiance after j reflections:  Assume perfectly diffuse BRDF (constant term)

IR Result Samples from Area Light Bounces

Making the Algorithm Practical  Limit number of samples virtual point lights and number of bounces  Proportional to reflectivity of the media  Assume less the perfect reflection ( < 1)  Compute light observed in terms of radiance from original light source propagated over some number of bounces

Implementation of Instant Radiocity Start with N particles at light  Determine samples remaining  Select a originator from light source  Trace through reflections  Loop until no samples are left  Next Slide : Light Power : Average Diffuse : Projection Direction

Implementation of Instant Radiocity  Render all 1 st 2 nd , 3 rd bounce VPLs together  All radiation from a sample particle rendered in one iteration : hit point : diffuse attenuation :random direction

Sampling  Random  Jitter  Halton

Omni ‐ Directional RSMs

Problems with IR  Requires many renders, possibly shadow maps  Requires many VPLs  100+ for static scene  Interactive rates reached with averaging renders  1000+ for dynamic scene

Problems with IR  Specular surfaces reveal locations of VPLs

Problems with IR  Render many detailed shadow maps

Incremental Instant Radiosity  Scenes can be rendered fairly accurately with only one bounce  Use Reflective Shadow Map to distribute first bounce Virtual Point Lights

Incremental Instant Radiosity  Reuse VPLs from previous renders  Budget 4 ‐ 8 new VPLs per new frame  Allows for moving light sources  Assumes a static scene (could be large)  Good frame rates 40+

Procedure: Find Bad VPLs  1. Determine valid VPLs  2. Remove invalid VPLs  Occluded from light source  Behind view of light source

Procedure: Find Bad VPLs  Generate Voronoi Diagram and Delaunay Triangulation for point set (left)  Some VPLS may be removed if they contribute to regions where virtual point light density is too great (right) Characterized by having short edge lengths 

Procedure: Seed VPLs  3.a Create VPLs within budget and render SMs Long edges are good places to seed new VPLs   4. Compute new Voronoi Area and weight intensity for VPLs accordingly

Procedure: Project  3.b Render Shadow maps for new VPLs Project new VPL distribution onto hemisphere 

Procedure: Render  5. Perform deferred render pass  store: position, normals, color  6. Tile the G ‐ Buffer (will reduce texture lookups)  7. Loop over tiles  use subset of VPLs to color each tile, results in noise

Procedure: Recombine + Smooth  8. Combine tiles back into full resolution image  9. Use spatially aware box filter (example on 4x4 tiling)  Same size as tiling  May require some tuning

IIR: Results

IIR: Problems  Diffuse surfaces only  View dependent  Dynamic scenes will have lagging shadows  This GI takes 70+ % of rendering budget  Does have nice overall render times

Imperfect Shadow Maps  Attempts indirect illumination for dynamic scenes  For many scenes: Indirect illumination varies smoothly  Each VPL has a small contribution

Imperfect Shadow Maps  Depth Renders do not need to be terribly accurate  Attempt to render 1000+ ISMs

Procedure: VPL Generation  Render Omni ‐ Directional RSM  Randomly seed VPLs by importance (some PDF)

Procedure: Pre ‐ Processing  Randomly sample geometry  Select triangles with probability proportional to area  Select a random 3D point on each selected triangle  Store barycentric coordinates with points  Used to calculate normal and reflectance

Procedure: ISM Generation  Each VPL selects a sub sample of points  Points are splat onto each depth map (GL_Points)  Dynamic scenes are handled by deforming points

Procedure: Pull Push Interpolation  Vertex buffer distributes points to ISMs evenly  Gaps are filled with push ‐ pull, errors average out

Procedure: Pull Push Interpolation  Interpolation performed in texture space  Parallel operation  Averaging of all ISM increases shadow smoothness

Procedure: Interleaved Sampling  Filter indirect light separately from direct light  Indirect light is interleaved  Same as in incremental instant radiocity  May need to tune parameters

Procedure: Multiple Bounces  Distribute from cube maps of omni ‐ directional light 1024 VPLs first bounce (left)  + 256 VPLs second bounce (middle)  + 256 VPLs third bounce (right) 

Problems  Works best with low gloss surfaces  Arbitrary tuning  Will not operate the same for every scene  For different kinds of lighting  As of 2009 offers only interactive frame rates  Does not scale well for large scenes  Indirect lighting still accounts for the majority of the render budget

Results: Complex Geometry  Extends well to arbitrary shapes  1024 VPLs  256x256 with subset of 4,000 points each

Results: Direct Lighting  Complex area lights 512 VPLS  256x256 with subset of 8,000 points each 

Results: Glossy  4096 VPLs, 64x64 resolution, subset of 2,000 points each  Problem with hard shadows

Results: Environmental Maps  Direct Environment Lighting 1024 VPLs  256x256 with subset of 8,000 points each 

Results: Sponza Scene  15 ms Rendering  7 ms VPL generation  4 ms Blur  44 ms ISM  11 ms Direct Lighting  8 ms Pull ‐ Push

Results: Accuracy Parameters can be manually adjusted to fit the scene (splat size, resolution, point samples)

Results: Render Time

References  Alexander Keller, Instant Radiosity, Proceedings of the 24th annual conference on Computer graphics and interactive techniques, p.49 ‐ 56, August 1997  Laine, S., Saransaari, H., Kontkanen, J., Lehtinen, J., and Aila, T. 2007. Incremental Instant Radiosity for Real ‐ Time Indirect Illumination. In Proc. of the Eurographics Symposium on Rendering , 277 ‐‐ 286.

References  T. Ritschel , T. Grosch , M. H. Kim , H. ‐ P. Seidel , C. Dachsbacher , J. Kautz, Imperfect shadow maps for efficient computation of indirect illumination, ACM Transactions on Graphics (TOG), v.27 n.5, December 2008  Carsten Dachsbacher , Jan Kautz, Real ‐ time global illumination for dynamic scenes, ACM SIGGRAPH 2009 Courses, p.1 ‐ 217, August 03 ‐ 07, 2009, New Orleans, Louisiana