1. Real-time Indirect Lighting Using Clustering Visibility Zhao, Tobias, Thorsten, Jan* *University College London

2. Motivation  Indirect lighting is Crucial for GI rendering. Check following 2 video clips. –Real-time performance for fully deformable/dynamic scene is open problem  Visibility computation is still the key bottleneck. –Approximate visibility for indirect lighting is quite reasonable from perception viewpoint. –Practical indirect lighting rendering for real game/film development. Video1 Video1 Rui Wang et al. EGSR07 Video2 Video2 Casten Dachsbacher et al. I3D06

3. state-of-the-art  Methods based on hierarchical radiosity  Interactive rendering performance  Moderate-sized scene  PRT-based methods  Require precomputation,  Static/rigid dynamic scene. [Casten et al. sig07] [Dong et al. pg07] [Iwasaki et al. egsr07][Wang et al. egsr07][Liu et al. eg07]

4. State-of-the-art  Image-based SH exponential –Deformable, low-freq lighting (SH) and only diffuse interreflection  Methods based on instant radiosity [Keller sig97] –The left two only support static/semi-dynamic scene –Tobias’ method support fully deformable scene, VPL + ISM [Sloan et al. pg07] [Laine et al. egsr07][Casten et al. I3d05,06][Tobias et al. Submitted to sigasia08

5. Rethinking Instant radiosity methods Direct Light VPL Receiver Pixel 1000 VPLs  Good quality 1000 VPLs equals to rendering 1000 SMs, Visibility computation becomes bottleneck! Cornell box  Can we utilize virtual area lights?  How to fast approximate the visibility function of virtual area lights? Key: Convolution Soft Shadow Map[SIG08] So, 1000 SMs  40 CSSMs!

6. Goal and Contributions  Goals –Real-time indirect lighting rendering for fully dynamic scenes. (Focus on indirect lighting)  Planned contribution –Novel: Separate visibility and shading computation of indirect lighting.  Approximate the indirect visibility by the cluster of VPLs  Virtual Area Light (VAL)  Shading still based on VPLs  keep the accuracy and all-freq indirect lighting (Caustics) –Delta: A simplified GPU-based k-means method for clustering VPLs in fully dynamic scene scheme (mesh animation).  Novel && practical clustering criteria for indirect lighting.

7. Our approach Direct Light VPL Cluster Receiver Pixel

8. Our algorithm steps:  Step 1: Render the scene by direct lighting and at the same time generate the VPLs.  Step 2: GPU-based clustering of VPLs  VALs  Step 3: For each VAL, approximate the visibility function  Based on convolution shadow map [Our SIG08 Paper]  Step 4: Per-pixel rendering: visibility based on VALs and shading based on all VPLs.

9. Intended Results  selling points –Real-time practical indirect lighting rendering for fully dynamic scene –Novel Virtual Area Light (VAL) concept applied into indirect lighting. –Our GPU-based clustering method is simple && practical, which can be applied in lots of graphics cases.  who might be interested? –Game/Film-tech developers

10. Example images

11. Validation  Side-by-side Comparison –Path tracing (ground truth) –Traditional instant radiosity methods (VPLs)  Perception experiment?

12. Major Steps  project team –Zhao, Tobias, Thorsten, Jan  Current status: –We have the code ready for step 1 and 4 –We have verified the kernel part for step 2 and 3  Unknown: Good clustering criterion, need experiments  which supporting algorithms are needed? –Good k-means clustering criterion  we want to get a square area from a group of points.  required hardware –Powerful PC equipped with powerful GPU

13. Questions and suggestions?  Possible venue: –EG09, I3D09?  Any suggestions? Gracious Thanks