Simple and Robust Iterative Importance Sampling of Virtual Point Lights Iliyan Georgiev Philipp Slusallek

Motivation Instant Radiosity (IR) – two-pass Previous approaches Cheap pre-processing Expensive rendering Previous approaches Bidirectional Instant Radiosity [Segovia et al. 06] Metropolis Instant Radiosity [Segovia et al. 07] Difficult to implement Multiple sampling strategies Many parameters Stratification problems “One-pixel image” assumption November 22, 2018

Our Method Simple extension of IR Probabilistically accept VPLs Generate VPLs from light sources only Probabilistically accept VPLs Proportionally to total contribution All VPLs bring the same power to the image “One-pixel image” assumption Minimum importance storage Filter VPLs on the fly November 22, 2018

Probabilistic VPL Acceptance VPL energy 𝐿 𝑖 = 𝐿 𝑖 𝑝 𝑖 𝑝 𝑖 = 𝐿 𝑖 𝑝 𝑖 0 1 𝜒 0, 𝑝 𝑖 𝑡 𝑑𝑡 One-sample Monte Carlo integration with 𝜉 𝐿 𝑖 = 𝐿 𝑖 𝑝 𝑖 , 𝜉< 𝑝 𝑖 &0, else Allows to control VPL density November 22, 2018

Importance Sampling Want 𝑁 VPLs with equal total image contribution Φ 𝑣 = Φ 𝑁 For each VPL candidate 𝑖 with energy 𝐿 𝑖 Estimate total image contribution Φ 𝑖 Russian roulette decision with 𝑝 𝑖 =min Φ 𝑖 Φ 𝑣 +𝜀,1 Accept with energy 𝐿 𝑖 𝑝 𝑖 Discard See paper for details November 22, 2018

Estimating Image Contribution Computing Φ 𝑖 Cast a number of camera rays Connect VPLs to camera points to estimate Φ 𝑖 Camera points – analog of importons Computing Φ 𝑣 Iteratively Start with Φ 𝑣 =0 Loop Render frame Accumulate Φ 𝑣 Single pass – path tracing, using VPLs, etc. November 22, 2018

Our Extension (0.07 acceptance) Results Instant Radiosity Our Extension (0.07 acceptance) November 22, 2018

Average acceptance probability: 0.28 Results Average acceptance probability: 0.28 November 22, 2018

Average acceptance probability: 0.23 Results Average acceptance probability: 0.23 November 22, 2018

Wrap Up Simple extension of IR Generate VPLs from light sources only Probabilistically accept VPLs on the fly No additional storage Easy to parallelize Few parameters “One-pixel image” assumption November 22, 2018

Iliyan Georgiev