Neural Importance Sampling of Many Lights

Figueiredo et al.

Evaluation

In this supplementary visualizer, we compare our approach against three other methods: Estevez and Kulla [2018] (ATS), Lin and Yuksel [2020] (SLCRT), Bitterli et al. [2020] (ReSTIR), and Wang et al. [2021] (VARL). We report three metrics: ꟻLIP (Andersson et al. [2020]), mean absolute percentage error (MAPE), and relative mean squared error (relMSE). relMSE is computed by discarding the top 0.01% errors to alleviate the contribution of outliers. For the virtual point light (VPL) scene (San Miguel) we use uniform light sampler for non-VPL light. relMSE is computed by discarding the top 0.01% errors to alleviate the contribution of outliers.

Equal Time:

Bathroom

Bedroom

Bistro

Living Room

Sanmiguel

Staircase

Staircase2

Zero Day

Equal Sample Count:

Bathroom

Bedroom

Bistro

Living Room

Sanmiguel

Staircase

Staircase2

Zero Day