Antithetic Sampling for Monte Carlo Differentiable Rendering

Supplemental Materials

Cheng Zhang^1,2, Zhao Dong², Michael Doggett^2,3, and Shuang Zhao^{1 1}University of California, Irvine          ²Facebook Reality Labs          ³Lund University

In this document, we show several inverse rendering examples to validate the benefits of our antithetic sampling technique. For each example, we provide a equal-time comparison between two sets of results with gradients estimated using and not using antithetic sampling method (combined with PSDR). The parameter RMSE plots show root-mean-square error of the optimized parameters. This information is not used by the optimizations.

The table below summarizes the performance statistics and optimization configurations for the inverse rendering examples. The reported Time is measured per iteration on a workstation with 8-core Intel i7-7820X CPU.

Scene	# Param.	# Iter.	Time (per iteration)
Glass bunny	1	200	125 s
Bunny	3	200	20.78 s
Mug	3	120	6.78 s
Einstein	100	300	1.87 s

Glass bunny

• This example contains a glass bunny rotating vertically. The smooth glass is approximated using rough dielectric BSDF with low roughness.
• We search for the rotation degree (DOF=1) of the bunny.

Bunny

• This example contains a diffuse bunny placed inside a glass cube, approximated using rough dielectric BSDF with low roughness.
• We search for the orientation (DOF=3) of the bunny.

Mug

• This example contains a glass mug with a small area light inside.
• We search for the position (DOF=3) of the small emitter by only looking at the caustic pattern on the ground

Einstein

• This example contains a glossy anisotropic reflector and an area light with a pattern of distorted Einstein
• We search for the height map (DOF=100) attached to the reflector to correct the distortion