Amortizing Samples in Physics-Based Inverse Rendering using ReSTIR

Supplemental Materials

In this supplemental material, we show animated comparisons among our method, path tracing (PT), and RIS [Talbot et al. 2005]. Both PT and RIS uses the PSDR [Zhang et al. 2020] formulation.

• In this scene, we put a picture inside a room. The picture is lit by 1 area light. We have another 2 area lights outside the room.
• We reconstruct the picture with a training set of 1 image.
• This is an equal-time comparison. We adjust the sample rate to make all methods run for approximately 12 minutes.

• In this scene, we put a picture with 3 wood sorrel leaves inside a room. The picture is lit by 2 spotlights lights and 1 fill light.
• We reconstruct the picture with a training set of 1 image.
• This is an equal-time comparison. We adjust the sample rate to make all methods run for roughly 1.7 hours.

• In this scene, we have an environment map around the kitty and try to reconstruct its texture map from a constant texture map.
• We optimize the albedo of the picture with a training set of 32 images.
• This is an equal-time comparison. We adjust the sample rate of all methods to make them run for approximately 30 minutes.

• In this example, we try to optimze the pose of a tree by merely looking at its shadow.
• We try different combinations of methods in forward and backward pass.
• We adjust the hyper-parameters of each method to make all of them run for roughly 16 minutes.

• In this case, we start from a disk and try to reconstruct a polygon by only looking at its shadow on a plane. The polygon is lit by a large area light.
• This is an equal-time comparison. We adjust the sample rate of all methods to make all of them run for 2 minutes.

• This example involves a coin. The object is lit by three spotlights and one blocked area light.
• Using 6 target images (with one shown), we optimize the shape of coin from an ellipsoid.
• This is an equal-time comparison. We adjust the sample rate of all methods to make all of them run for 18 minutes.

• This example involves an octagon. The object is lit by 17 spotlights.
• Using 17 target images (with one shown), we optimize the shape of coin from a sphere by merely looking at its shadow.
• This is an equal-time comparison. We adjust the sample rate of all methods to make all of them run for 105 minutes.

• This example involves a world map relief in a large room with a window. The relief is lit by a large area light outside that is partially blocked.
• Using 6 target images (with one shown), we optimize the shape of relief.
• This is an equal-time comparison. We adjust the sample rate of all methods to make all of them run for 90 minutes.