Sample-based Monte Carlo Denoising using a Kernel-Splatting Network

Status
Read
Field
Monte Carlo Rendering
Denoising
Ray Tracing
Conference / Journal
SIGGRAPH
Year
2019
Link
Created
2020/08/31 00:56
Novel point of SBMC is...
1.
Sample-wise reconstruction allows more expressive contributions for reconstruction than pixel-wise reconstruction
2.
Splatting helps the model to be trained without shortcut (reducing range of kernel to reduce loss) compared to gathering.
3.
Splatting is permutation-invariant, which allows sample-wise reconstruction to show better performance.
1.
Scene generation
Geometry → ShapeNet, SunCG dataset
Camera → FOV, depth-of-field effect , shutter speed, motion blur (translation motion)
Material → materials from PBRT v2, bias to diffuse materials, random texture & bumps & UV scaling, Describable Textures Dataset
Lighting → HDRI Haven environmental maps
Rejection sampling → remove too much simple dataset
2.
Interesting Results
By maintaining albedo, it could denoise effects such as motion blur & defocus
Beneficial on low-sample, time increases linearly to samples
Scene complexity does not affect the runtime.