28. deep adaptive sampling and reconstruction for low sampling count

Department: Computer Science & Engineering
Research Institute Affiliation: Center for Visual Computing
Faculty Advisor(s): Ravi Ramamoorthi

Primary Student
Name: Alexandr Kuznetsov
Email: a1kuznet@ucsd.edu
Phone: 858-822-1483
Grad Year: 2021

Student Collaborators
Nima Khademi Kalantari, nkhademi@ucsd.edu

Monte Carlo rendering is a standard method in 3D graphic for rendering high quality images. However , Monte Carlo method requires a large number of sample to get a noise free image, causing a long rendering time. We developed an algorithm for adaptively rendering and filtering with a very low sample count. Our method first renders a scene with 1 sample per pixel and uses it to generate a sampling map via a neural network. Because we are using just 1 sample per pixel, previous methods fail to produce a good sampling map. Using our sampling map, we distribute just 3 additional spp and then filter the resulting render with another neural network. In order to train sampling map network, we need to be able to differentiate the error with respect to the sampling count. Our main technical contribution is differentiation of second input image with respect to the sampling map. By doing so, we are able to optimize the sampling map, throwing additional samples where it matters the most for the given denoiser. Moreover, our system is fully differentiable. We are able to better handle complex effects such as depth of field, global illumination and soft shadows and features not captured by auxiliary bufferers.

Industry Application Area(s)
Software, Analytics

« Back to Posters or Search Results