Tampere University of Technology

TUTCRIS Research Portal

Machine Learning Is the Solution Also for Foveated Path Tracing Reconstruction

Research output: Chapter in Book/Report/Conference proceedingConference contributionScientificpeer-review

Details

Original languageEnglish
Title of host publicationProceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 1: GRAPP
PublisherSCITEPRESS
Pages361-367
ISBN (Electronic)978-989758402-2
DOIs
Publication statusPublished - Feb 2020
Publication typeA4 Article in a conference publication
EventInternational conference on computer graphics theory and applications - Valletta, Malta
Duration: 27 Feb 202029 Feb 2020

Conference

ConferenceInternational conference on computer graphics theory and applications
CountryMalta
CityValletta
Period27/02/2029/02/20

Abstract

Real-time photorealistic rendering requires a lot of computational power. Foveated rendering reduces the work by focusing the effort to where the user is looking, but the very sparse sampling in the periphery requires fast reconstruction algorithms with good quality. The problem is even more complicated in the field of foveated path tracing where the sparse samples are also noisy. In this position paper we argue that machine learning and data-driven methods play an important role in the future of real-time foveated rendering. In order to show initial proofs to support this opinion, we propose a preliminary machine learning based method which is able to improve the reconstruction quality of foveated path traced image by using spatio-temporal input data. Moreover, the method is able to run in the same reduced foveated resolution as the path tracing setup. The reconstruction using the preliminary network is about 2.9ms per 658 × 960 frame on a GeForce RTX 2080 Ti GPU.

Publication forum classification

Field of science, Statistics Finland