Title:	SAMPLE SCENES FROM A DISTRIBUTION: A CONTENT CREATION PIPELINE FOR REALISTIC RENDERING FOR NEURAL NETWORKS TRAINING
Author(s):	Vadim Sanzharov, Vladimir Frolov, Alexey Voloboy and Vladimir Galaktionov
ISBN:	978-989-8704-21-4
Editors:	Yingcai Xiao, Ajith P. Abraham and Jörg Roth
Year:	2020
Edition:	Single
Keywords:	Realistic Rendering, Procedural Pipeline, 3D Scene Generator, GPU Rendering, CNN Training Datasets
Type:	Full
First Page:	115
Last Page:	123
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	In this paper we present our experience with development of a content creation pipeline targeted at generation of realistic image sequences with highly variable content. Our technique allows rendering of a single 3D object or a 3D scene in variety of appearances which includes changing of geometry, materials and lighting. In our work we were able to generate datasets for individual 3D objects and create procedural generator of interior scenes. Our solution is highly controllable and allows generating datasets with desired distribution of features in a reproducible manner. Using synthetic data in training, we have got increase of accuracy in CNN-based models comparing to usage of real-life data only. During our work we had to significantly improve the content creation pipeline for the existing open source GPU rendering system adapting it for our tasks. In this paper we suggest new approach for content creation which we call “sampling scenes from a distribution”.
	Go Back