Deep Intrinsic Image Decomposition Using Joint Parallel Learning

Yuan Yuan, Bin Sheng, Ping Li, Lei Bi, Jinman Kim, Enhua Wu

Research output: Chapter in book / Conference proceedingConference article published in proceeding or bookAcademic researchpeer-review

3 Citations (Scopus)

Abstract

Intrinsic image decomposition is a highly ill-posed problem in computer vision referring to extract albedo and shading from an image. In this paper, we regard it as an image-to-image translation issue and propose a novel thought, which makes use of parallel convolutional neural networks (ParCNN) to learn albedo and shading with different spatial features and data distributions, respectively. At the same time, the energy is preserved as much as possible under the constraint of image reconstruction loss shared by the two networks. Moreover, we add the gradient prior based on the traditional image formation process into the loss function, which can lead to a performance improvement of our basic learning model by jointing advantages of the physically-based method and the data-driven method. We choose MPI Sintel dataset for model training and testing. Quantitative and qualitative evaluation results outperform the state-of-the-art methods.

Original languageEnglish
Title of host publicationAdvances in Computer Graphics - 36th Computer Graphics International Conference, CGI 2019, Proceedings
EditorsMarina Gavrilova, Jian Chang, Nadia Magnenat Thalmann, Nadia Magnenat Thalmann, Eckhard Hitzer, Hiroshi Ishikawa
PublisherSpringer-Verlag
Pages336-341
Number of pages6
ISBN (Print)9783030225131
DOIs
Publication statusPublished - 1 Jan 2019
Externally publishedYes
Event36th Computer Graphics International Conference, CGI 2019 - Calgary, Canada
Duration: 17 Jun 201920 Jun 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11542 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference36th Computer Graphics International Conference, CGI 2019
Country/TerritoryCanada
CityCalgary
Period17/06/1920/06/19

Keywords

  • Gradient priors
  • Intrinsic image decomposition
  • ParCNN

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Cite this