Sublabel-Accurate Multilabeling Meets Product Label Spaces

Zhenzhang Ye, Bjoern Haefner, Yvain Quéau, Thomas Möllenhoff, Daniel Cremers

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


Functional lifting methods are a promising approach to determine optimal or near-optimal solutions to difficult nonconvex variational problems. Yet, they come with increased memory demands, limiting their practicability. To overcome this drawback, this paper presents a combination of two approaches designed to make liftings more scalable, namely product-space relaxations and sublabel-accurate discretizations. Our main contribution is a simple way to solve the resulting semi-infinite optimization problem with a sampling strategy. We show that despite its simplicity, our approach significantly outperforms baseline methods, in the sense that it finds solutions with lower energies given the same amount of memory. We demonstrate our empirical findings on the nonconvex optical flow and manifold-valued denoising problems.

Original languageEnglish
Title of host publicationPattern Recognition - 43rd DAGM German Conference, DAGM GCPR 2021, Proceedings
EditorsChristian Bauckhage, Juergen Gall, Alexander Schwing
PublisherSpringer Science and Business Media Deutschland GmbH
Number of pages15
ISBN (Print)9783030926588
StatePublished - 2021
Event43rd DAGM German Conference on Pattern Recognition, DAGM GCPR 2021 - Virtual, Online
Duration: 28 Sep 20211 Oct 2021

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13024 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference43rd DAGM German Conference on Pattern Recognition, DAGM GCPR 2021
CityVirtual, Online


  • Convex relaxation
  • Global optimization
  • Manifold-valued problems
  • Variational methods


Dive into the research topics of 'Sublabel-Accurate Multilabeling Meets Product Label Spaces'. Together they form a unique fingerprint.

Cite this