On the effects of normalization in adaptive MRF hierarchies

Albert Y.C. Chen, Jason J. Corso

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

Abstract

In this paper, we analyze the effects of energy normalization in adaptive-hierarchy-based energy minimization methods. Adaptive hierarchies provide a convenient multi-level abstraction of the underlying MRF. They have been shown to both accelerate computation and help avoid local minima. However, the standard recursive way of accumulating energy throughout the hierarchy causes energy terms to grow at different rates. Consequently, the faster-growing term, typically the unary term, dominates the overall energy at coarser level nodes, which hinders larger-scale energy/label change from happening. To solve the problem, we first investigate the theory and construction of adaptive hierarchies, then we analyze the theoretical bounds and expected values of its energy terms. Based on these analyses, we design and experimentally analyze three different energy-normalizing schemes. Our experiments show that properly normalized energies facilitate better use of the hierarchies during optimization: we observe an average improvement in the speed by 15% with the same accuracy.

Original languageEnglish
Title of host publicationComputational Modeling of Objects Represented in Images - Second International Symposium, CompIMAGE 2010, Proceedings
Pages275-286
Number of pages12
DOIs
StatePublished - 2010
Event2nd International Symposium on Computational Modeling of Objects Represented in Images, Fundamentals, Methods and Applications, CompIMAGE 2010 - Buffalo, NY, United States
Duration: 5 May 20107 May 2010

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume6026 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference2nd International Symposium on Computational Modeling of Objects Represented in Images, Fundamentals, Methods and Applications, CompIMAGE 2010
Country/TerritoryUnited States
CityBuffalo, NY
Period5/05/107/05/10

Fingerprint

Dive into the research topics of 'On the effects of normalization in adaptive MRF hierarchies'. Together they form a unique fingerprint.

Cite this