An alternating direction method for optical flow estimation with lp regularization

Naftali Zon, Nahum Kiryati

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

Abstract

We consider the optical flow estimation problem with lp sub-quadratic regularization, where 0 ≤ p ≤ 1. As in other image analysis tasks based on functional minimization, sub-quadratic regularization is expected to admit discontinuities and avoid oversmoothing of the estimated optical flow field. The problem is mathematically challenging, since the regularization term is non-differentiable. It is harder than the l1 case, that can be addressed via Moreau proximal mapping with a closed form solution. In this paper, we propose a novel approach, based on variable splitting and the Alternating Direction Method of Multipliers (ADMM). We exemplify that our method can outperform optical flow with l1 regularization, but this is not the essence of this paper. The contribution is in demonstrating that state of the art optimization methods can be harnessed to solve a mathematically-challenging class of important image processing problems, and to highlight crucial numerical aspects that are often obscured in the image processing literature.

Original languageEnglish
Title of host publication2016 IEEE International Conference on the Science of Electrical Engineering, ICSEE 2016
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9781509021529
DOIs
StatePublished - 4 Jan 2017
Event2016 IEEE International Conference on the Science of Electrical Engineering, ICSEE 2016 - Eilat, Israel
Duration: 16 Nov 201618 Nov 2016

Publication series

Name2016 IEEE International Conference on the Science of Electrical Engineering, ICSEE 2016

Conference

Conference2016 IEEE International Conference on the Science of Electrical Engineering, ICSEE 2016
Country/TerritoryIsrael
CityEilat
Period16/11/1618/11/16

Fingerprint

Dive into the research topics of 'An alternating direction method for optical flow estimation with lp regularization'. Together they form a unique fingerprint.

Cite this