BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Cambridge Image Analysis Seminars
SUMMARY:Global MAP-Optimality by Shrinking the Combinatori
al Search Area with Convex Relaxation - Bogdan Sav
chynskyy\, University of Heidelberg
DTSTART;TZID=Europe/London:20140521T150000
DTEND;TZID=Europe/London:20140521T160000
UID:TALK52499AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/52499
DESCRIPTION:We consider energy minimization for undirected gra
phical models\, also known as the MAP-inference pr
oblem for Markov random fields. Although\ncombinat
orial methods\, which return a provably optimal in
tegral solution of the problem\, made a significan
t progress in the past decade\, they are still typ
ically unable to cope with large-scale datasets. O
n the other hand\, large scale datasets are often
defined on sparse graphs and convex relaxation met
hods\, such as linear programming relaxations then
provide good approximations to integral solutions
. We propose a novel method of combining combinato
rial and convex programming techniques to obtain a
global solution of the initial combinatorial prob
lem. Based on the information obtained from the so
lution of the convex\nrelaxation\, our method conf
ines application of the combinatorial solver to a
small fraction of the initial graphical model\, wh
ich allows to optimally solve much larger problems
. We demonstrate the efficacy of our approach on a
computer vision energy minimization benchmark.
LOCATION:MR 5\, Centre for Mathematical Sciences
CONTACT:Dr Jan Lellmann
END:VEVENT
END:VCALENDAR