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:Frontiers in Artificial Intelligence Series
SUMMARY:Probabilistic and Deep Models for 3D Reconstructio
n - Andreas Geiger\, Max Planck Institute for Inte
lligent Systems
DTSTART;TZID=Europe/London:20170926T130000
DTEND;TZID=Europe/London:20170926T140000
UID:TALK80361AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/80361
DESCRIPTION:3D reconstruction from multiple 2D images is an in
herently ill-posed problem. Prior knowledge is req
uired to resolve ambiguities and probabilistic mod
els are desirable to capture the ambiguities in th
e reconstructed model. In this talk\, I will prese
nt two recent results tackling these two aspects.
First\, I will introduce a probabilistic framework
for volumetric 3D reconstruction where the recons
truction problem is cast as inference in a Markov
random field using ray potentials. Our main contri
bution is a discrete-continuous inference algorith
m which computes marginal distributions of each vo
xel's occupancy and appearance. I will show that t
he proposed algorithm allows for Bayes optimal pre
dictions with respect to a natural reconstruction
loss. I will further demonstrate several extension
s which integrate non-local CAD priors into the re
construction process. In the second part of my tal
k\, I will present a novel framework for deep lear
ning with 3D data called OctNet which enables 3D C
NNs on high-dimensional inputs. I will demonstrate
the utility of the OctNet representation on sever
al 3D tasks including classification\, orientation
estimation and point cloud labeling. Finally\, I
will present an extension of OctNet called OctNetF
usion which jointly predicts the space partitionin
g function with the output representation\, result
ing in an end-to-end trainable model for volumetri
c depth map fusion.
LOCATION:Auditorium\, Microsoft Research Ltd\, 21 Station R
oad\, Cambridge\, CB1 2FB
CONTACT:Microsoft Research Cambridge Talks Admins
END:VEVENT
END:VCALENDAR