BEGIN:VCALENDAR
VERSION:2.0
PRODID:www.dresden-science-calendar.de
METHOD:PUBLISH
CALSCALE:GREGORIAN
X-MICROSOFT-CALSCALE:GREGORIAN
X-WR-TIMEZONE:Europe/Berlin
BEGIN:VTIMEZONE
TZID:Europe/Berlin
X-LIC-LOCATION:Europe/Berlin
BEGIN:DAYLIGHT
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
DTSTART:19810329T030000
RRULE:FREQ=YEARLY;INTERVAL=1;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
DTSTART:19961027T030000
RRULE:FREQ=YEARLY;INTERVAL=1;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:DSC-13939
DTSTART;TZID=Europe/Berlin:20180116T160000
SEQUENCE:1516088843
TRANSP:OPAQUE
DTEND;TZID=Europe/Berlin:20180116T170000
URL:https://www.dresden-science-calendar.de/calendar/de/detail/13939
LOCATION:TUD Andreas-Pfitzmann-Bau\, Nöthnitzer Straße 4601069 Dresden
SUMMARY:Krull: Learning Sampling-Based 6D Object Pose Estimation
CLASS:PUBLIC
DESCRIPTION:Speaker: Dipl.-Medieninf. Alexander Krull\nInstitute of Speaker
 : Institut für Künstliche Intelligenz\, Professur für Computer Vision\n
 Topics:\nInformatik\n Location:\n  Name: TUD Andreas-Pfitzmann-Bau (APB 10
 04 (Ratssaal))\n  Street: Nöthnitzer Straße 46\n  City: 01069 Dresden\n 
  Phone: \n  Fax: \nDescription: The task of 6D object pose estimation\, i.
 e. of estimating an object's position (three degrees of freedom) and orien
 tation (three degrees of freedom) from images is an essential building blo
 ck of many modern applications\, such as robotic grasping\, autonomous dri
 ving\, or augmented reality. Automatic pose estimation systems have to ove
 rcome a variety of visual ambiguities\, including texture-less objects\, c
 lutter\, and occlusion. Since many applications demand real time performan
 ce the efficient use of computational resources is an additional challenge
 . In this thesis\, we will take a probabilistic stance on trying to overco
 me said issues. We build on a highly successful automatic pose estimation 
 framework based on predicting pixel-wise correspondences between the camer
 a coordinate system and the local coordinate system of the object. These d
 ense correspondences are used to generate a pool of hypotheses\, which in 
 turn serve as a starting point in a final search procedure. We will presen
 t three systems that each use probabilistic modeling and sampling to impro
 ve upon different aspects of the framework. The goal of the first system\,
  System I\, is to enable pose tracking\, i.e. estimating the pose of an ob
 ject in a sequence of frames instead of a single image. By including infor
 mation from previous frames tracking systems can resolve many visual ambig
 uities and reduce computation time. System I is a particle filter (PF) app
 roach. The PF represents its belief about the pose in each frame by propag
 ating a set of samples through time. Our system uses the process of hypoth
 esis generation from the original framework as part of a proposal distribu
 tion that efficiently concentrates samples in the appropriate areas. In Sy
 stem II\, we focus on the problem of evaluating the quality of pose hypoth
 eses. This task plays an essential role in the final search procedure of t
 he original framework. We use a convolutional neural network (CNN) to asse
 ss the quality of an hypothesis by comparing rendered and observed images.
  To train the CNN we view it as part of an energy based probability distri
 bution in pose space. This probabilistic perspective allows us to train th
 e system under the maximum likelihood paradigm. We use a sampling approach
  to approximate the required gradients. The resulting system for pose esti
 mation yields superior results in particular for highly occluded objects. 
 In System III\, we take the idea of machine learning a step further. Inste
 ad of learning to predict an hypothesis quality measure\, to be used in a 
 search procedure\, we present a way of learning the search procedure itsel
 f. We train a reinforcement learning (RL) agent\, termed PoseAgent\, to st
 eer the search process and make optimal use of a given computational budge
 t. PoseAgent dynamically decides which hypothesis should be refined next\,
  and which one should ultimately be output as the system's estimate. Since
  the search procedure includes discrete non-differentiable choices\, train
 ing of the system via gradient descent is not easily possible. To solve th
 e problem\, we model PoseAgent's behavior as non-deterministic stochastic 
 policy\, which is ultimately governed by a CNN. This allows us to use a sa
 mpling-based stochastic policy gradient training procedure. We believe tha
 t some of the ideas developed in this thesis\, such as the sampling driven
  probabilistically motivated training of a CNN for the comparison of image
 s or the search procedure implemented by PoseAgent have the potential to b
 e applied in fields beyond pose stimation as well.
DTSTAMP:20260717T104554Z
CREATED:20180104T080156Z
LAST-MODIFIED:20180116T074723Z
END:VEVENT
END:VCALENDAR